Parallel Computing Toolbox™UserGuide
Parallel Computing Toolbox™UserGuide
R2011b
Product enhancement suggestions Bug reports Documentation error reports Order status, license renewals, passcodes Sales, pricing, and general information
508-647-7000 (Phone) 508-647-7001 (Fax) The MathWorks, Inc. 3 Apple Hill Drive Natick, MA 01760-2098
For contact information about worldwide offices, see the MathWorks Web site. Parallel Computing Toolbox Users Guide COPYRIGHT 20042011 by The MathWorks, Inc.
The software described in this document is furnished under a license agreement. The software may be used or copied only under the terms of the license agreement. No part of this manual may be photocopied or reproduced in any form without prior written consent from The MathWorks, Inc. FEDERAL ACQUISITION: This provision applies to all acquisitions of the Program and Documentation by, for, or through the federal government of the United States. By accepting delivery of the Program or Documentation, the government hereby agrees that this software or documentation qualifies as commercial computer software or commercial computer software documentation as such terms are used or defined in FAR 12.212, DFARS Part 227.72, and DFARS 252.227-7014. Accordingly, the terms and conditions of this Agreement and only those rights specified in this Agreement, shall pertain to and govern the use, modification, reproduction, release, performance, display, and disclosure of the Program and Documentation by the federal government (or other entity acquiring for or through the federal government) and shall supersede any conflicting contractual terms or conditions. If this License fails to meet the governments needs or is inconsistent in any respect with federal procurement law, the government agrees to return the Program and Documentation, unused, to The MathWorks, Inc.
Trademarks
MATLAB and Simulink are registered trademarks of The MathWorks, Inc. See www.mathworks.com/trademarks for a list of additional trademarks. Other product or brand names may be trademarks or registered trademarks of their respective holders.
Patents
MathWorks products are protected by one or more U.S. patents. Please see www.mathworks.com/patents for more information.
Revision History
November 2004 March 2005 September 2005 November 2005 March 2006 September 2006 March 2007 September 2007 March 2008 October 2008 March 2009 September 2009 March 2010 September 2010 April 2011 September 2011
Online only Online only Online only Online only Online only Online only Online only Online only Online only Online only Online only Online only Online only Online only Online only Online only
New for Version 1.0 (Release 14SP1+) Revised for Version 1.0.1 (Release 14SP2) Revised for Version 1.0.2 (Release 14SP3) Revised for Version 2.0 (Release 14SP3+) Revised for Version 2.0.1 (Release 2006a) Revised for Version 3.0 (Release 2006b) Revised for Version 3.1 (Release 2007a) Revised for Version 3.2 (Release 2007b) Revised for Version 3.3 (Release 2008a) Revised for Version 4.0 (Release 2008b) Revised for Version 4.1 (Release 2009a) Revised for Version 4.2 (Release 2009b) Revised for Version 4.3 (Release 2010a) Revised for Version 5.0 (Release 2010b) Revised for Version 5.1 (Release 2011a) Revised for Version 5.2 (Release 2011b)
Contents
Getting Started
1
Product Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Key Problems Addressed by Parallel Computing . . . . . Running Parallel for-Loops (parfor) . . . . . . . . . . . . . . . . . . . Executing Batch Jobs in Parallel . . . . . . . . . . . . . . . . . . . . . Partitioning Large Data Sets . . . . . . . . . . . . . . . . . . . . . . . . Introduction to Parallel Solutions . . . . . . . . . . . . . . . . . . . Interactively Run a Loop in Parallel . . . . . . . . . . . . . . . . . . Run a Batch Job . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Run a Batch Parallel Loop . . . . . . . . . . . . . . . . . . . . . . . . . . Run Scripts as Batch Jobs from the Current Folder Browser . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Distributing Arrays and Running SPMD . . . . . . . . . . . . . . Determining Product Installation and Versions . . . . . . 1-2 1-3 1-3 1-4 1-4 1-5 1-5 1-7 1-8 1-10 1-11 1-13
2
Getting Started with parfor . . . . . . . . . . . . . . . . . . . . . . . . parfor-Loops in MATLAB . . . . . . . . . . . . . . . . . . . . . . . . . . . Deciding When to Use parfor . . . . . . . . . . . . . . . . . . . . . . . . Creating a parfor-Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Differences Between for-Loops and parfor-Loops . . . . . . . . Reduction Assignments: Values Updated by Each Iteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Displaying Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Programming Considerations . . . . . . . . . . . . . . . . . . . . . . . 2-2 2-2 2-3 2-4 2-6 2-7 2-8 2-9
MATLAB Path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Error Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Using Objects in parfor Loops . . . . . . . . . . . . . . . . . . . . . . . Performance Considerations . . . . . . . . . . . . . . . . . . . . . . . . . Compatibility with Earlier Versions of MATLAB Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Advanced Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . About Programming Notes . . . . . . . . . . . . . . . . . . . . . . . . . . Classification of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . Improving Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2-9 2-9 2-10 2-15 2-15 2-16 2-17 2-17 2-17 2-32
3
Executing Simultaneously on Multiple Data Sets . . . . . Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . When to Use spmd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Setting Up MATLAB Resources Using matlabpool . . . . . . . Defining an spmd Statement . . . . . . . . . . . . . . . . . . . . . . . . Displaying Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Accessing Data with Composites . . . . . . . . . . . . . . . . . . . . Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Creating Composites in spmd Statements . . . . . . . . . . . . . . Variable Persistence and Sequences of spmd . . . . . . . . . . . Creating Composites Outside spmd Statements . . . . . . . . . Distributing Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Distributed Versus Codistributed Arrays . . . . . . . . . . . . . . Creating Distributed Arrays . . . . . . . . . . . . . . . . . . . . . . . . Creating Codistributed Arrays . . . . . . . . . . . . . . . . . . . . . . . Programming Tips . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . MATLAB Path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Error Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2 3-2 3-2 3-3 3-4 3-6 3-7 3-7 3-7 3-9 3-10 3-12 3-12 3-12 3-13 3-15 3-15 3-15 3-15
vi
Contents
4
pmode Versus spmd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Run Parallel Jobs Interactively Using pmode . . . . . . . . Parallel Command Window . . . . . . . . . . . . . . . . . . . . . . . . . Running pmode Interactive Jobs on a Cluster . . . . . . . . Plotting Distributed Data Using pmode . . . . . . . . . . . . . . pmode Limitations and Unexpected Results . . . . . . . . . . Using Graphics in pmode . . . . . . . . . . . . . . . . . . . . . . . . . . . pmode Troubleshooting . . . . . . . . . . . . . . . . . . . . . . . . . . . . Connectivity Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hostname Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Socket Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-2 4-3 4-10 4-15 4-16 4-18 4-18 4-19 4-19 4-19 4-19
5
Nondistributed Versus Distributed Arrays . . . . . . . . . . . Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nondistributed Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Codistributed Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Working with Codistributed Arrays . . . . . . . . . . . . . . . . . How MATLAB Software Distributes Arrays . . . . . . . . . . . . Creating a Codistributed Array . . . . . . . . . . . . . . . . . . . . . . Local Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Obtaining information About the Array . . . . . . . . . . . . . . . Changing the Dimension of Distribution . . . . . . . . . . . . . . . Restoring the Full Array . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-2 5-2 5-2 5-4 5-5 5-5 5-7 5-11 5-12 5-13 5-14
vii
Indexing into a Codistributed Array . . . . . . . . . . . . . . . . . . 2-Dimensional Distribution . . . . . . . . . . . . . . . . . . . . . . . . . Using a for-Loop Over a Distributed Range (for-drange) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Parallelizing a for-Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Codistributed Arrays in a for-drange Loop . . . . . . . . . . . . . Using MATLAB Functions on Codistributed Arrays . . .
5-15 5-17
Programming Overview
6
How Parallel Computing Products Run a Job . . . . . . . . Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Toolbox and Server Components . . . . . . . . . . . . . . . . . . . . . Life Cycle of a Job . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Create Simple Distributed Jobs . . . . . . . . . . . . . . . . . . . . . Evaluate a Basic Function . . . . . . . . . . . . . . . . . . . . . . . . . . Program a Basic Job with a Local Scheduler . . . . . . . . . . . Getting Help . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Parallel Configurations for Cluster Access . . . . . . . . . . . Defining Configurations . . . . . . . . . . . . . . . . . . . . . . . . . . . . Exporting and Importing Configurations . . . . . . . . . . . . . . Validate Configurations . . . . . . . . . . . . . . . . . . . . . . . . . . . . Applying Configurations in Client Code . . . . . . . . . . . . . . . Job Monitor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Job Monitor GUI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Manage Jobs Using the Job Monitor . . . . . . . . . . . . . . . . . . Identify Task Errors Using the Job Monitor . . . . . . . . . . . . Programming Tips . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Program Development Guidelines . . . . . . . . . . . . . . . . . . . . Current Working Directory of a MATLAB Worker . . . . . . . Writing to Files from Workers . . . . . . . . . . . . . . . . . . . . . . . 6-2 6-2 6-3 6-8 6-10 6-10 6-10 6-12 6-14 6-14 6-21 6-22 6-25 6-27 6-27 6-28 6-29 6-30 6-30 6-31 6-32
viii
Contents
Saving or Sending Objects . . . . . . . . . . . . . . . . . . . . . . . . . . Using clear functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Running Tasks That Call Simulink Software . . . . . . . . . . . Using the pause Function . . . . . . . . . . . . . . . . . . . . . . . . . . . Transmitting Large Amounts of Data . . . . . . . . . . . . . . . . . Interrupting a Job . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Speeding Up a Job . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Profiling Parallel Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Collecting Parallel Profile Data . . . . . . . . . . . . . . . . . . . . . . Viewing Parallel Profile Data . . . . . . . . . . . . . . . . . . . . . . . . Parallel Profiler Demos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Benchmarking Performance . . . . . . . . . . . . . . . . . . . . . . . . Demos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . HPC Challenge Benchmarks . . . . . . . . . . . . . . . . . . . . . . . . Troubleshooting and Debugging . . . . . . . . . . . . . . . . . . . . Object Data Size Limitations . . . . . . . . . . . . . . . . . . . . . . . . File Access and Permissions . . . . . . . . . . . . . . . . . . . . . . . . . No Results or Failed Job . . . . . . . . . . . . . . . . . . . . . . . . . . . . Connection Problems Between the Client and Job Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . SFTP Error: Received Message Too Long . . . . . . . . . . . . . .
6-32 6-33 6-33 6-33 6-33 6-33 6-34 6-35 6-35 6-35 6-36 6-45 6-47 6-47 6-47 6-48 6-48 6-48 6-50 6-51 6-52
7
Evaluate Functions Synchronously . . . . . . . . . . . . . . . . . Scope of dfeval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Arguments of dfeval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Example Use dfeval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Evaluate Functions Asynchronously 7-2 7-2 7-3 7-4 7-8
................
ix
8
Use a Local Scheduler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Create and Run Jobs with a Local Scheduler . . . . . . . . . . . Local Scheduler Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . Use a Job Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Creating and Running Jobs with a Job Manager . . . . . . . . Share Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Manage Objects in the Job Manager . . . . . . . . . . . . . . . . . . Use a Fully Supported Third-Party Scheduler . . . . . . . . Create and Run Jobs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Share Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Manage Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Use the Generic Scheduler Interface . . . . . . . . . . . . . . . . Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . MATLAB Client Submit Function . . . . . . . . . . . . . . . . . . . . Example Write the Submit Function . . . . . . . . . . . . . . . MATLAB Worker Decode Function . . . . . . . . . . . . . . . . . . . Example Write the Decode Function . . . . . . . . . . . . . . . . Example Program and Run a Job in the Client . . . . . . . Supplied Submit and Decode Functions . . . . . . . . . . . . . . . Manage Jobs with Generic Scheduler . . . . . . . . . . . . . . . . . Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-2 8-2 8-7 8-8 8-8 8-14 8-17 8-21 8-21 8-29 8-31 8-34 8-34 8-35 8-39 8-40 8-43 8-43 8-48 8-50 8-53
9
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Use a Supported Scheduler . . . . . . . . . . . . . . . . . . . . . . . . . Schedulers and Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . Code the Task Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . Code in the Client . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-2 9-4 9-4 9-4 9-5
Contents
Use the Generic Scheduler Interface . . . . . . . . . . . . . . . . Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Code in the Client . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Further Notes on Parallel Jobs . . . . . . . . . . . . . . . . . . . . . Number of Tasks in a Parallel Job . . . . . . . . . . . . . . . . . . . . Avoid Deadlock and Other Dependency Errors . . . . . . . . . .
GPU Computing
10
When to Use a GPU for Matrix Operations . . . . . . . . . . . Capabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Demos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Using GPUArray . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Transfer Data Between Workspace and GPU . . . . . . . . . . . Create GPU Data Directly . . . . . . . . . . . . . . . . . . . . . . . . . . Examine GPUArray Characteristics . . . . . . . . . . . . . . . . . . Built-In Functions That Support GPUArray . . . . . . . . . . . . Execute MATLAB Code on a GPU . . . . . . . . . . . . . . . . . . . MATLAB Code vs. GPUArray Objects . . . . . . . . . . . . . . . . Running Your MATLAB Functions on the GPU . . . . . . . . . Example: Running Your MATLAB Code . . . . . . . . . . . . . . . Supported MATLAB Code . . . . . . . . . . . . . . . . . . . . . . . . . . 10-2 10-2 10-2 10-3 10-4 10-4 10-5 10-7 10-7 10-10 10-10 10-10 10-11 10-12
Identify and Select a GPU from Multiple GPUs . . . . . . . 10-15 Example: Selecting a GPU . . . . . . . . . . . . . . . . . . . . . . . . . . 10-15 Executing CUDA or PTX Code on the GPU . . . . . . . . . . . Creating Kernels from CU Files . . . . . . . . . . . . . . . . . . . . . . Running the Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Determining Input and Output Correspondence . . . . . . . . Kernel Object Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . Specifying Entry Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . Providing C Prototype Input . . . . . . . . . . . . . . . . . . . . . . . . 10-17 10-17 10-17 10-18 10-19 10-20 10-21
xi
Complete Kernel Workflow . . . . . . . . . . . . . . . . . . . . . . . . . . 10-22 GPU Characteristics and Limitations . . . . . . . . . . . . . . . . 10-25 Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-25 Complex Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-25
Object Reference
11
Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Schedulers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Generic Scheduler Interface Tools . . . . . . . . . . . . . . . . . . Jobs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Workers 11-2 11-2 11-3 11-3 11-3 11-4 11-4
..........................................
12
Function Reference
13
Parallel Code Execution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-2
xii
Contents
Parallel Code on a MATLAB Pool . . . . . . . . . . . . . . . . . . . . Configuration, Input, and Output . . . . . . . . . . . . . . . . . . . . Interactive Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Distributed and Codistributed Arrays . . . . . . . . . . . . . . . Toolbox Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Overloaded MATLAB Functions . . . . . . . . . . . . . . . . . . . . . Jobs and Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Job Creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Job Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Task Execution Information . . . . . . . . . . . . . . . . . . . . . . . . . Object Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Interlab Communication Within a Parallel Job . . . . . . .
13-2 13-2 13-3 13-3 13-3 13-4 13-6 13-6 13-7 13-8 13-9 13-9
14
Property Reference
15
Job Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Schedulers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jobs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15-2 15-3 15-5 15-7
xiii
Workers
..........................................
15-8
16
Glossary
Index
xiv
Contents
1
Getting Started
Product Overview on page 1-2 Key Problems Addressed by Parallel Computing on page 1-3 Introduction to Parallel Solutions on page 1-5 Determining Product Installation and Versions on page 1-13
Getting Started
Product Overview
Parallel Computing Toolbox lets you solve computationally and data-intensive problems using multicore processors, GPUs, and computer clusters. High-level constructsparallel for-loops, special array types, and parallelized numerical algorithmslet you parallelize MATLAB applications without CUDA or MPI programming. You can use the toolbox with Simulink to run multiple simulations of a model in parallel. The toolbox provides twelve workers (MATLAB computational engines) to execute applications locally on a multicore desktop. Without changing the code, you can run the same application on a computer cluster or a grid computing service (using MATLAB Distributed Computing Server). You can run parallel applications interactively or in batch. MATLAB Distributed Computing Server software allows you to run as many MATLAB workers on a remote cluster of computers as your licensing allows. You can also use MATLAB Distributed Computing Server to run workers on your client machine if you want to run more than twelve local workers. Most MathWorks products let you code in such a way as to run applications in parallel. For example, Simulink models can run simultaneously in parallel, as described in Running Parallel Simulations. MATLAB Compiler software lets you build and deploy parallel applications, as shown in Deploying Applications Created Using Parallel Computing Toolbox. Several MathWorks products now offer built-in support for the parallel computing products, without requiring extra coding. For the current list of these products and their parallel functionality, see:
http://www.mathworks.com/products/parallel-computing/builtin-parallel-support.html
1-2
Many iterations A sweep might take a long time because it comprises many iterations. Each iteration by itself might not take long to execute, but to complete thousands or millions of iterations in serial could take a long time. Long iterations A sweep might not have a lot of iterations, but each iteration could take a long time to run.
Typically, the only difference between iterations is defined by different input data. In these cases, the ability to run separate sweep iterations simultaneously can improve performance. Evaluating such iterations in parallel is an ideal way to sweep through large or multiple data sets. The only restriction on parallel loops is that no iterations be allowed to depend on any other iterations. Test suites with independent segments For applications that run a series of unrelated tasks, you can run these tasks simultaneously on separate resources. You might not have used a for-loop for a case such as this comprising distinctly different tasks, but a parfor-loop could offer an appropriate solution. Parallel Computing Toolbox software improves the performance of such loop execution by allowing several MATLAB workers to execute individual loop iterations simultaneously. For example, a loop of 100 iterations could run on
1-3
Getting Started
a cluster of 20 MATLAB workers, so that simultaneously, the workers each execute only five iterations of the loop. You might not get quite 20 times improvement in speed because of communications overhead and network traffic, but the speedup should be significant. Even running local workers all on the same machine as the client, you might see significant performance improvement on a multicore/multiprocessor machine. So whether your loop takes a long time to run because it has many iterations or because each iteration takes a long time, you can improve your loop speed by distributing iterations to MATLAB workers.
1-4
waveform:
for i=1:1024 A(i) = sin(i*2*pi/1024); end plot(A)
2 To interactively run code that contains a parallel loop, you first open a
MATLAB pool. This reserves a collection of MATLAB worker sessions to run your loop iterations. The MATLAB pool can consist of MATLAB sessions running on your local machine or on a remote cluster:
matlabpool open local 3
3 With the MATLAB pool reserved, you can modify your code to run your loop
1-5
Getting Started
The only difference in this loop is the keyword parfor instead of for. After the loop runs, the results look the same as those generated from the previous for-loop.
MATLAB workers parfor MATLAB client
Because the iterations run in parallel in other MATLAB sessions, each iteration must be completely independent of all other iterations. The worker calculating the value for A(100) might not be the same worker calculating A(500). There is no guarantee of sequence, so A(900) might be calculated before A(400). (The MATLAB Editor can help identify some problems with parfor code that might not contain independent iterations.) The only place where the values of all the elements of the array A are available is in the MATLAB client, after the data returns from the MATLAB workers and the loop completes.
4 When you are finished with your code, close the MATLAB pool and release
the workers:
matlabpool close
For more information on parfor-loops, see Chapter 2, Parallel for-Loops (parfor). The examples in this section run on three local workers. With parallel configurations, you can control how many workers run your loops, and whether the workers are local or remote. For more information on parallel configurations, see Parallel Configurations for Cluster Access on page 6-14.
1-6
You can run Simulink models in parallel loop iterations with the sim command inside your loop. For more information and examples of using Simulink with parfor, see Running Parallel Simulations in the Simulink documentation.
edit mywave
2 In the MATLAB Editor, enter the text of the for-loop:
5 The batch command does not block MATLAB, so you must wait for the job
the workspace of the client, where you can view the results:
load(job, 'A')
1-7
Getting Started
plot(A)
7 When the job is complete, permanently remove its data:
destroy(job)
edit mywave
2 Modify the script so that the for statement is a parfor statement:
that the script should use a MATLAB pool for the parallel loop:
job = batch('mywave', 'matlabpool', 3)
This command specifies that three workers (in addition to the one running the batch script) are to evaluate the loop iterations. Therefore, this example uses a total of four local workers, including the one worker running the batch script.
1-8
MATLAB workers
parfor
The results look the same as before, however, there are two important differences in execution: The work of defining the parfor-loop and accumulating its results are offloaded to another MATLAB session (batch). The loop iterations are distributed from one MATLAB worker to another set of workers running simultaneously (matlabpool and parfor), so the loop might run faster than having only one worker execute it.
6 When the job is complete, permanently remove its data:
destroy(job)
1-9
Getting Started
When you run a batch job from the browser, this also opens the Job Monitor. The Job Monitor is a tool that lets you track your job in the scheduler queue. For more information about the Job Monitor and its capabilities, see Job Monitor on page 6-27.
1-10
Now MM is a distributed array, equivalent to M, and you can manipulate or access its elements in the same way as any other array.
M2 = 2*MM; % M2 is also distributed, calculation performed on workers x = M2(1,1) % x on the client is set to first element of M2
When you are finished and have no further need of data from the labs, you can close the MATLAB pool. Data on the labs does not persist from one instance of a MATLAB pool to another.
matlabpool close
This code creates an individual 4-by-4 matrix, R, of random numbers on each lab in the pool.
1-11
Getting Started
Composites
Following an spmd statement, in the client context, the values from the block are accessible, even though the data is actually stored on the labs. On the client, these variables are called Composite objects. Each element of a composite is a symbol referencing the value (data) on a lab in the pool. Note that because a variable might not be defined on every lab, a Composite might have undefined elements. Continuing with the example from above, on the client, the Composite R has one element for each lab:
X = R{3}; % Set X to the value of R from lab 3.
The line above retrieves the data from lab 3 to assign the value of X. The following code sends data to lab 3:
X = X + 2; R{3} = X; % Send the value of X from the client to lab 3.
If the MATLAB pool remains open between spmd statements and the same labs are used, the data on each lab persists from one spmd statement to another.
spmd R = R + labindex end % Use values of R from previous spmd.
A typical use for spmd is to run the same code on a number of labs, each of which accesses a different set of data. For example:
spmd INP = load(['somedatafile' num2str(labindex) '.mat']); RES = somefun(INP) end
Then the values of RES on the labs are accessible from the client as RES{1} from lab 1, RES{2} from lab 2, etc. There are two forms of indexing a Composite, comparable to indexing a cell array:
1-12
AA{n} returns the values of AA from lab n. AA(n) returns a cell array of the content of AA from lab n. When you are finished with all spmd execution and have no further need of data from the labs, you can close the MATLAB pool.
matlabpool close
Although data persists on the labs from one spmd block to another as long as the MATLAB pool remains open, data does not persist from one instance of a MATLAB pool to another. For more information about using distributed arrays, spmd, and Composites, see Chapter 3, Single Program Multiple Data (spmd).
When you enter this command, MATLAB displays information about the version of MATLAB you are running, including a list of all toolboxes installed on your system and their version numbers. If you want to run your applications on a cluster, see your system administrator to verify that the version of Parallel Computing Toolbox you are using is the same as the version of MATLAB Distributed Computing Server installed on your cluster.
1-13
Getting Started
1-14
2
Parallel for-Loops (parfor)
Getting Started with parfor on page 2-2 Programming Considerations on page 2-9 Advanced Topics on page 2-17
parfor-Loops in MATLAB
The basic concept of a parfor-loop in MATLAB software is the same as the standard MATLAB for-loop: MATLAB executes a series of statements (the loop body) over a range of values. Part of the parfor body is executed on the MATLAB client (where the parfor is issued) and part is executed in parallel on MATLAB workers. The necessary data on which parfor operates is sent from the client to workers, where most of the computation happens, and the results are sent back to the client and pieced together. Because several MATLAB workers can be computing concurrently on the same loop, a parfor-loop can provide significantly better performance than its analogous for-loop. Each execution of the body of a parfor-loop is an iteration. MATLAB workers evaluate iterations in no particular order, and independently of each other. Because each iteration is independent, there is no guarantee that the iterations are synchronized in any way, nor is there any need for this. If the number of workers is equal to the number of loop iterations, each worker performs one iteration of the loop. If there are more iterations than workers, some workers perform more than one loop iteration; in this case, a worker might receive multiple iterations at once to reduce communication time.
2-2
2-3
Creating a parfor-Loop
Set Up MATLAB Resources Using matlabpool
You use the function matlabpool to reserve a number of MATLAB workers for executing a subsequent parfor-loop. Depending on your scheduler, the workers might be running remotely on a cluster, or they might run locally on your MATLAB client machine. You identify a scheduler and cluster by selecting a parallel configuration. For a description of how to manage and use configurations, see Parallel Configurations for Cluster Access on page 6-14. To begin the examples of this section, allocate local MATLAB workers for the evaluation of your loop iterations:
matlabpool
This command starts the number of MATLAB worker sessions defined by the default parallel configuration. If the local configuration is your default and does not specify the number of workers, this starts one worker per core (maximum of twelve) on your local MATLAB client machine. Note If matlabpool is not running, a parfor-loop runs serially on the client without regard for iteration sequence.
2-4
Notice that each element of A is equal to its index. The parfor-loop works because each element depends only upon its iteration of the loop, and upon no other iterations. for-loops that merely repeat such independent tasks are ideally suited candidates for parfor-loops.
2-5
Although the elements of A come out the same in both of these examples, the value of d does not. In the for-loop above on the left, the iterations execute in sequence, so afterward d has the value it held in the last iteration of the loop. In the parfor-loop on the right, the iterations execute in parallel, not in sequence, so it would be impossible to assign d a definitive value at the end of the loop. This also applies to the loop variable, i. Therefore, parfor-loop behavior is defined so that it does not affect the values d and i outside the loop at all, and their values remain the same before and after the loop. So, a parfor-loop requires that each iteration be independent of the other iterations, and that all code that follows the parfor-loop not depend on the loop iteration sequence.
2-6
If the loop iterations operate in random sequence, you might expect the concatenation sequence in the example on the right to be nonconsecutive. However, MATLAB recognizes the concatenation operation and yields deterministic results. The next example, which attempts to compute Fibonacci numbers, is not a valid parfor-loop because the value of an element of f in one iteration depends on the values of other elements of f calculated in other iterations.
f = zeros(1,50); f(1) = 1; f(2) = 2; parfor n = 3:50 f(n) = f(n-1) + f(n-2); end
When you are finished with your loop examples, clear your workspace and close or release your pool of workers:
clear matlabpool close
The following sections provide further information regarding programming considerations and limitations for parfor-loops.
2-7
Displaying Output
When running a parfor-loop on a MATLAB pool, all command-line output from the workers displays in the client Command Window, except output from variable assignments. Because the workers are MATLAB sessions without displays, any graphical output (for example, figure windows) from the pool does not display at all.
2-8
Programming Considerations
Programming Considerations
In this section... MATLAB Path on page 2-9 Error Handling on page 2-9 Limitations on page 2-10 Using Objects in parfor Loops on page 2-15 Performance Considerations on page 2-15 Compatibility with Earlier Versions of MATLAB Software on page 2-16
MATLAB Path
All workers executing a parfor-loop must have the same MATLAB path configuration as the client, so that they can execute any functions called in the body of the loop. Therefore, whenever you use cd, addpath, or rmpath on the client, it also executes on all the workers, if possible. For more information, see the matlabpool reference page. When the workers are running on a different platform than the client, use the function pctRunOnAll to properly set the MATLAB path on all workers.
Error Handling
When an error occurs during the execution of a parfor-loop, all iterations that are in progress are terminated, new ones are not initiated, and the loop terminates. Errors and warnings produced on workers are annotated with the worker ID and displayed in the clients Command Window in the order in which they are received by the client MATLAB. The behavior of lastwarn is unspecified at the end of the parfor if used within the loop body.
2-9
Limitations
Unambiguous Variable Names
If you use a name that MATLAB cannot unambiguously distinguish as a variable inside a parfor-loop, at parse time MATLAB assumes you are referencing a function. Then at run-time, if the function cannot be found, MATLAB generates an error. (See Naming Variables in the MATLAB documentation.) For example, in the following code f(5) could refer either to the fifth element of an array named f, or to a function named f with an argument of 5. If f is not clearly defined as a variable in the code, MATLAB looks for the function f on the path when the code runs.
parfor i=1:n ... a = f(5); ... end
Transparency
The body of a parfor-loop must be transparent, meaning that all references to variables must be visible (i.e., they occur in the text of the program). In the following example, because X is not visible as an input variable in the parfor body (only the string 'X' is passed to eval), it does not get transferred to the workers. As a result, MATLAB issues an error at run time:
X = 5; parfor ii = 1:4 eval('X'); end
Similarly, you cannot clear variables from a workers workspace by executing clear inside a parfor statement:
parfor ii= 1:4 <statements...> clear('X') % cannot clear: transparency violation <statements...> end
2-10
Programming Considerations
As a workaround, you can free up most of the memory used by a variable by setting its value to empty, presumably when it is no longer needed in your parfor statement:
parfor ii= 1:4 <statements...> X = []; <statements...> end
Examples of some other functions that violate transparency are evalc, evalin, and assignin with the workspace argument specified as 'caller'; save and load, unless the output of load is assigned. MATLAB does successfully execute eval and evalc statements that appear in functions called from the parfor body.
A corresponding parfor-loop does not allow B to reference a function handle. So you can work around the problem with feval:
B = @sin; parfor ii = 1:100 A(ii) = feval(B, ii); end
2-11
Nondistributable Functions
If you use a function that is not strictly computational in nature (e.g., input, plot, keyboard) in a parfor-loop or in any function called by a parfor-loop, the behavior of that function occurs on the worker. The results might include hanging the worker process or having no visible effect at all.
Nested Functions
The body of a parfor-loop cannot make reference to a nested function. However, it can call a nested function by means of a function handle.
Nested Loops
The body of a parfor-loop cannot contain another parfor-loop. But it can call a function that contains another parfor-loop. However, because a worker cannot open a MATLAB pool, a worker cannot run the inner nested parfor-loop in parallel. This means that only one level of nested parfor-loops can run in parallel. If the outer loop runs in parallel on a MATLAB pool, the inner loop runs serially on each worker. If the outer loop runs serially in the client (e.g., parfor specifying zero workers), the function that contains the inner loop can run the inner loop in parallel on workers in a pool. The body of a parfor-loop can contain for-loops. You can use the inner loop variable for indexing the sliced array, but only if you use the variable in plain form, not part of an expression. For example:
A = zeros(4,5); parfor j = 1:4 for k = 1:5 A(j,k) = j + k; end end A
2-12
Programming Considerations
Limitations of Nested for-Loops. For proper variable classification, the range of a for-loop nested in a parfor must be defined by constant numbers or variables. In the following example, the code on the left does not work because the for-loop upper limit is defined by a function call. The code on the right works around this by defining a broadcast or constant variable outside the parfor first:
A = zeros(100, 200); parfor i = 1:size(A, 1) for j = 1:size(A, 2) A(i, j) = plus(i, j); end end
A = zeros(100, 200); n = size(A, 2); parfor i = 1:size(A,1) for j = 1:n A(i, j) = plus(i, j); end end
When using the nested for-loop variable for indexing the sliced array, you must use the variable in plain form, not as part of an expression. For example, the following code on the left does not work, but the code on the right does:
If you use a nested for-loop to index into a sliced array, you cannot use that array elsewhere in the parfor-loop. For example, in the following example, the code on the left does not work because A is sliced and indexed inside the nested for-loop; the code on the right works because v is assigned to A outside the nested loop:
2-13
A = zeros(4, 10); parfor i = 1:4 for j = 1:10 A(i, j) = i + j; end disp(A(i, 1)) end
A = zeros(4, 10); parfor i = 1:4 v = zeros(1, 10); for j = 1:10 v(j) = i + j; end disp(v(1)) A(i, :) = v; end
Inside a parfor, if you use multiple for-loops (not nested inside each other) to index into a single sliced array, they must loop over the same range of values. In the following example, the code on the left does not work because j and k loop over different values; the code on the right works to index different portions of the sliced array A:
A = zeros(4, 10); parfor i = 1:4 for j = 1:5 A(i, j) = i + j; end for k = 6:10 A(i, k) = pi; end end
A = zeros(4, 10); parfor i = 1:4 for j = 1:10 if j < 6 A(i, j) = i + j; else A(i, j) = pi; end end end
2-14
Programming Considerations
Handle Classes
Changes made to handle classes on the workers during loop iterations are not automatically propagated to the client.
P-Code Scripts
You can call P-code script files from within a parfor-loop, but P-code script cannot contain a parfor-loop.
Performance Considerations
Slicing Arrays
If a variable is initialized before a parfor-loop, then used inside the parfor-loop, it has to be passed to each MATLAB worker evaluating the loop iterations. Only those variables used inside the loop are passed from the client workspace. However, if all occurrences of the variable are indexed by the loop variable, each worker receives only the part of the array it needs. For more information, see Where to Create Arrays on page 2-32.
2-15
Current Syntax
for i = drange(range) loop body . . end parfor i = range loop body . . end
Not Implemented
2-16
Advanced Topics
Advanced Topics
In this section... About Programming Notes on page 2-17 Classification of Variables on page 2-17 Improving Performance on page 2-32
Classification of Variables
Overview on page 2-17 Loop Variable on page 2-18 Sliced Variables on page 2-19 Broadcast Variables on page 2-23 Reduction Variables on page 2-23 Temporary Variables on page 2-30
Overview
When a name in a parfor-loop is recognized as referring to a variable, it is classified into one of the following categories. A parfor-loop generates an
2-17
error if it contains any variables that cannot be uniquely categorized or if any variables violate their category restrictions. Classification Loop Sliced Broadcast Reduction Temporary Description Serves as a loop index for arrays An array whose segments are operated on by different iterations of the loop A variable defined before the loop whose value is used inside the loop, but never assigned inside the loop Accumulates a value across iterations of the loop, regardless of iteration order Variable created inside the loop, but unlike sliced or reduction variables, not available outside the loop
Loop Variable
The following restriction is required, because changing i in the parfor body invalidates the assumptions MATLAB makes about communication between the client and workers.
2-18
Advanced Topics
Required (static): Assignments to the loop variable are not allowed. This example attempts to modify the value of the loop variable i in the body of the loop, and thus is invalid:
parfor i = 1:n i = i + 1; a(i) = i; end
Sliced Variables
A sliced variable is one whose value can be broken up into segments, or slices, which are then operated on separately by workers and by the MATLAB client. Each iteration of the loop works on a different slice of the array. Using sliced variables is important because this type of variable can reduce communication between the client and workers. Only those slices needed by a worker are sent to it, and only when it starts working on a particular range of indices. In the next example, a slice of A consists of a single element of that array:
parfor i = 1:length(A) B(i) = f(A(i)); end
Characteristics of a Sliced Variable. A variable in a parfor-loop is sliced if it has all of the following characteristics. A description of each characteristic follows the list: Type of First-Level Indexing The first level of indexing is either parentheses, (), or braces, {}. Fixed Index Listing Within the first-level parenthesis or braces, the list of indices is the same for all occurrences of a given variable. Form of Indexing Within the list of indices for the variable, exactly one index involves the loop variable. Shape of Array In assigning to a sliced variable, the right-hand side of the assignment is not [] or '' (these operators indicate deletion of elements).
2-19
Type of First-Level Indexing. For a sliced variable, the first level of indexing is enclosed in either parentheses, (), or braces, {}. This table lists the forms for the first level of indexing for arrays sliced and not sliced. Reference for Variable Not Sliced
A.x A.(...)
After the first level, you can use any type of valid MATLAB indexing in the second and further levels. The variable A shown here on the left is not sliced; that shown on the right is sliced:
A.q{i,12} A{i,12}.q
Fixed Index Listing. Within the first-level parentheses or braces of a sliced variables indexing, the list of indices is the same for all occurrences of a given variable. The variable A shown here on the left is not sliced because A is indexed by i and i+1 in different places; that shown on the right is sliced:
The example above on the right shows some occurrences of a sliced variable with first-level parenthesis indexing and with first-level brace indexing in the same loop. This is acceptable. Form of Indexing. Within the list of indices for a sliced variable, one of these indices is of the form i, i+k, i-k, k+i, or k-i, where i is the loop variable and
2-20
Advanced Topics
k is a constant or a simple (nonindexed) broadcast variable; and every other index is a constant, a simple broadcast variable, colon, or end.
With i as the loop variable, the A variables shown here on the left are not sliced; those on the right are sliced:
When you use other variables along with the loop variable to index an array, you cannot set these variables inside the loop. In effect, such variables are constant over the execution of the entire parfor statement. You cannot combine the loop variable with itself to form an index expression. Shape of Array. A sliced variable must maintain a constant shape. The variable A shown here on either line is not sliced:
A(i,:) = []; A(end + 1) = i;
The reason A is not sliced in either case is because changing the shape of a sliced array would violate assumptions governing communication between the client and workers. Sliced Input and Output Variables. All sliced variables have the characteristics of being input or output. A sliced variable can sometimes have both characteristics. MATLAB transmits sliced input variables from the client to the workers, and sliced output variables from workers back to the client. If a variable is both input and output, it is transmitted in both directions.
2-21
However, if it is clear that in every iteration, every reference to an array element is set before it is used, the variable is not a sliced input variable. In this example, all the elements of A are set, and then only those fixed values are used:
parfor ii = 1:n if someCondition A(ii) = 32; else A(ii) = 17; end loop code that uses A(ii) end
Even if a sliced variable is not explicitly referenced as an input, implicit usage might make it so. In the following example, not all elements of A are necessarily set inside the parfor-loop, so the original values of the array are received, held, and then returned from the loop, making A both a sliced input and output variable.
A = 1:10; parfor ii = 1:10 if rand < 0.5 A(ii) = 0; end end
2-22
Advanced Topics
Broadcast Variables
A broadcast variable is any variable other than the loop variable or a sliced variable that is not affected by an assignment inside the loop. At the start of a parfor-loop, the values of any broadcast variables are sent to all workers. Although this type of variable can be useful or even essential, broadcast variables that are large can cause a lot of communication between client and workers. In some cases it might be more efficient to use temporary variables for this purpose, creating and assigning them inside the loop.
Reduction Variables
MATLAB supports an important exception, called reductions, to the rule that loop iterations must be independent. A reduction variable accumulates a value that depends on all the iterations together, but is independent of the iteration order. MATLAB allows reduction variables in parfor-loops. Reduction variables appear on both side of an assignment statement, such as any of the following, where expr is a MATLAB expression.
X = X + expr X = X - expr X = expr + X
See Associativity in Reduction Assignments in Further Considerations with Reduction Variables on page 2-25
X = expr .* X X = expr * X X = expr & X X = expr | X X = [expr, X] X = [expr; X] X = {expr, X} X = {expr; X} X = min(expr, X) X = max(expr, X)
X = X .* expr X = X * expr X = X & expr X = X | expr X = [X, expr] X = [X; expr] X = {X, expr} X = {X; expr} X = min(X, expr) X = max(X, expr)
2-23
X = union(expr, X) X = intersect(expr, X)
Each of the allowed statements listed in this table is referred to as a reduction assignment, and, by definition, a reduction variable can appear only in assignments of this type. The following example shows a typical usage of a reduction variable X:
X = ...; parfor i = 1:n X = X + d(i); end % Do some initialization of X
This loop is equivalent to the following, where each d(i) is calculated by a different iteration:
X = X + d(1) + ... + d(n)
If the loop were a regular for-loop, the variable X in each iteration would get its value either before entering the loop or from the previous iteration of the loop. However, this concept does not apply to parfor-loops: In a parfor-loop, the value of X is never transmitted from client to workers or from worker to worker. Rather, additions of d(i) are done in each worker, with i ranging over the subset of 1:n being performed on that worker. The results are then transmitted back to the client, which adds the workers partial sums into X. Thus, workers do some of the additions, and the client does the rest. Basic Rules for Reduction Variables. The following requirements further define the reduction assignments associated with a given variable. Required (static): For any reduction variable, the same reduction function or operation must be used in all reduction assignments for that variable. The parfor-loop on the left is not valid because the reduction assignment uses + in one instance, and [,] in another. The parfor-loop on the right is valid:
2-24
Advanced Topics
parfor i = 1:n if testLevel(k) A = A + i; else A = [A, 4+i]; end % loop body continued end
parfor i = 1:n if testLevel(k) A = A + i; else A = A + i + 5*k; end % loop body continued end
Required (static): If the reduction assignment uses * or [,], then in every reduction assignment for X, X must be consistently specified as the first argument or consistently specified as the second. The parfor-loop on the left below is not valid because the order of items in the concatenation is not consistent throughout the loop. The parfor-loop on the right is valid:
parfor i = 1:n if testLevel(k) A = [A, 4+i]; else A = [r(i), A]; end % loop body continued end
parfor i = 1:n if testLevel(k) A = [A, 4+i]; else A = [A, r(i)]; end % loop body continued end
Further Considerations with Reduction Variables. This section provide more detail about reduction assignments, associativity, commutativity, and overloading of reduction functions. Reduction Assignments. In addition to the specific forms of reduction assignment listed in the table in Reduction Variables on page 2-23, the only other (and more general) form of a reduction assignment is
X = f(X, expr) X = f(expr, X)
2-25
Required (static): f can be a function or a variable. If it is a variable, it must not be affected by the parfor body (in other words, it is a broadcast variable). If f is a variable, then for all practical purposes its value at run time is a function handle. However, this is not strictly required; as long as the right-hand side can be evaluated, the resulting value is stored in X. The parfor-loop below on the left will not execute correctly because the statement f = @times causes f to be classified as a temporary variable and therefore is cleared at the beginning of each iteration. The parfor on the right is correct, because it does not assign to f inside the loop:
f = @(x,k)x * k; parfor i = 1:n a = f(a,i); % loop body continued f = @times; % Affects f end
Note that the operators && and || are not listed in the table in Reduction Variables on page 2-23. Except for && and ||, all the matrix operations of MATLAB have a corresponding function f, such that u op v is equivalent to f(u,v). For && and ||, such a function cannot be written because u&&v and u||v might or might not evaluate v, but f(u,v) always evaluates v before calling f. This is why && and || are excluded from the table of allowed reduction assignments for a parfor-loop. Every reduction assignment has an associated function f. The properties of f that ensure deterministic behavior of a parfor statement are discussed in the following sections. Associativity in Reduction Assignments. Concerning the function f as used in the definition of a reduction variable, the following practice is recommended, but does not generate an error if not adhered to. Therefore, it is up to you to ensure that your code meets this recommendation.
2-26
Advanced Topics
Recommended: To get deterministic behavior of parfor-loops, the reduction function f must be associative. To be associative, the function f must satisfy the following for all a, b, and c:
f(a,f(b,c)) = f(f(a,b),c)
The classification rules for variables, including reduction variables, are purely syntactic. They cannot determine whether the f you have supplied is truly associative or not. Associativity is assumed, but if you violate this, different executions of the loop might result in different answers. Note While the addition of mathematical real numbers is associative, addition of floating-point numbers is only approximately associative, and different executions of this parfor statement might produce values of X with different round-off errors. This is an unavoidable cost of parallelism. For example, the statement on the left yields 1, while the statement on the right returns 1 + eps:
(1 + eps/2) + eps/2 1 + (eps/2 + eps/2)
With the exception of the minus operator (-), all the special cases listed in the table in Reduction Variables on page 2-23 have a corresponding (perhaps approximately) associative function. MATLAB calculates the assignment X = X - expr by using X = X + (-expr). (So, technically, the function for calculating this reduction assignment is plus, not minus.) However, the assignment X = expr - X cannot be written using an associative function, which explains its exclusion from the table. Commutativity in Reduction Assignments. Some associative functions, including +, .*, min, and max, intersect, and union, are also commutative. That is, they satisfy the following for all a and b:
f(a,b) = f(b,a)
Examples of noncommutative functions are * (because matrix multiplication is not commutative for matrices in which both dimensions have size greater than one), [,], [;], {,}, and {;}. Noncommutativity is the reason that consistency
2-27
in the order of arguments to these functions is required. As a practical matter, a more efficient algorithm is possible when a function is commutative as well as associative, and parfor is optimized to exploit commutativity. Recommended: Except in the cases of *, [,], [;], {,}, and {;}, the function f of a reduction assignment should be commutative. If f is not commutative, different executions of the loop might result in different answers. Unless f is a known noncommutative built-in, it is assumed to be commutative. There is currently no way to specify a user-defined, noncommutative function in parfor. Overloading in Reduction Assignments. Most associative functions f have an identity element e, so that for any a, the following holds true:
f(e,a) = a = f(a,e)
Examples of identity elements for some functions are listed in this table. Function
+ * and .* min max [,], [;], and union
Identity Element
0 1 Inf -Inf []
MATLAB uses the identity elements of reduction functions when it knows them. So, in addition to associativity and commutativity, you should also keep identity elements in mind when overloading these functions. Recommended: An overload of +, *, .*, min, max, union, [,], or [;] should be associative if it is used in a reduction assignment in a parfor. The overload must treat the respective identity element given above (all with class double) as an identity element.
2-28
Advanced Topics
Recommended: An overload of +, .*, min, max, union, or intersect should be commutative. There is no way to specify the identity element for a function. In these cases, the behavior of parfor is a little less efficient than it is for functions with a known identity element, but the results are correct. Similarly, because of the special treatment of X = X - expr, the following is recommended. Recommended: An overload of the minus operator (-) should obey the mathematical law that X - (y + z) is equivalent to (X - y) - z. Example: Using a Custom Reduction Function. Suppose each iteration of a loop performs some calculation, and you are interested in finding which iteration of a loop produces the maximum value. This is a reduction exercise that makes an accumulation across multiple iterations of a loop. Your reduction function must compare iteration results, until finally the maximum value can be determined after all iterations are compared. First consider the reduction function itself. To compare an iterations result against anothers, the function requires as input the current iterations result and the known maximum result from other iterations so far. Each of the two inputs is a vector containing an iterations result data and iteration number.
function mc = comparemax(A, B) % Custom reduction function for 2-element vector input if A(1) >= B(1) % Compare the two input data values mc = A; % Return the vector with the larger result else mc = B; end
Inside the loop, each iteration calls the reduction function (comparemax), passing in a pair of 2-element vectors: The accumulated maximum and its iteration index (this is the reduction variable, cummax)
2-29
The iterations own calculation value and index If the data value of the current iteration is greater than the maximum in cummmax, the function returns a vector of the new value and its iteration number. Otherwise, the function returns the existing maximum and its iteration number. The code for the loop looks like the following, with each iteration calling the reduction function comparemax to compare its own data [dat i] to that already accumulated in cummax.
% First element of cummax is maximum data value % Second element of cummax is where (iteration) maximum occurs cummax = [0 0]; % Initialize reduction variable parfor ii = 1:100 dat = rand(); % Simulate some actual computation cummax = comparemax(cummax, [dat ii]); end disp(cummax);
Temporary Variables
A temporary variable is any variable that is the target of a direct, nonindexed assignment, but is not a reduction variable. In the following parfor-loop, a and d are temporary variables:
a = 0; z = 0; r = rand(1,10); parfor i = 1:10 a = i; z = z + i; if i <= 5 d = 2*a; end end
% Variable a is temporary
% Variable d is temporary
In contrast to the behavior of a for-loop, MATLAB effectively clears any temporary variables before each iteration of a parfor-loop. To help ensure the independence of iterations, the values of temporary variables cannot
2-30
Advanced Topics
be passed from one iteration of the loop to another. Therefore, temporary variables must be set inside the body of a parfor-loop, so that their values are defined separately for each iteration. MATLAB does not send temporary variables back to the client. A temporary variable in the context of the parfor statement has no effect on a variable with the same name that exists outside the loop, again in contrast to ordinary for-loops. Uninitialized Temporaries. Because temporary variables are cleared at the beginning of every iteration, MATLAB can detect certain cases in which any iteration through the loop uses the temporary variable before it is set in that iteration. In this case, MATLAB issues a static error rather than a run-time error, because there is little point in allowing execution to proceed if a run-time error is guaranteed to occur. This kind of error often arises because of confusion between for and parfor, especially regarding the rules of classification of variables. For example, suppose you write
b = true; parfor i = 1:n if b && some_condition(i) do_something(i); b = false; end ... end
This loop is acceptable as an ordinary for-loop, but as a parfor-loop, b is a temporary variable because it occurs directly as the target of an assignment inside the loop. Therefore it is cleared at the start of each iteration, so its use in the condition of the if is guaranteed to be uninitialized. (If you change parfor to for, the value of b assumes sequential execution of the loop, so that do_something(i) is executed for only the lower values of i until b is set false.) Temporary Variables Intended as Reduction Variables. Another common cause of uninitialized temporaries can arise when you have a variable that you intended to be a reduction variable, but you use it elsewhere in the loop, causing it technically to be classified as a temporary variable. For example:
2-31
If the only occurrences of s were the two in the first statement of the body, it would be classified as a reduction variable. But in this example, s is not a reduction variable because it has a use outside of reduction assignments in the line s > whatever. Because s is the target of an assignment (in the first statement), it is a temporary, so MATLAB issues an error about this fact, but points out the possible connection with reduction. Note that if you change parfor to for, the use of s outside the reduction assignment relies on the iterations being performed in a particular order. The point here is that in a parfor-loop, it matters that the loop does not care about the value of a reduction variable as it goes along. It is only after the loop that the reduction value becomes usable.
Improving Performance
Where to Create Arrays
With a parfor-loop, it might be faster to have each MATLAB worker create its own arrays or portions of them in parallel, rather than to create a large array in the client before the loop and send it out to all the workers separately. Having each worker create its own copy of these arrays inside the loop saves the time of transferring the data from client to workers, because all the workers can be creating it at the same time. This might challenge your usual practice to do as much variable initialization before a for-loop as possible, so that you do not needlessly repeat it inside the loop. Whether to create arrays before the parfor-loop or inside the parfor-loop depends on the size of the arrays, the time needed to create them, whether the workers need all or part of the arrays, the number of loop iterations that each worker performs, and other factors. While many for-loops can be
2-32
Advanced Topics
directly converted to parfor-loops, even in these cases there might be other issues involved in optimizing your code.
Then enter the following examples. (If you are viewing this documentation in the MATLAB help browser, highlight each segment of code below, right-click, and select Evaluate Selection in the context menu to execute the block in MATLAB. That way the time measurement will not include the time required to paste or type.)
tic; n = 200; M = magic(n); R = rand(n); parfor i = 1:n A(i) = sum(M(i,:).*R(n+1-i,:)); end toc
tic; n = 200; parfor i = 1:n M = magic(n); R = rand(n); A(i) = sum(M(i,:).*R(n+1-i,:)); end toc
Running on a remote cluster, you might find different behavior as workers can simultaneously create their arrays, saving transfer time. Therefore, code that is optimized for local workers might not be optimized for cluster workers, and vice versa.
2-33
2-34
3
Single Program Multiple Data (spmd)
Executing Simultaneously on Multiple Data Sets on page 3-2 Accessing Data with Composites on page 3-7 Distributing Arrays on page 3-12 Programming Tips on page 3-15
Introduction
The single program multiple data (spmd) language construct allows seamless interleaving of serial and parallel programming. The spmd statement lets you define a block of code to run simultaneously on multiple labs. Variables assigned inside the spmd statement on the labs allow direct access to their values from the client by reference via Composite objects. This chapter explains some of the characteristics of spmd statements and Composite objects.
3-2
Programs that take a long time to execute spmd lets several labs compute solutions simultaneously. Programs operating on large data sets spmd lets the data be distributed to multiple labs.
This command starts the number of MATLAB worker sessions defined by the default parallel configuration. If the local configuration is your default and does not specify the number of workers, this starts one worker per core (maximum of twelve) on your local MATLAB client machine. If you do not want to use default settings, you can specify in the matlabpool statement which configuration or how many labs to use. For example, to use only three labs with your default configuration, type:
matlabpool 3
This command returns a value indicating the number of labs in the current pool. If the command returns 0, there is currently no pool open.
3-3
Note If there is no MATLAB pool open, an spmd statement runs locally in the MATLAB client without any parallel execution, provided you have Parallel Computing Toolbox software installed. In other words, it runs in your client session as though it were a single lab. When you are finished using a MATLAB pool, close it with the command:
matlabpool close
The block of code represented by <statements> executes in parallel simultaneously on all labs in the MATLAB pool. If you want to limit the execution to only a portion of these labs, specify exactly how many labs to run on:
spmd (n) <statements> end
This statement requires that n labs run the spmd code. n must be less than or equal to the number of labs in the open MATLAB pool. If the pool is large enough, but n labs are not available, the statement waits until enough labs are available. If n is 0, the spmd statement uses no labs, and runs locally on the client, the same as if there were not a pool currently open. You can specify a range for the number of labs:
spmd (m, n) <statements> end
In this case, the spmd statement requires a minimum of m labs, and it uses a maximum of n labs.
3-4
If it is important to control the number of labs that execute your spmd statement, set the exact number in the configuration or with the spmd statement, rather than using a range. For example, create a random matrix on three labs:
matlabpool spmd (3) R = rand(4,4); end matlabpool close
Note All subsequent examples in this chapter assume that a MATLAB pool is open and remains open between sequences of spmd statements. Unlike a parfor-loop, the labs used for an spmd statement each have a unique value for labindex. This lets you specify code to be run on only certain labs, or to customize execution, usually for the purpose of accessing unique data. For example, create different sized arrays depending on labindex:
spmd (3) if labindex==1 R = rand(9,9); else R = rand(4,4); end end
Load unique data on each lab according to labindex, and use the same function on each lab to compute a result from the data:
spmd (3) labdata = load(['datafile_' num2str(labindex) '.ascii']) result = MyFunction(labdata) end
The labs executing an spmd statement operate simultaneously and are aware of each other. As with a parallel job, you are allowed to directly
3-5
control communications between the labs, transfer data between them, and use codistributed arrays among them. For a list of toolbox functions that facilitate these capabilities, see the Function Reference sections Interlab Communication Within a Parallel Job on page 13-9 and Distributed and Codistributed Arrays on page 13-3. For example, use a codistributed array in an spmd statement:
spmd (3) RR = rand(30, codistributor()); end
Each lab has a 30-by-10 segment of the codistributed array RR. For more information about codistributed arrays, see Chapter 5, Math with Codistributed Arrays.
Displaying Output
When running an spmd statement on a MATLAB pool, all command-line output from the workers displays in the client Command Window. Because the workers are MATLAB sessions without displays, any graphical output (for example, figure windows) from the pool does not display at all.
3-6
Introduction
Composite objects in the MATLAB client session let you directly access data values on the labs. Most often you assigned these variables within spmd statements. In their display and usage, Composites resemble cell arrays. There are two ways to create Composites: Using the Composite function on the client. Values assigned to the Composite elements are stored on the labs. Defining variables on labs inside an spmd statement. After the spmd statement, the stored values are accessible on the client as Composites.
end MM{1} % In the client, MM is a Composite with one element per lab 8 1 6
3-7
3 4 MM{2} 16 5 9 4
5 9
7 2
2 11 7 14
3 10 6 15
13 8 12 1
A variable might not be defined on every lab. For the labs on which a variable is not defined, the corresponding Composite element has no value. Trying to read that element throws an error.
spmd if labindex > 1 HH = rand(4); end end HH Lab 1: No data Lab 2: class = double, size = [4 Lab 3: class = double, size = [4 4] 4]
You can also set values of Composite elements from the client. This causes a transfer of data, storing the value on the appropriate lab even though it is not executed within an spmd statement:
MM{3} = eye(4);
In this case, MM must already exist as a Composite, otherwise MATLAB interprets it as a cell array. Now when you do enter an spmd statement, the value of the variable MM on lab 3 is as set:
spmd if labindex == 3, MM, end end Lab 3: MM = 1 0 0 0
3-8
0 0 0
1 0 0
0 1 0
0 0 1
Data transfers from lab to client when you explicitly assign a variable in the client workspace using a Composite element:
M = MM{1} % Transfer data from lab 1 to variable M on the client 8 3 4 1 5 9 6 7 2
Assigning an entire Composite to another Composite does not cause a data transfer. Instead, the client merely duplicates the Composite as a reference to the appropriate data stored on the labs:
NN = MM % Set entire Composite equal to another, without transfer
However, accessing a Composites elements to assign values to other Composites does result in a transfer of data from the labs to the client, even if the assignment then goes to the same lab. In this case, NN must already exist as a Composite:
NN{1} = MM{1} % Transfer data to the client and then to lab
3-9
matlabpool open local 4 spmd AA = end AA(:) % [1] [2] [3] [4] spmd AA = end AA(:) % [2] [4] [6] [8] clear AA labindex; % Initial setting Composite
By default, this creates a Composite with an element for each lab in the MATLAB pool. You can also create Composites on only a subset of the labs in the pool. See the Composite reference page for more details. The elements of the Composite can now be set as usual on the client, or as variables inside an spmd statement. When you set an element of a Composite, the data is immediately transferred to the appropriate lab:
for ii = 1:numel(PP)
3-10
3-11
Distributing Arrays
In this section... Distributed Versus Codistributed Arrays on page 3-12 Creating Distributed Arrays on page 3-12 Creating Codistributed Arrays on page 3-13
3-12
Distributing Arrays
Use any of the overloaded distributed object methods to directly construct a distributed array on the labs. This technique does not require that the array already exists in the client, thereby reducing client workspace memory requirements. These overloaded functions include distributed.eye, distributed.rand, etc. For a full list, see the distributed object reference page. Create a codistributed array inside an spmd statement, then access it as a distributed array outside the spmd statement. This lets you use distribution schemes other than the default. The first two of these techniques do not involve spmd in creating the array, but you can see how spmd might be used to manipulate arrays created this way. For example: Create an array in the client workspace, then make it a distributed array:
matlabpool open local 2 W = ones(6,6); W = distributed(W); % Distribute to the labs spmd T = W*2; % Calculation performed on labs, in parallel. % T and W are both codistributed arrays here. end T % View results in client. whos % T and W are both distributed arrays here. matlabpool close
3-13
Create a distributed array outside an spmd statement, then access it as a codistributed array inside the spmd statement running on the same MATLAB pool. In this example, you create a codistributed array inside an spmd statement, using a nondefault distribution scheme. First, define 1-D distribution along the third dimension, with 4 parts on lab 1, and 12 parts on lab 2. Then create a 3-by-3-by-16 array of zeros.
matlabpool open local 2 spmd codist = codistributor1d(3, [4, 12]); Z = codistributed.zeros(3, 3, 16, codist); Z = Z + labindex; end Z % View results in client. % Z is a distributed array here. matlabpool close
For more details on codistributed arrays, see Chapter 5, Math with Codistributed Arrays, and Chapter 4, Interactive Parallel Computation with pmode.
3-14
Programming Tips
Programming Tips
In this section... MATLAB Path on page 3-15 Error Handling on page 3-15 Limitations on page 3-15
MATLAB Path
All labs executing an spmd statement must have the same MATLAB path configuration as the client, so that they can execute any functions called in their common block of code. Therefore, whenever you use cd, addpath, or rmpath on the client, it also executes on all the labs, if possible. For more information, see the matlabpool reference page. When the labs are running on a different platform than the client, use the function pctRunOnAll to properly set the MATLAB path on all labs.
Error Handling
When an error occurs on a lab during the execution of an spmd statement, the error is reported to the client. The client tries to interrupt execution on all labs, and throws an error to the user. Errors and warnings produced on labs are annotated with the lab ID and displayed in the clients Command Window in the order in which they are received by the MATLAB client. The behavior of lastwarn is unspecified at the end of an spmd if used within its body.
Limitations
Transparency
The body of an spmd statement must be transparent, meaning that all references to variables must be visible (i.e., they occur in the text of the program).
3-15
In the following example, because X is not visible as an input variable in the spmd body (only the string 'X' is passed to eval), it does not get transferred to the labs. As a result, MATLAB issues an error at run time:
X = 5; spmd eval('X'); end
Similarly, you cannot clear variables from a workers workspace by executing clear inside an spmd statement:
spmd; clear('X'); end
To clear a specific variable from a worker, clear its Composite from the client workspace. Alternatively, you can free up most of the memory used by a variable by setting its value to empty, presumably when it is no longer needed in your spmd statement:
spmd <statements....> X = []; end
Examples of some other functions that violate transparency are evalc, evalin, and assignin with the workspace argument specified as 'caller'; save and load, unless the output of load is assigned. MATLAB does successfully execute eval and evalc statements that appear in functions called from the spmd body.
Nested Functions
Inside a function, the body of an spmd statement cannot make any direct reference to a nested function. However, it can call a nested function by means of a variable defined as a function handle to the nested function. Because the spmd body executes on workers, variables that are updated by nested functions called inside an spmd statement do not get updated in the workspace of the outer function.
3-16
Programming Tips
Anonymous Functions
The body of an spmd statement cannot define an anonymous function. However, it can reference an anonymous function by means of a function handle.
Nested parfor-Loops
The body of a parfor-loop cannot contain an spmd statement, and an spmd statement cannot contain a parfor-loop.
3-17
3-18
4
Interactive Parallel Computation with pmode
This chapter describes interactive pmode in the following sections: pmode Versus spmd on page 4-2 Run Parallel Jobs Interactively Using pmode on page 4-3 Parallel Command Window on page 4-10 Running pmode Interactive Jobs on a Cluster on page 4-15 Plotting Distributed Data Using pmode on page 4-16 pmode Limitations and Unexpected Results on page 4-18 pmode Troubleshooting on page 4-19
4-2
This starts four local labs, creates a parallel job to run on those labs, and opens the Parallel Command Window.
You can control where the command history appears. For this exercise, the position is set by clicking Window > History Position > Above Prompt, but you can set it according to your own preference.
2 To illustrate that commands at the pmode prompt are executed on all labs,
4-3
3 Set a variable at the pmode prompt. Notice that the value is set on all
the labs.
P>> x = pi
4 A variable does not necessarily have the same value on every lab. The
labindex function returns the ID particular to each lab working on this parallel job. In this example, the variable x exists with a different value in
4-4
P>> segment = [1 2; 3 4; 5 6]
4-5
7 Assign a unique value to the array on each lab, dependent on the lab
8 Until this point in the example, the variant arrays are independent, other
than having the same name. Use the codistributed.build function to aggregate the array segments into a coherent array, distributed among the labs.
P>> codist = codistributor1d(2, [2 2 2 2], [3 8]) P>> whole = codistributed.build(segment, codist)
This combines four separate 3-by-2 arrays into one 3-by-8 codistributed array. The codistributor1d object indicates that the array is distributed along its second dimension (columns), with 2 columns on each of the four labs. On each lab, segment provided the data for the local portion of the whole array.
9 Now, when you operate on the codistributed array whole, each lab handles
the calculations on only its portion, or segment, of the array, not the whole array.
P>> whole = whole + 1000
4-6
10 Although the codistributed array allows for operations on its entirety, you
can use the getLocalPart function to access the portion of a codistributed array on a particular lab.
P>> section = getLocalPart(whole)
11 If you need the entire array in one workspace, use the gather function.
Notice, however, that this gathers the entire array into the workspaces of all the labs. See the gather reference page for the syntax to gather the array into the workspace of only one lab.
12 Because the labs ordinarily do not have displays, if you want to perform
any graphical tasks involving your data, such as plotting, you must do this from the client workspace. Copy the array to the client workspace by typing the following commands in the MATLAB (client) Command Window.
pmode lab2client combined 1
4-7
arrays. For example, the eye function creates an identity matrix. Now you can create a codistributed identity matrix with the following commands in the Parallel Command Window.
P>> distobj = codistributor1d(); P>> I = eye(6, distobj) P>> getLocalPart(I)
Calling the codistributor1d function without arguments specifies the default distribution, which is by columns in this case, distributed as evenly as possible.
4-8
the redistribute function. In this example, the argument 1 to codistributor1d specifies distribution of the array along the first dimension (rows).
P>> distobj = codistributor1d(1); P>> I = redistribute(I, distobj) P>> getLocalPart(I)
4-9
four labs start on your local machine and a parallel job is created to run on them. The first time you run pmode with this configuration, you get a tiled display of the four labs.
4-10
The Parallel Command Window offers much of the same functionality as the MATLAB desktop, including command line, output, and command history. When you select one or more lines in the command history and right-click, you see the following context menu.
You have several options for how to arrange the tiles showing your lab outputs. Usually, you will choose an arrangement that depends on the format of your data. For example, the data displayed until this point in this section, as in the previous figure, is distributed by columns. It might be convenient to arrange the tiles side by side.
4-11
This arrangement results in the following figure, which might be more convenient for viewing data distributed by columns.
Alternatively, if the data is distributed by rows, you might want to stack the lab tiles vertically. For the following figure, the data is reformatted with the command
P>> distobj = codistributor('1d',1); P>> I = redistribute(I, distobj)
4-12
You can control the relative positions of the command window and the lab output. The following figure shows how to set the output to display beside the input, rather than above it.
2. Select tab
4-13
You can have multiple labs send their output to the same tile or tab. This allows you to have fewer tiles or tabs than labs.
In this case, the window provides shading to help distinguish the outputs from the various labs.
4-14
carefully locate your scheduler on that page and note that pmode sessions run as jobs described as parallel applications that use inter-worker communication. Many of the jobs properties are determined by a configuration. For more details about creating and using configurations, see Parallel Configurations for Cluster Access on page 6-14. The general form of the command to start a pmode session is
pmode start <config-name> <num-labs>
where <config-name> is the name of the configuration you want to use, and <num-labs> is the number of labs you want to run the pmode job on. If <num-labs> is omitted, the number of labs is determined by the configuration. Coordinate with your system administrator when creating or using a configuration. If you omit <config-name>, pmode uses the default configuration (see the defaultParallelConfig reference page). For details on all the command options, see the pmode reference page.
4-15
one lab.
2 Transfer the whole array from any lab to the MATLAB client with pmode
lab2client.
3 Plot the data from the client workspace.
The following example illustrates this technique. Create a 1-by-100 codistributed array of 0s. With four labs, each lab has a 1-by-25 segment of the whole array.
P>> D = zeros(1,100,codistributor1d()) Lab Lab Lab Lab 1: 2: 3: 4: This This This This lab lab lab lab stores stores stores stores D(1:25). D(26:50). D(51:75). D(76:100).
Use a for-loop over the distributed range to populate the array so that it contains a sine wave. Each lab does one-fourth of the array.
P>> for i = drange(1:100) D(i) = sin(i*2*pi/100); end;
Gather the array so that the whole array is contained in the workspace of lab 1.
P>> P = gather(D, 1);
4-16
Transfer the array from the workspace of lab 1 to the MATLAB client workspace, then plot the array from the client. Note that both commands are entered in the MATLAB (client) Command Window.
pmode lab2client P 1 plot(P)
This is not the only way to plot codistributed data. One alternative method, especially useful when running noninteractive parallel jobs, is to plot the data to a file, then view it from a later MATLAB session.
4-17
4-18
pmode Troubleshooting
pmode Troubleshooting
In this section... Connectivity Testing on page 4-19 Hostname Resolution on page 4-19 Socket Connections on page 4-19
Connectivity Testing
For testing connectivity between the client machine and the machines of your compute cluster, you can use Admin Center. For more information about Admin Center, including how to start it and how to test connectivity, see Admin Center in the MATLAB Distributed Computing Server documentation.
Hostname Resolution
If a lab cannot resolve the hostname of the computer running the MATLAB client, use pctconfig to change the hostname by which the client machine advertises itself.
Socket Connections
If a lab cannot open a socket connection to the MATLAB client, try the following: Use pctconfig to change the hostname by which the client machine advertises itself. Make sure that firewalls are not preventing communication between the lab and client machines. Use pctconfig to change the clients pmodeport property. This determines the port that the labs will use to contact the client in the next pmode session.
4-19
4-20
5
Math with Codistributed Arrays
This chapter describes the distribution or partition of data across several labs, and the functionality provided for operations on that data in spmd statements, parallel jobs, and pmode. The sections are as follows. Nondistributed Versus Distributed Arrays on page 5-2 Working with Codistributed Arrays on page 5-5 Using a for-Loop Over a Distributed Range (for-drange) on page 5-21 Using MATLAB Functions on Codistributed Arrays on page 5-25
Introduction
All built-in data types and data structures supported by MATLAB software are also supported in the MATLAB parallel computing environment. This includes arrays of any number of dimensions containing numeric, character, logical values, cells, or structures; but not function handles or user-defined objects. In addition to these basic building blocks, the MATLAB parallel computing environment also offers different types of arrays.
Nondistributed Arrays
When you create a nondistributed array, MATLAB constructs a separate array in the workspace of each lab and assigns a common variable to them. Any operation performed on that variable affects all individual arrays assigned to it. If you display from lab 1 the value assigned to this variable, all labs respond by showing the array of that name that resides in their workspace. The state of a nondistributed array depends on the value of that array in the workspace of each lab: Replicated Arrays on page 5-2 Variant Arrays on page 5-3 Private Arrays on page 5-4
Replicated Arrays
A replicated array resides in the workspaces of all labs, and its size and content are identical on all labs. When you create the array, MATLAB assigns it to the same variable on all labs. If you display in spmd the value assigned to this variable, all labs respond by showing the same array.
5-2
Variant Arrays
A variant array also resides in the workspaces of all labs, but its content differs on one or more labs. When you create the array, MATLAB assigns a different value to the same variable on all labs. If you display the value assigned to this variable, all labs respond by showing their version of the array.
spmd, A = magic(3) + labindex - 1, end LAB 1 8 3 4 1 5 9 6 7 2 | | | | 9 4 5 LAB 2 2 6 10 7 9 3 | | 10 | 5 | 6 LAB 3 3 7 11 8 9 4 | | 11 | 6 | 7 LAB 4 4 8 12 9 10 5
A replicated array can become a variant array when its value becomes unique on each lab.
spmd B = magic(3); B = B + labindex; end %replicated on all labs %now a variant array, different on each lab
5-3
Private Arrays
A private array is defined on one or more, but not all labs. You could create this array by using the lab index in a conditional statement, as shown here:
spmd if labindex >= 3, A = magic(3) + labindex - 1, end end LAB 1 LAB 2 LAB 3 LAB 4 | | | A is | A is | 10 3 8 | 11 4 9 undefined | undefined | 5 7 9 | 6 8 10 | 6 11 4 | 7 12 5
Codistributed Arrays
With replicated and variant arrays, the full content of the array is stored in the workspace of each lab. Codistributed arrays, on the other hand, are partitioned into segments, with each segment residing in the workspace of a different lab. Each lab has its own array segment to work with. Reducing the size of the array that each lab has to store and process means a more efficient use of memory and faster processing, especially for large data sets. This example distributes a 3-by-10 replicated array A over four labs. The resulting array D is also 3-by-10 in size, but only a segment of the full array resides on each lab.
spmd A = [11:20; 21:30; 31:40]; D = codistributed(A); getLocalPart(D) end LAB 1 11 21 31 12 22 32 13 23 33 | | | | 14 24 34 LAB 2 15 25 35 16 26 36 | | | | LAB 3 17 27 37 18 28 38 | | | | LAB 4 19 29 39 20 30 40
For more details on using codistributed arrays, see Working with Codistributed Arrays on page 5-5.
5-4
5-5
spmd A = zeros(80, 1000); D = codistributed(A); Lab 1: This lab stores Lab 2: This lab stores Lab 3: This lab stores Lab 4: This lab stores end
Each lab has access to all segments of the array. Access to the local segment is faster than to a remote segment, because the latter requires sending and receiving data between labs and thus takes more time.
5-6
To see the actual data in the local segment of the array, use the getLocalPart function.
5-7
18 28 38 48
The next line uses the codistributed function to construct a single 4-by-8 matrix D that is distributed along the second dimension of the array:
spmd D = codistributed(A); getLocalPart(D) end 1: Local Part 11 12 21 22 31 32 41 42 | 2: Local Part | 13 14 | 23 24 | 33 34 | 43 44 | 3: Local Part | 15 16 | 25 26 | 35 36 | 45 46 | 4: Local Part | 17 18 | 27 28 | 37 38 | 47 48
Arrays A and D are the same size (4-by-8). Array A exists in its full size on each lab, while only a segment of array D exists on each lab.
spmd, size(A), size(D), end
Examining the variables in the client workspace, an array that is codistributed among the labs inside an spmd statement, is a distributed array from the perspective of the client outside the spmd statement. Variables that are not codistributed inside the spmd, are Composites in the client outside the spmd.
5-8
whos Name A D
See the codistributed function reference page for syntax and usage information.
Now combine these segments into an array that is distributed by the first dimension (rows). The array is now 16-by-250, with a 4-by-250 segment residing on each lab:
spmd D = codistributed.build(A, codistributor1d(1,[4 4 4 4],[16 250])) end Lab 1: This lab stores D(1:4,:).
5-9
You could also use replicated arrays in the same fashion, if you wanted to create a codistributed array whose segments were all identical to start with. See the codistributed function reference page for syntax and usage information.
5-10
codistributed.sprand(m, n, density, codist) codistributed.sprandn(m, n, density, codist) codistributed.true(m, n, ..., codist) codistributed.zeros(m, n, ..., classname, codist)
Local Arrays
That part of a codistributed array that resides on each lab is a piece of a larger array. Each lab can work on its own segment of the common array, or it can make a copy of that segment in a variant or private array of its own. This local copy of a codistributed array segment is called a local array.
Each lab recognizes that the codistributed array D is 3-by-80. However, notice that the size of the local part, L, is 3-by-20 on each lab, because the 80 columns of D are distributed over four labs.
5-11
the local variant arrays into a single array distributed along the specified dimension. Continuing the previous example, take the local variant arrays L and put them together as segments to build a new codistributed array X.
spmd codist = codistributor1d(2,[20 20 20 20],[3 80]); X = codistributed.build(L, codist); size(X) end
5-12
The Dimension value of 2 means the array X is distributed by columns (dimension 2); and the Partition value of [20 20 20 20] means that twenty columns reside on each of the four labs. To get these properties programmatically, return the output of
getCodistributor to a variable, then use dot notation to access each
property:
spmd C = getCodistributor(X); part = C.Partition dim = C.Dimension end
5-13
size(getLocalPart(D)) end
Create a new codistributed array distributed by rows from an existing one already distributed by columns:
spmd X = redistribute(D, codistributor1d(1)); size(getLocalPart(X)) end
18 28 38 48
19 29 39 49
20 30 40 50
5-14
end
Restore the undistributed segments to the full array form by gathering the segments:
spmd, X = gather(D), X = 11 12 13 21 22 23 31 32 33 41 42 43 spmd, 4 size(X), 10 end end 14 24 34 44 15 25 35 45 16 26 36 46 17 27 37 47 18 28 38 48 19 29 39 49 20 30 40 50
5-15
The MATLAB colon operator and end keyword are two of the basic tools for indexing into nondistributed arrays. For codistributed arrays, MATLAB provides a version of the colon operator, called codistributed.colon. This actually is a function, not a symbolic operator like colon. Note When using arrays to index into codistributed arrays, you can use only replicated or codistributed arrays for indexing. The toolbox does not check to ensure that the index is replicated, as that would require global communications. Therefore, the use of unsupported variants (such as labindex) to index into codistributed arrays might create unexpected results.
If you run this code on a pool of four workers you get this result:
Lab 1: Element is in position 225000 on lab 1.
If you run this code on a pool of five workers you get this result:
Lab 2: Element is in position 25000 on lab 2.
5-16
Notice if you use a pool of a different size, the element ends up in a different location on a different lab, but the same code can be used to locate the element.
2-Dimensional Distribution
As an alternative to distributing by a single dimension of rows or columns, you can distribute a matrix by blocks using '2dbc' or two-dimensional block-cyclic distribution. Instead of segments that comprise a number of complete rows or columns of the matrix, the segments of the codistributed array are 2-dimensional square blocks. For example, consider a simple 8-by-8 matrix with ascending element values. You can create this array in an spmd statement, parallel job, or pmode. This example uses pmode for a visual display.
P>> A = reshape(1:64, 8, 8)
Suppose you want to distribute this array among four labs, with a 4-by-4 block as the local part on each lab. In this case, the lab grid is a 2-by-2 arrangement of the labs, and the block size is a square of four elements on a side (i.e., each block is a 4-by-4 square). With this information, you can define the codistributor object:
5-17
Now you can use this codistributor object to distribute the original matrix:
P>> AA = codistributed(A, DIST)
This distributes the array among the labs according to this scheme:
LAB 1 LAB 2
1 2 3 4 5 6 7 8
9 10 11 12 13 14 15 16
17 18 19 20 21 22 23 24
LAB 3
25 26 27 28 29 30 31 32
33 34 35 36 37 38 39 40
41 42 43 44 45 46 47 48
LAB 4
49 50 51 52 53 54 55 56
57 58 59 60 61 62 63 64
If the lab grid does not perfectly overlay the dimensions of the codistributed array, you can still use '2dbc' distribution, which is block cyclic. In this case, you can imagine the lab grid being repeatedly overlaid in both dimensions until all the original matrix elements are included. Using the same original 8-by-8 matrix and 2-by-2 lab grid, consider a block size of 3 instead of 4, so that 3-by-3 square blocks are distributed among the labs. The code looks like this:
P>> DIST = codistributor2dbc([2 2], 3) P>> AA = codistributed(A, DIST)
The first row of the lab grid is distributed to lab 1 and lab 2, but that contains only six of the eight columns of the original matrix. Therefore, the next two columns are distributed to lab 1. This process continues until all columns in
5-18
the first rows are distributed. Then a similar process applies to the rows as you proceed down the matrix, as shown in the following distribution scheme:
Original matrix 1 2 3 4 5 6 7 8
LAB 3 LAB 1
9 10 11 12 13 14 15 16
17 18 19 20 21 22 23 24
LAB 1
25 26 27 28 29 30 31 32
LAB 2
33 34 35
41 42 43 44 45 46 47 48
LAB 2
49 50 51 52 53 54 55 56
57 58 59 60 61 62 63 64
LAB 1 LAB 2 LAB 3 LAB 4 LAB 1 LAB 2
LAB 4
36 37 38 39 40
LAB 3
LAB 4
LAB 3
LAB 4
The diagram above shows a scheme that requires four overlays of the lab grid to accommodate the entire original matrix. The following pmode session shows the code and resulting distribution of data to each of the labs:
5-19
The following points are worth noting: '2dbc' distribution might not offer any performance enhancement unless the block size is at least a few dozen. The default block size is 64. The lab grid should be as close to a square as possible. Not all functions that are enhanced to work on '1d' codistributed arrays work on '2dbc' codistributed arrays.
5-20
Note Using a for-loop over a distributed range (drange) is intended for explicit indexing of the distributed dimension of codistributed arrays (such as inside an spmd statement or a parallel job). For most applications involving parallel for-loops you should first try using parfor loops. See Chapter 2, Parallel for-Loops (parfor).
Parallelizing a for-Loop
If you already have a coarse-grained application to perform, but you do not want to bother with the overhead of defining jobs and tasks, you can take advantage of the ease-of-use that pmode provides. Where an existing program might take hours or days to process all its independent data sets, you can shorten that time by distributing these independent computations over your cluster. For example, suppose you have the following serial code:
results = zeros(1, numDataSets); for i = 1:numDataSets load(['\\central\myData\dataSet' int2str(i) '.mat']) results(i) = processDataSet(i); end plot(1:numDataSets, results); save \\central\myResults\today.mat results
The following changes make this code operate in parallel, either interactively in spmd or pmode, or in a parallel job:
results = zeros(1, numDataSets, codistributor()); for i = drange(1:numDataSets) load(['\\central\myData\dataSet' int2str(i) '.mat'])
5-21
results(i) = processDataSet(i); end res = gather(results, 1); if labindex == 1 plot(1:numDataSets, res); print -dtiff -r300 fig.tiff; save \\central\myResults\today.mat res end
Note that the length of the for iteration and the length of the codistributed array results need to match in order to index into results within a for drange loop. This way, no communication is required between the labs. If results was simply a replicated array, as it would have been when running the original code in parallel, each lab would have assigned into its part of results, leaving the remaining parts of results 0. At the end, results would have been a variant, and without explicitly calling labSend and labReceive or gcat, there would be no way to get the total results back to one (or all) labs. When using the load function, you need to be careful that the data files are accessible to all labs if necessary. The best practice is to use explicit paths to files on a shared file system. Correspondingly, when using the save function, you should be careful to only have one lab save to a particular file (on a shared file system) at a time. Thus, wrapping the code in if labindex == 1 is recommended. Because results is distributed across the labs, this example uses gather to collect the data onto lab 1. A lab cannot plot a visible figure, so the print function creates a viewable file of the plot.
5-22
require a lab to access portions of a codistributed array from another lab will generate an error. To illustrate this characteristic, you can try the following example, in which one for loop works, but the other does not. At the pmode prompt, create two codistributed arrays, one an identity matrix, the other set to zeros, distributed across four labs.
D = eye(8, 8, codistributor()) E = zeros(8, 8, codistributor())
By default, these arrays are distributed by columns; that is, each of the four labs contains two columns of each array. If you use these arrays in a for-drange loop, any calculations must be self-contained within each lab. In other words, you can only perform calculations that are limited within each lab to the two columns of the arrays that the labs contain. For example, suppose you want to set each column of array E to some multiple of the corresponding column of array D:
for j = drange(1:size(D,2)); E(:,j) = j*D(:,j); end
This statement sets the j-th column of E to j times the j-th column of D. In effect, while D is an identity matrix with 1s down the main diagonal, E has the sequence 1, 2, 3, etc., down its main diagonal. This works because each lab has access to the entire column of D and the entire column of E necessary to perform the calculation, as each lab works independently and simultaneously on two of the eight columns. Suppose, however, that you attempt to set the values of the columns of E according to different columns of D:
for j = drange(1:size(D,2)); E(:,j) = j*D(:,j+1); end
This method fails, because when j is 2, you are trying to set the second column of E using the third column of D. These columns are stored in different labs, so an error occurs, indicating that communication between the labs is not allowed.
5-23
Restrictions
To use for-drange on a codistributed array, the following conditions must exist: The codistributed array uses a 1-dimensional distribution scheme (not 2dbc). The distribution complies with the default partition scheme. The variable over which the for-drange loop is indexing provides the array subscript for the distribution dimension. All other subscripts can be chosen freely (and can be taken from for-loops over the full range of each dimension). To loop over all elements in the array, you can use for-drange on the dimension of distribution, and regular for-loops on all other dimensions. The following example executes in an spmd statement running on a MATLAB pool of 4 labs:
spmd PP = codistributed.zeros(6,8,12); RR = rand(6,8,12,codistributor()) % Default distribution: % by third dimension, evenly across 4 labs. for ii = 1:6 for jj = 1:8 for kk = drange(1:12) PP(ii,jj,kk) = RR(ii,jj,kk) + labindex; end end end end
5-24
For example,
help codistributed/normest
The following table lists the enhanced MATLAB functions that operate on codistributed arrays. Type of Function Data functions Data type functions Function Names
cumprod, cumsum, fft, max, min, prod, sum arrayfun, cast, cell2mat, cell2struct, celldisp, cellfun, char, double, fieldnames, int16, int32, int64, int8, logical, num2cell, rmfield, single, struct2cell, swapbytes, typecast, uint16, uint32, uint64, uint8 abs, acos, acosd, acosh, acot, acotd, acoth, acsc, acscd, acsch, angle, asec, asecd, asech, asin, asind, asinh, atan, atan2, atand, atanh, ceil, complex, conj, cos, cosd, cosh, cot, cotd, coth, csc, cscd, csch, exp, expm1, fix, floor, hypot, imag, isreal, log, log10, log1p, log2, mod, nextpow2, nthroot, pow2, real, reallog, realpow, realsqrt, rem, round, sec, secd, sech, sign, sin, sind, sinh, sqrt, tan, tand, tanh
5-25
Function Names
cat, diag, eps, find, isempty, isequal, isequalwithequalnans, isfinite, isinf, isnan, length, meshgrid, ndgrid, ndims, numel, reshape, size, sort, tril, triu chol, eig, inv, lu, norm, normest, qr, svd all, and (&), any, bitand, bitor, bitxor, ctranspose ('), end, eq (==), ge (>=), gt (>), horzcat ([]), ldivide (.\), le (<=), lt (<), minus (-), mldivide (\), mrdivide (/), mtimes (*), ne (~=), not (~), or (|), plus (+), power (.^), rdivide (./), subsasgn, subsindex, subsref, times (.*), transpose (.'), uminus (-), uplus (+), vertcat ([;]), xor full, issparse, nnz, nonzeros, nzmax, sparse, spfun, spones dot
5-26
6
Programming Overview
This chapter provides information you need for programming with Parallel Computing Toolbox software. Further details of evaluating functions in a cluster, programming distributed jobs, and programming parallel jobs are covered in later chapters. This chapter describes features common to programming all kinds of jobs. The sections are as follows. How Parallel Computing Products Run a Job on page 6-2 Create Simple Distributed Jobs on page 6-10 Parallel Configurations for Cluster Access on page 6-14 Job Monitor on page 6-27 Programming Tips on page 6-30 Profiling Parallel Code on page 6-35 Benchmarking Performance on page 6-47 Troubleshooting and Debugging on page 6-48
Programming Overview
Overview
Parallel Computing Toolbox and MATLAB Distributed Computing Server software let you solve computationally and data-intensive problems using MATLAB and Simulink on multicore and multiprocessor computers. Parallel processing constructs such as parallel for-loops and code blocks, distributed arrays, parallel numerical algorithms, and message-passing functions let you implement task-parallel and data-parallel algorithms at a high level in MATLAB without programming for specific hardware and network architectures. A job is some large operation that you need to perform in your MATLAB session. A job is broken down into segments called tasks. You decide how best to divide your job into tasks. You could divide your job into identical tasks, but tasks do not have to be identical. The MATLAB session in which the job and its tasks are defined is called the client session. Often, this is on the machine where you program MATLAB. The client uses Parallel Computing Toolbox software to perform the definition of jobs and tasks. MATLAB Distributed Computing Server software is the product that performs the execution of your job by evaluating each of its tasks and returning the result to your client session. The job manager is the part of the engine that coordinates the execution of jobs and the evaluation of their tasks. The job manager distributes the tasks for evaluation to the servers individual MATLAB sessions called workers. Use of the MathWorks job manager is optional; the distribution of tasks to workers can also be performed by a third-party scheduler, such as Microsoft Windows HPC Server (including CCS) or Platform LSF schedulers.
6-2
See the Glossary on page Glossary-1 for definitions of the parallel computing terms used in this manual.
MATLAB Worker
MATLAB Distributed Computing Server
MATLAB Client
Parallel Computing Toolbox
MATLAB Worker
MATLAB Distributed Computing Server
MATLAB Worker
MATLAB Distributed Computing Server
6-3
Programming Overview
Each worker is given a task from the running job by the job manager, executes the task, returns the result to the job manager, and then is given another task. When all tasks for a running job have been assigned to workers, the job manager starts running the next job with the next available worker. A MATLAB Distributed Computing Server software setup usually includes many workers that can all execute tasks simultaneously, speeding up execution of large MATLAB jobs. It is generally not important which worker executes a specific task. The workers evaluate tasks one at a time, returning the results to the job manager. The job manager then returns the results of all the tasks in the job to the client session. Note For testing your application locally or other purposes, you can configure a single computer as client, worker, and job manager. You can also have more than one worker session or more than one job manager session on a machine.
Worker
Client
All Results
Job
Worker
Client
All Results
Worker
A large network might include several job managers as well as several client sessions. Any client session can create, run, and access jobs on any job manager, but a worker session is registered with and dedicated to only one job manager at a time. The following figure shows a configuration with multiple job managers.
6-4
Worker
Client
Worker
Worker
Worker
Client
Worker
Worker
Local Scheduler
A feature of Parallel Computing Toolbox software is the ability to run a local scheduler and up to twelve workers on the client machine, so that you can run distributed and parallel jobs without requiring a remote cluster or MATLAB Distributed Computing Server software. In this case, all the processing required for the client, scheduling, and task evaluation is performed on the same computer. This gives you the opportunity to develop, test, and debug your distributed or parallel application before running it on your cluster.
Third-Party Schedulers
As an alternative to using the MathWorks job manager, you can use a third-party scheduler. This could be a Microsoft Windows HPC Server (including CCS), Platform LSF scheduler, PBS Pro scheduler, TORQUE scheduler, mpiexec, or a generic scheduler. Choosing Between a Third-Party Scheduler and Job Manager. You should consider the following when deciding to use a scheduler or the MathWorks job manager for distributing your tasks: Does your cluster already have a scheduler?
6-5
Programming Overview
If you already have a scheduler, you may be required to use it as a means of controlling access to the cluster. Your existing scheduler might be just as easy to use as a job manager, so there might be no need for the extra administration involved. Is the handling of parallel computing jobs the only cluster scheduling management you need? The MathWorks job manager is designed specifically for MathWorks parallel computing applications. If other scheduling tasks are not needed, a third-party scheduler might not offer any advantages. Is there a file sharing configuration on your cluster already? The MathWorks job manager can handle all file and data sharing necessary for your parallel computing applications. This might be helpful in configurations where shared access is limited. Are you interested in batch mode or managed interactive processing? When you use a job manager, worker processes usually remain running at all times, dedicated to their job manager. With a third-party scheduler, workers are run as applications that are started for the evaluation of tasks, and stopped when their tasks are complete. If tasks are small or take little time, starting a worker for each one might involve too much overhead time. Are there security concerns? Your own scheduler may be configured to accommodate your particular security requirements. How many nodes are on your cluster? If you have a large cluster, you probably already have a scheduler. Consult your MathWorks representative if you have questions about cluster size and the job manager. Who administers your cluster? The person administering your cluster might have a preference for how jobs are scheduled. Do you need to monitor your jobs progress or access intermediate data? A job run by the job manager supports events and callbacks, so that particular functions can run as each job and task progresses from one state to another.
6-6
In a mixed-platform environment, system administrators should be sure to follow the proper installation instructions for the local machine on which you are installing the software.
mdce Service
If you are using the MathWorks job manager, every machine that hosts a worker or job manager session must also run the mdce service. The mdce service controls the worker and job manager sessions and recovers them when their host machines crash. If a worker or job manager machine crashes, when the mdce service starts up again (usually configured to start at machine boot time), it automatically restarts the job manager and worker sessions to resume their sessions from before the system crash. These processes are covered more fully in the MATLAB Distributed Computing Server System Administrators Guide.
6-7
Programming Overview
Scheduler
Queued Pending
Job Job Job Job Job
Running
Job Job
createJob
Client
Job Job
submit
Job
Finished
Job Job Job Job
Worker
getAllOutputArguments
Stages of a Job
The following table describes each stage in the life cycle of a job. Job Stage Pending Description You create a job on the scheduler with the createJob function in your client session of Parallel Computing Toolbox software. The jobs first state is pending. This is when you define the job by adding tasks to it. When you execute the submit function on a job, the scheduler places the job in the queue, and the jobs state is queued. The scheduler executes jobs in the queue in the sequence in which they are submitted, all jobs moving up the queue as the jobs before them are finished. You can change the order of the jobs in the queue with the promote and demote functions.
Queued
6-8
Description When a job reaches the top of the queue, the scheduler distributes the jobs tasks to worker sessions for evaluation. The jobs state is running. If more workers are available than necessary for a jobs tasks, the scheduler begins executing the next job. In this way, there can be more than one job running at a time. When all of a jobs tasks have been evaluated, a job is moved to the finished state. At this time, you can retrieve the results from all the tasks in the job with the function getAllOutputArguments. When using a third-party scheduler, a job might fail if the scheduler encounters an error when attempting to execute its commands or access necessary files. When a jobs data has been removed from its data location or from the job manager, the state of the job in the client is destroyed. This state is available only as long as the job object remains in the client.
Finished
Failed
Destroyed
Note that when a job is finished, it remains in the job manager or DataLocation directory, even if you clear all the objects from the client session. The job manager or scheduler keeps all the jobs it has executed, until you restart the job manager in a clean state. Therefore, you can retrieve information from a job at a later time or in another client session, so long as the job manager has not been restarted with the -clean option. To permanently remove completed jobs from the job manager or schedulers data location, use the destroy function.
6-9
Programming Overview
This example runs the job as three tasks in three separate MATLAB worker sessions, reporting the results back to the session from which you ran dfeval. For more information about dfeval and in what circumstances you can use it, see Chapter 7, Evaluate Functions in a Cluster.
6-10
1 Identify a scheduler. Use findResource to indicate that you are using the
local scheduler and create the object sched, which represents the scheduler. (For more information, see Find a Job Manager on page 8-8 or Create and Run Jobs on page 8-21.)
sched = findResource('scheduler', 'type', 'local')
2 Create a job. Create job j on the scheduler. (For more information, see
array that is passed as an input argument. (For more information, see Create Tasks on page 8-12.)
createTask(j, @sum, 1, {[1 1]}) createTask(j, @sum, 1, {[2 2]}) createTask(j, @sum, 1, {[3 3]})
4 Submit the job to the scheduler queue for evaluation. The scheduler then
distributes the jobs tasks to MATLAB workers that are available for evaluating. The local scheduler actually starts a MATLAB worker session for each task, up to twelve at one time. (For more information, see Submit a Job to the Job Queue on page 8-13.)
submit(j);
5 Wait for the job to complete, then get the results from all the tasks of the
job. (For more information, see Retrieve the Jobs Results on page 8-14.)
waitForState(j) results = getAllOutputArguments(j) results = [2] [4] [6]
6 Destroy the job. When you have the results, you can permanently remove
6-11
Programming Overview
Getting Help
Command-Line Help on page 6-12 Help Browser on page 6-13
Command-Line Help
You can get command-line help on the toolbox object functions by using the syntax
help distcomp.objectType/functionName
The available choices for objectType are listed in the Chapter 11, Object Reference. List Available Functions. To find the functions available for each type of object, type
methods(obj)
where obj is an object of one of the available types. For example, to see the functions available for job manager objects, type
jm = findResource('scheduler','type','jobmanager'); methods(jm)
6-12
Help Browser
You can open the Help browser with the doc command. To open the browser on a specific reference page for a function or property, type
doc distcomp/RefName
where RefName is the name of the function or property whose reference page you want to read. For example, to open the Help browser on the reference page for the createJob function, type
doc distcomp/createJob
To open the Help browser on the reference page for the UserData property, type
doc distcomp/UserData
6-13
Programming Overview
Defining Configurations
Configurations allow you to define certain parameters and properties, then have your settings applied when creating objects in the MATLAB client. The functions that support the use of configurations are batch (also supports default configuration) createJob (also supports default configuration) createMatlabPoolJob (also supports default configuration) createParallelJob (also supports default configuration) createTask dfeval dfevalasync findResource matlabpool (also supports default configuration) pmode (also supports default configuration) set
6-14
You create and modify configurations through the Configurations Manager. You access the Configurations Manager using the Parallel pull-down menu on the MATLAB desktop. Select Parallel > Manage Configurations to open the Configurations Manger.
The first time you open the Configurations Manager, it lists only one configuration called local, which at first is the default configuration and has only default settings.
The following example provides instructions on how to create and modify configurations using the Configurations Manager and its menus and dialog boxes.
6-15
Programming Overview
that you want a new configuration whose type of scheduler is a job manager.
6-16
the following figure. In the Scheduler tab, enter the host name for the machine on which the job manager is running and the name of the job manager. If you are entering information for an actual job manager already running on your network, enter the appropriate text. If you are unsure about job manager names and locations on your network, ask your system administrator for help.
Note Fields that indicate Unset or that you leave empty, have no effect on their property values. For those properties, the configuration does not alter the values that you had set programmatically before applying the configuration.
6-17
Programming Overview
3 In the Jobs tab, enter 4 and 4 for the maximum and minimum number of
workers. This specifies that for jobs using this configuration, they require at least four workers and use no more than four workers. Therefore, the job runs on exactly four workers, even if it has to wait until four workers are available before starting.
4 Click OK to save the configuration and close the dialog box. Your new
6-18
duplicate an existing configuration and modify only the parts you need to change:
a In the Configurations Manager, right-click the configuration
The duplicate configuration is created with a default name using the original name along with the extension .copy1.
b Double-click the new configuration to open its properties dialog. c Change the name of the new configuration to MyJMconfig2. d Edit the description field to change its text to My job manager and any
workers.
6-19
Programming Overview
6 Select the Jobs tab. Remove the 4 from each of the fields for minimum and
maximum workers.
You now have two configurations that differ only in the number of workers required for running a job.
6-20
After creating a job, you can apply either configuration to that job as a way of specifying how many workers it should run on.
want to export.
2 Click File > Export. (Alternatively, you can right-click the configuration
file. The default file name is the same as the name of the configuration it contains, with a .mat extension appended; these do not need to be the same, so you can alter the names if you want to.
6-21
Programming Overview
Configurations saved in this way can then be imported by other MATLAB software users:
1 In the Configuration Manager, click File > Import. 2 In the Import Configuration dialog box, browse to find the .mat file for the
configuration you want to import. Select the file and click Import. The imported configuration appears in your Configurations Manager list. Note that the list contains the configuration name, which is not necessarily the file name. If you already have a configuration with the same name as the one you are importing, the imported configuration gets an extension added to its name so you can distinguish it. You can also import configurations programmatically with the
importParallelConfig function. For details and examples, see the importParallelConfig reference page.
Validate Configurations
The Configurations Manager includes a tool for validating configurations.
6-22
want to test in the list of those available. Note that you can highlight a configuration this way without changing the selected default configuration. So a configuration selected for validation does not need to be your default configuration.
3 Click Start Validation.
The Configuration Validation tool attempts four operations to validate the chosen configuration: Uses findResource to locate the scheduler Runs a distributed job using the configuration Runs a parallel job using the configuration Runs a MATLAB pool job using the configuration While the tests are running, the Configurations Manager displays their progress as shown here.
6-23
Programming Overview
You can adjust the timeout allowed for each stage of the testing. If your cluster does not have enough workers available to perform the validation, the test times out and returns a failure. Note You cannot run a configuration validation if you have a MATLAB pool open. The configuration listing displays the overall validation result for each configuration. The following figure shows overall validation results for one configuration that passed and one that failed. The selected configuration is the one that failed.
Note When using an mpiexec scheduler, a failure is expected for the Distributed Job stage. It is normal for the test then to proceed to the Parallel Job and Matlabpool stages. For each stage of the validation testing, you can click Details to get more information about that stage. This information includes any error messages, debug logs, and other data that might be useful in diagnosing problems or helping to determine proper configuration or network settings. The Configuration Validation tool keeps the test results available until the current MATLAB session closes.
6-24
Finding Schedulers
When executing the findResource function, you can use configurations to identify a particular scheduler and apply property values. For example,
jm = findResource('scheduler', 'Configuration', 'our_jobmanager')
This command finds the scheduler defined by the settings of the configuration named our_jobmanager and sets property values on the scheduler object based on settings in the configuration. The advantage of configurations is that you can alter your scheduler choices without changing your MATLAB application code, merely by changing the configuration settings
6-25
Programming Overview
For a third-party scheduler such as Platform LSF, the command might look like
lsfsched = findResource('scheduler', 'Configuration', 'my_lsf_config');
Creating Jobs
Because the properties of scheduler, job, and task objects can be defined in a configuration, you do not have to define them in your application. Therefore, the code itself can accommodate any type of scheduler. For example,
job1 = createJob(sched, 'Configuration', 'MyConfig');
The configuration defined as MyConfig must define any and all properties necessary and appropriate for your scheduler and configuration, and the configuration must not include any parameters inconsistent with your setup. All changes necessary to use a different scheduler can now be made in the configuration, without any modification needed in the application.
Notice that the Configuration property of a job indicates the configuration that was applied to the job.
get(job1, 'Configuration') our_jobmanager_config
When you apply a configuration to an object, all the properties defined in that configuration get applied to the object, and the objects Configuration property is set to reflect the name of the configuration that you applied. If you later directly change any of the objects individual properties, the objects Configuration property is cleared.
6-26
Job Monitor
Job Monitor
In this section... Job Monitor GUI on page 6-27 Manage Jobs Using the Job Monitor on page 6-28 Identify Task Errors Using the Job Monitor on page 6-29
The job monitor lists all the jobs that exist for the scheduler specified in the selected configuration. You can choose any one of your configurations (those available in your current session Configurations Manager), and whether to display jobs from all users or only your own jobs.
6-27
Programming Overview
Cancel Stops a running job and changes its state to 'finished'. If the job is pending or queued, the state changes to 'finished' without its ever running. This is the same as the command-line cancel function for the job. Destroy Deletes the jobs data and removes it from the queue. This is the same as the command-line destroy function for the job. Assign Job to Workspace This creates a job object in the MATLAB workspace so that you can access the job and its properties from the command line. This is accomplished by the findJob command, which is reflected in the command window. Show Errors This displays all the tasks that generated an error in that job, with their error properties.
6-28
Job Monitor
If you save this script in a file named invert_me.m, you can try to run the script as a batch job:
defaultParallelConfiguration('local') batch('invert_me')
When updated after the job runs, the Job Monitor includes the job created by the batch command, with an error icon ( ) for this job. Right-click the job in the list, and select Show Errors. For all the tasks with an error in that job, the task information, including properties related to the error, display in the MATLAB command window:
Task ID 1 from Job ID 2 Information =================================== State : finished Function : @parallel.internal.cluster.executeScript StartTime : Tue Jun 28 11:46:28 EDT 2011 Running Duration : 0 days 0h 0m 1s - Task Result Properties ErrorIdentifier : MATLAB:square ErrorMessage : Matrix must be square. Error Stack : invert_me.m at 2 : executeScript.m at 24
6-29
Programming Overview
Programming Tips
In this section... Program Development Guidelines on page 6-30 Current Working Directory of a MATLAB Worker on page 6-31 Writing to Files from Workers on page 6-32 Saving or Sending Objects on page 6-32 Using clear functions on page 6-33 Running Tasks That Call Simulink Software on page 6-33 Using the pause Function on page 6-33 Transmitting Large Amounts of Data on page 6-33 Interrupting a Job on page 6-33 Speeding Up a Job on page 6-34
functions so that as you progress, you are not trying to debug the functions and the distribution at the same time. Run your functions in a single instance of MATLAB software on your local computer. For programming suggestions, see Techniques for Improving Performance in the MATLAB documentation.
2 Decide whether you need a distributed or parallel job. If your
application involves large data sets on which you need simultaneous calculations performed, you might benefit from a parallel job with
6-30
Programming Tips
distributed arrays. If your application involves looped or repetitive calculations that can be performed independently of each other, a distributed job might be appropriate.
3 Modify your code for division. Decide how you want your code divided.
For a distributed job, determine how best to divide it into tasks; for example, each iteration of a for-loop might define one task. For a parallel job, determine how best to take advantage of parallel processing; for example, a large array can be distributed across all your labs.
4 Use pmode to develop parallel functionality. Use pmode with the local
scheduler to develop your functions on several workers (labs) in parallel. As you progress and use pmode on the remote cluster, that might be all you need to complete your work.
5 Run the distributed or parallel job with a local scheduler. Create a
parallel or distributed job, and run the job using the local scheduler with several local workers. This verifies that your code is correctly set up for batch execution, and in the case of a distributed job, that its computations are properly divided into tasks.
6 Run the distributed job on only one cluster node. Run your
distributed job with one task to verify that remote distribution is working between your client and the cluster, and to verify file and path dependencies.
7 Run the distributed or parallel job on multiple cluster nodes. Scale
up your job to include as many tasks as you need for a distributed job, or as many workers (labs) as you need for a parallel job. Note The client session of MATLAB must be running the Java Virtual Machine (JVM) to use Parallel Computing Toolbox software. Do not start MATLAB with the -nojvm flag.
6-31
Programming Overview
where CHECKPOINTBASE is defined in the mdce_def file, HOSTNAME is the name of the node on which the worker is running, and WORKERNAME is the name of the MATLAB worker session. For example, if the worker named worker22 is running on host nodeA52, and its CHECKPOINTBASE value is C:\TEMP\MDCE\Checkpoint, the starting current directory for that worker session is
C:\TEMP\MDCE\Checkpoint\nodeA52_worker22_mlworker_log\work
6-32
Programming Tips
added on the workers. Other than in the task function code, typical ways of loading these objects might be in taskStartup, jobStartup, and in the case of workers in a MATLAB pool, in poolStartup and using pctRunOnAll.
clears all Parallel Computing Toolbox objects from the current MATLAB session. They still remain in the job manager. For information on recreating these objects in the client session, see Recover Objects on page 8-18.
Interrupting a Job
Because jobs and tasks are run outside the client session, you cannot use Ctrl+C (^C) in the client session to interrupt them. To control or interrupt
6-33
Programming Overview
the execution of jobs and tasks, use such functions as cancel, destroy, demote, promote, pause, and resume.
Speeding Up a Job
You might find that your code runs slower on multiple workers than it does on one desktop computer. This can occur when task startup and stop time is not negligible relative to the task run time. The most common mistake in this regard is to make the tasks too small, i.e., too fine-grained. Another common mistake is to send large amounts of input or output data with each task. In both of these cases, the time it takes to transfer data and initialize a task is far greater than the actual time it takes for the worker to evaluate the task function.
6-34
Introduction
The parallel profiler provides an extension of the profile command and the profile viewer specifically for parallel jobs, to enable you to see how much time each lab spends evaluating each function and how much time communicating or waiting for communications with the other labs. Before using the parallel profiler, familiarize yourself with the standard profiler and its views, as described in Profiling for Improving Performance. Note The parallel profiler works on parallel jobs, including inside pmode. It does not work on parfor-loops.
Now the profiler is collecting information about the execution of code on each lab and the communications between the labs. Such information includes: Execution time of each function on each lab Execution time of each line of code in each function
6-35
Programming Overview
Amount of data transferred between each lab Amount of time each lab spends waiting for communications With the parallel profiler on, you can proceed to execute your code while the profiler collects the data. In the pmode Parallel Command Window, to find out if the profiler is on, type:
P>> mpiprofile status
For a complete list of options regarding profiler data details, clearing data, etc., see the mpiprofile reference page.
The remainder of this section is an example that illustrates some of the features of the parallel profile viewer. This example executes in a pmode session running on four local labs. Initiate pmode by typing in the MATLAB Command Window:
pmode start local 4
When the Parallel Command Window (pmode) starts, type the following code at the pmode prompt:
P>> P>> P>> P>> P>> P>> R1 = rand(16, codistributor()) R2 = rand(16, codistributor()) mpiprofile on P = R1*R2 mpiprofile off mpiprofile viewer
6-36
The last command opens the Profiler window, first showing the Parallel Profile Summary (or function summary report) for lab 1.
The function summary report displays the data for each function executed on a lab in sortable columns with the following headers: Column Header Calls Total Time Self Time Total Comm Time Self Comm Waiting Time Description How many times the function was called on this lab The total amount of time this lab spent executing this function The time this lab spent inside this function, not within children or subfunctions The total time this lab spent transferring data with other labs, including waiting time to receive data The time this lab spent during this function waiting to receive data from other labs
6-37
Programming Overview
Column Header Total Interlab Data Computation Time Ratio Total Time Plot
Description The amount of data transferred to and from this lab for this function The ratio of time spent in computation for this function vs. total time (which includes communication time) for this function Bar graph showing relative size of Self Time, Self Comm Waiting Time, and Total Time for this function on this lab
6-38
Click the name of any function in the list for more details about the execution of that function. The function detail report for codistributed.mtimes includes this listing:
The code that is displayed in the report is taken from the client. If the code has changed on the client since the parallel job ran on the labs, or if the labs are running a different version of the functions, the display might not accurately reflect what actually executed. You can display information for each lab, or use the comparison controls to display information for several labs simultaneously. Two buttons provide Automatic Comparison Selection, allowing you to compare the data from the labs that took the most versus the least amount of time to execute the code, or data from the labs that spent the most versus the least amount of time in performing interlab communication. Manual Comparison Selection allows you to compare data from specific labs or labs that meet certain criteria.
6-39
Programming Overview
The following listing from the summary report shows the result of using the Automatic Comparison Selection of Compare (max vs. min TotalTime). The comparison shows data from lab 3 compared to lab 1 because these are the labs that spend the most versus least amount of time executing the code.
6-40
The following figure shows a summary of all the functions executed during the profile collection time. The Manual Comparison Selection of max Time Aggregate means that data is considered from all the labs for all functions to determine which lab spent the maximum time on each function. Next to each functions name is the lab that took the longest time to execute that function. The other columns list the data from that lab.
6-41
Programming Overview
The next figure shows a summary report for the labs that spend the most versus least time for each function. A Manual Comparison Selection of max Time Aggregate against min Time >0 Aggregate generated this summary. Both aggregate settings indicate that the profiler should consider data from all labs for all functions, for both maximum and minimum. This report lists the data for codistributed.mtimes from labs 3 and 1, because they spent the maximum and minimum times on this function. Similarly, other functions are listed.
6-42
Click on a function name in the summary listing of a comparison to get a detailed comparison. The detailed comparison for codistributed.mtimes looks like this, displaying line-by-line data from both labs:
6-43
Programming Overview
To see plots of communication data, select Plot All PerLab Communication in the Show Figures menu. The top portion of the plot view report plots how much data each lab receives from each other lab for all functions.
6-44
To see only a plot of interlab communication times, select Plot CommTimePerLab in the Show Figures menu.
Plots like those in the previous two figures can help you determine the best way to balance work among your labs, perhaps by altering the partition scheme of your codistributed arrays.
6-45
Programming Overview
6-46
Benchmarking Performance
Benchmarking Performance
In this section... Demos on page 6-47 HPC Challenge Benchmarks on page 6-47
Demos
Several benchmarking demos can help you understand and evaluate performance of the parallel computing products. You can access these demos in the Help Browser under the Parallel Computing Toolbox node: expand the node for Demos then Benchmarks.
file is self-documented with explanatory comments. These files are not self-contained demos, but rather require that you know enough about your cluster to be able to provide the necessary information when using these files.
6-47
Programming Overview
6-48
The worker that ran the task did not have access to the function function_name. One solution is to make sure the location of the functions file, function_name.m, is included in the jobs PathDependencies property. Another solution is to transfer the function file to the worker by adding function_name.m to the FileDependencies property of the job.
In determining the cause of this error, consider the following questions: What is the workers current directory? Can the worker find the file or directory? What user is the worker running as? Does the worker have permission to read or write the file in question?
6-49
Programming Overview
If using a generic scheduler, make sure the submit function redirects error messages to a log file. Possible causes of the problem are The MATLAB worker failed to start due to licensing errors, the executable is not on the default path on the worker machine, or is not installed in the location where the scheduler expected it to be. MATLAB could not read/write the job input/output files in the schedulers data location. The data location may not be accessible to all the worker nodes, or the user that MATLAB runs as does not have permission to read/write the job files. If using a generic scheduler
The environment variable MDCE_DECODE_FUNCTION was not defined before the MATLAB worker started. The decode function was not on the workers path. The passphrase to smpd was incorrect or missing. The smpd daemon was not running on all the specified machines.
If using mpiexec
This code displays the nonempty error messages of the tasks found in the job object yourjob.
6-50
Debug Logs
If you are using a supported third-party scheduler, you can use the getDebugLog function to read the debug log from the scheduler for a particular job or task. For example, find the failed job on your LSF scheduler, and read its debug log.
sched = findResource('scheduler', 'type', 'lsf') failedjob = findJob(sched, 'State', 'failed'); message = getDebugLog(sched, failedjob(1))
the most likely reasons for this failure are The client cannot contact the job manager host via multicast. Try to fully specify where to look for the job manager by using the LookupURL property in your call to findResource:
findResource('scheduler','type','jobmanager', ... 'LookupURL','JobMgrHostName')
6-51
Programming Overview
The job manager is currently not running. Firewalls do not allow traffic from the client to the job manager. The client and the job manager are not running the same version of the software. The client and the job manager cannot resolve each others short hostnames.
6-52
endif endif
The telling symptom is the phrase "Received message is too long:" followed by a very large number. The sftp server starts a shell, usually bash or tcsh, to set your standard read and write permissions appropriately before transferring files. The server initializes the shell in the standard way, calling files like .bashrc and .cshrc. This problem happens if your shell emits text to standard out when it starts. That text is transferred back to the sftp client running inside MATLAB, and is interpreted as the size of the sftp servers response message. To work around this error, locate the shell startup file code that is emitting the text, and either remove it or bracket it within if statements to see if the sftp server is starting the shell:
if ($?TERM != 0) then if ("$TERM" != "dumb") then /your command/ endif endif
You can test this outside of MATLAB with a standard UNIX or Windows sftp command-line client before trying again in MATLAB. If the problem is not fixed, the error message persists:
> sftp yourSubmitMachine Connecting to yourSubmitMachine... Received message too long 1718579042
6-53
Programming Overview
6-54
7
Evaluate Functions in a Cluster
In many cases, the tasks of a job are all the same, or there are a limited number of different kinds of tasks in a job. Parallel Computing Toolbox software offers a solution for these cases that alleviates you from having to define individual tasks and jobs when evaluating a function in a cluster of workers. The two ways of evaluating a function on a cluster are described in the following sections: Evaluate Functions Synchronously on page 7-2 Evaluate Functions Asynchronously on page 7-8
Scope of dfeval
When you evaluate a function in a cluster of computers with dfeval, you provide basic required information, such as the function to be evaluated, the number of tasks to divide the job into, and the variable into which the results are returned. Synchronous (sync) evaluation in a cluster means that your MATLAB session is blocked until the evaluation is complete and the results are assigned to the designated variable. So you provide the necessary information, while Parallel Computing Toolbox software handles all the job-related aspects of the function evaluation. When executing the dfeval function, the toolbox performs all these steps of running a job:
1 Finds a job manager or scheduler 2 Creates a job 3 Creates tasks in that job 4 Submits the job to the queue in the job manager or scheduler 5 Retrieves the results from the job 6 Destroys the job
By allowing the system to perform all the steps for creating and running jobs with a single function call, you do not have access to the full flexibility offered by Parallel Computing Toolbox software. However, this narrow functionality meets the requirements of many straightforward applications. To focus the scope of dfeval, the following limitations apply:
7-2
You can pass property values to the job object; but you cannot set any task-specific properties, including callback functions, unless you use configurations. All the tasks in the job must have the same number of input arguments. All the tasks in the job must have the same number of output arguments. If you are using a third-party scheduler instead of the job manager, you must use configurations in your call to dfeval. See Parallel Configurations for Cluster Access on page 6-14, and the reference page for dfeval. You do not have direct access to the job manager, job, or task objects, i.e., there are no objects in your MATLAB workspace to manipulate (though you can get them using findResource and the properties of the scheduler object). Note that dfevalasync returns a job object. Without access to the objects and their properties, you do not have control over the handling of errors.
Arguments of dfeval
Suppose the function myfun accepts three input arguments, and generates two output arguments. To run a job with four tasks that call myfun, you could type
[X, Y] = dfeval(@myfun, {a1 a2 a3 a4}, {b1 b2 b3 b4}, {c1 c2 c3 c4});
The number of elements of the input argument cell arrays determines the number of tasks in the job. All input cell arrays must have the same number of elements. In this example, there are four tasks. Because myfun returns two arguments, the results of your job will be assigned to two cell arrays, X and Y. These cell arrays will have four elements each, for the four tasks. The first element of X will have the first output argument from the first task, the first element of Y will have the second argument from the first task, and so on. The following table shows how the job is divided into tasks and where the results are returned.
7-3
Results
X{1}, Y{1} X{2}, Y{2} X{3}, Y{3} X{4}, Y{4}
So using one dfeval line would be equivalent to the following code, except that dfeval can run all the statements simultaneously on separate machines.
[X{1}, [X{2}, [X{3}, [X{4}, Y{1}] Y{2}] Y{3}] Y{4}] = = = = myfun(a1, myfun(a2, myfun(a3, myfun(a4, b1, b2, b3, b4, c1); c2); c3); c4);
For further details and examples of the dfeval function, see the dfeval reference page.
You can use dfeval to run this function on four sets of data using four tasks in a single job. The input data can be represented by the four vectors,
[1 2 6] [10 20 60] [100 200 600] [1000 2000 6000]
7-4
A quick look at the first set of data tells you that its mean is 3, while its median is 2. So,
[x,y] = averages(1,2,6) x = 3 y = 2
When calling dfeval, its input requires that the data be grouped together such that the first input argument to each task function is in the first cell array argument to dfeval, all second input arguments to the task functions are grouped in the next cell array, and so on. Because we want to evaluate four sets of data with four tasks, each of the three cell arrays will have four elements. In this example, the first arguments for the task functions are 1, 10, 100, and 1000. The second inputs to the task functions are 2, 20, 200, and 2000. With the task inputs arranged thus, the call to dfeval looks like this.
[A, B] = dfeval(@averages, {1 10 100 1000}, ... {2 20 200 2000}, {6 60 600 6000}, 'jobmanager', ... 'MyJobManager', 'FileDependencies', {'averages.m'}) A = [ 3] [ 30] [ 300] [3000] B = [ 2] [ 20] [ 200] [2000]
Notice that the first task evaluates the first element of the three cell arrays. The results of the first task are returned as the first elements of each of the two output values. In this case, the first task returns a mean of 3 and median of 2. The second task returns a mean of 30 and median of 20.
7-5
If the original function were written to accept one input vector, instead of three input values, it might make the programming of dfeval simpler. For example, suppose your task function were
function [mean_, median_] = avgs (V) % AVGS Return mean and median of input vector mean_ = mean(V); median_ = median(V);
Now the function requires only one argument, so a call to dfeval requires only one cell array. Furthermore, each element of that cell array can be a vector containing all the values required for an individual task. The first vector is sent as a single argument to the first task, the second vector to the second task, and so on.
[A,B] = dfeval(@avgs, {[1 2 6] [10 20 60] ... [100 200 600] [1000 2000 6000]}, 'jobmanager', ... 'MyJobManager', 'FileDependencies', {'avgs.m'}) A = [ 3] [ 30] [ 300] [3000] B = [ 2] [ 20] [ 200] [2000]
If you cannot vectorize your function, you might have to manipulate your data arrangement for using dfeval. Returning to our original data in this example, suppose you want to start with data in three vectors.
v1 v2 v3 v4 = = = = [1 2 6]; [10 20 60]; [100 200 600]; [1000 2000 6000];
7-6
Now you can use these cell arrays as your input arguments for dfeval.
[A, B] = dfeval(@averages, c1, c2, c3, 'jobmanager', ... 'MyJobManager', 'FileDependencies', {'averages.m'}) A = [ 3] [ 30] [ 300] [3000] B = [ 2] [ 20] [ 200] [2000]
7-7
Note that you have to specify the number of output arguments that each task will return (2, in this example). The MATLAB session does not wait for the job to execute, but returns the prompt immediately. Instead of assigning results to cell array variables, the function creates a job object in the MATLAB workspace that you can use to access job status and results. You can use the MATLAB session to perform other operations while the job is being run on the cluster. When you want to get the jobs results, you should make sure it is finished before retrieving the data.
waitForState(job1, 'finished') results = getAllOutputArguments(job1) results = [ 3] [ 30] [ 300] [3000]
The structure of the output arguments is now slightly different than it was for dfeval. The getAllOutputArguments function returns all output arguments from all tasks in a single cell array, with one row per task. In this example,
7-8
each row of the cell array results will have two elements. So, results{1,1} contains the first output argument from the first task, results{1,2} contains the second argument from the first task, and so on. For further details and examples of the dfevalasync function, see the dfevalasync reference page.
7-9
7-10
8
Program Distributed Jobs
A distributed job is one whose tasks do not directly communicate with each other. The tasks do not need to run simultaneously, and a worker might run several tasks of the same job in succession. Typically, all tasks perform the same or similar functions on different data sets in an embarrassingly parallel configuration. The following sections describe how to program distributed jobs: Use a Local Scheduler on page 8-2 Use a Job Manager on page 8-8 Use a Fully Supported Third-Party Scheduler on page 8-21 Use the Generic Scheduler Interface on page 8-34
8-2
location. You can find existing jobs using the findJob function or the Jobs property of the scheduler object.
Create a Job
You create a job with the createJob function. This statement creates a job in the schedulers data location, creates the job object job1 in the client session, and if you omit the semicolon at the end of the command, displays some information about the job.
job1 = createJob(sched) Job ID 1 Information ==================== UserName State SubmitTime StartTime Running Duration - Data Dependencies FileDependencies : {} PathDependencies : {} - Associated Task(s) Number Pending Number Running Number Finished TaskID of errors : 0 : 0 : 0 : : eng864 : pending : : :
You can use the get function to see all the properties of this job object.
8-3
get(job1) Configuration: Name: ID: UserName: Tag: State: CreateTime: SubmitTime: StartTime: FinishTime: Tasks: FileDependencies: PathDependencies: JobData: Parent: UserData:
'' 'Job1' 1 'eng864' '' 'pending' 'Mon Jan 08 15:40:18 EST 2007' '' '' '' [0x1 double] {0x1 cell} {0x1 cell} [] [1x1 distcomp.localscheduler] []
Note that the jobs State property is pending. This means the job has not yet been submitted (queued) for running, so you can now add tasks to it. The schedulers display now indicates the existence of your job, which is the pending one.
sched Local Scheduler Information =========================== Type ClusterOsType DataLocation HasSharedFilesystem - Assigned Jobs Number Number Number Number Pending Queued Running Finished : : : : 1 0 0 0 : : : : local pc C:\WINNT\Profiles\eng864\App... true
8-4
Create Tasks
After you have created your job, you can create tasks for the job using the createTask function. Tasks define the functions to be evaluated by the workers during the running of the job. Often, the tasks of a job are all identical. In this example, five tasks will each generate a 3-by-3 matrix of random numbers.
createTask(job1, @rand, 1, {{3,3} {3,3} {3,3} {3,3} {3,3}});
The local scheduler starts up to twelve workers and distributes the tasks of job1 to its workers for evaluation.
8-5
After the job is complete, you can repeat the commands to examine the updated status of the scheduler, job, and task objects:
sched job1 get(job1,'Tasks')
8-6
8-7
8-8
To find a specific job manager, use parameter-value pairs for matching. In this example, MyJobManager is the name of the job manager, while MyJMhost is the hostname of the machine running the job manager lookup service.
jm = findResource('scheduler','type','jobmanager', ... 'Name','MyJobManager','LookupURL','MyJMhost') jm = Jobmanager Information ====================== Type : jobmanager ClusterOsType : 'pc' DataLocation : database on MyJobManager@MyJMhost - Assigned Jobs Number Number Number Number Pending Queued Running Finished : : : : 0 0 0 0
- Authentication and Security UserName : myloginname SecurityLevel : 0 - Jobmanager Specific Properties Name Hostname HostAddress State ClusterSize NumberOfIdleWorkers NumberOfBusyWorkers : : : : : : : MyJobManager MyJMhost 123.123.123.123 running 2 2 0
You can view all the accessible properties of the job manager object with the get function:
get(jm)
8-9
If your network supports multicast, you can omit property values to search on, and findResource returns all available job managers.
all_managers = findResource('scheduler','type','jobmanager')
You can then examine the properties of each job manager to identify which one you want to use.
for i = 1:length(all_managers) get(all_managers(i)) end
When you have identified the job manager you want to use, you can isolate it and create a single object.
jm = all_managers(3)
Create a Job
You create a job with the createJob function. Although this command executes in the client session, it actually creates the job on the job manager, jm, and creates a job object, job1, in the client session.
job1 = createJob(jm)job1 = Job ID 1 Information ==================== UserName AuthorizedUsers State SubmitTime StartTime Running Duration - Data Dependencies FileDependencies : {} PathDependencies : {} - Associated Task(s) : myloginname : {} : pending : : :
8-10
: 0 : 0 : 0 :
- Jobmanager Dependent Properties MaximumNumberOfWorkers MinimumNumberOfWorkers Timeout RestartWorker QueuedFcn RunningFcn FinishedFcn : : : : : : : Inf 1 Inf false
Use get to see all the accessible properties of this job object.
get(job1)
Note that the jobs State property is pending. This means the job has not been queued for running yet, so you can now add tasks to it. The job managers display now includes one pending job.
jm jm = Jobmanager Information ====================== Type : jobmanager ClusterOsType : 'pc' DataLocation : database on MyJobManager@MyJMhost - Assigned Jobs Number Pending Number Queued Number Running : 1 : 0 : 0
8-11
Number Finished : 0 - Authentication and Security UserName : myloginname SecurityLevel : 0 - Jobmanager Specific Properties Name Hostname HostAddress State ClusterSize NumberOfIdleWorkers NumberOfBusyWorkers : : : : : : : MyJobManager MyJMhost 123.123.123.123 running 2 2 0
You can transfer files to the worker by using the FileDependencies property of the job object. For details, see the FileDependencies reference page and Share Code on page 8-14.
Create Tasks
After you have created your job, you can create tasks for the job using the createTask function. Tasks define the functions to be evaluated by the workers during the running of the job. Often, the tasks of a job are all identical. In this example, each task will generate a 3-by-3 matrix of random numbers.
createTask(job1, createTask(job1, createTask(job1, createTask(job1, createTask(job1, @rand, @rand, @rand, @rand, @rand, 1, 1, 1, 1, 1, {3,3}); {3,3}); {3,3}); {3,3}); {3,3});
8-12
Alternatively, you can create the five tasks with one call to createTask by providing a cell array of five cell arrays defining the input arguments to each task.
T = createTask(job1, @rand, 1, {{3,3} {3,3} {3,3} {3,3} {3,3}});
The job manager distributes the tasks of job1 to its registered workers for evaluation. Each worker performs the following steps for task evaluation:
1 Receive FileDependencies and PathDependencies from the job. Place
You can specify this function in FileDependencies or PathDependencies. If the same worker evaluates subsequent tasks for this job, jobStartup does not run between tasks.
3 Run the taskStartup function. You can specify this function in
evaluation that the worker performs, so it could occur multiple times on a worker for each job.
4 If the worker is part of forming a new MATLAB pool, run the poolStartup function. (This occurs when executing matlabpool open or when running
property.
8-13
Share Code
Because the tasks of a job are evaluated on different machines, each machine must have access to all the files needed to evaluate its tasks. The basic mechanisms for sharing code are explained in the following sections: Access Files Directly on page 8-15
8-14
Pass Data Between Sessions on page 8-15 Pass MATLAB Code for Startup and Finish on page 8-16
These files can be passed to the worker by the jobs FileDependencies or PathDependencies property. Otherwise, the version of each of these files that is used is the one highest on the workers path. Access to files among shared resources can depend upon permissions based on the user name. You can set the user name with which the job manager and worker services of MATLAB Distributed Computing Server software run by setting the MDCEUSER value in the mdce_def file before starting the services. For Microsoft Windows operating systems, there is also MDCEPASS for providing the account password for the specified user. For an explanation of service default settings and the mdce_def file, see Defining the Script Defaults in the MATLAB Distributed Computing Server System Administrators Guide.
8-15
could include MATLAB code necessary for task evaluation, or the input data for processing or output data resulting from task evaluation. All these properties are described in detail in their own reference pages: InputArguments This property of each task contains the input data provided to the task constructor. This data gets passed into the function when the worker performs its evaluation. OutputArguments This property of each task contains the results of the functions evaluation. JobData This property of the job object contains data that gets sent to every worker that evaluates tasks for that job. This property works efficiently because the data is passed to a worker only once per job, saving time if that worker is evaluating more than one task for the job. FileDependencies This property of the job object lists all the directories and files that get zipped and sent to the workers. At the worker, the data is unzipped, and the entries defined in the property are added to the path of the MATLAB worker session. PathDependencies This property of the job object provides pathnames that are added to the MATLAB workers path, reducing the need for data transfers in a shared file system. There is a default maximum amount of data that can be sent in a single call for setting properties. This limit applies to the OutputArguments property as well as to data passed into a job as input arguments or FileDependencies. If the limit is exceeded, you get an error message. For more information about this data transfer size limit, see Object Data Size Limitations on page 6-48.
8-16
taskStartup.m automatically executes on a worker each time the worker begins evaluation of a task. poolStartup.m automatically executes on a worker each time the worker is included in a newly started MATLAB pool. taskFinish.m automatically executes on a worker each time the worker completes evaluation of a task. Empty versions of these files are provided in the directory
matlabroot/toolbox/distcomp/user
You can edit these files to include whatever MATLAB code you want the worker to execute at the indicated times. Alternatively, you can create your own versions of these files and pass them to the job as part of the FileDependencies property, or include the path names to their locations in the PathDependencies property. The worker gives precedence to the versions provided in the FileDependencies property, then to those pointed to in the PathDependencies property. If any of these files is not included in these properties, the worker uses the version of the file in the toolbox/distcomp/user directory of the workers MATLAB installation.
8-17
Recover Objects
A client session of Parallel Computing Toolbox software can access any of the objects in MATLAB Distributed Computing Server software, whether the current client session or another client session created these objects. You create job manager and worker objects in the client session by using the findResource function. These client objects refer to sessions running in the engine.
jm = findResource('scheduler','type','jobmanager', ... 'Name','Job_Mgr_123','LookupURL','JobMgrHost')
If your network supports multicast, you can find all available job managers by omitting any specific property information.
jm_set = findResource('scheduler','type','jobmanager')
The array jm_set contains all the job managers accessible from the client session. You can index through this array to determine which job manager is of interest to you.
jm = jm_set(2)
When you have access to the job manager by the object jm, you can create objects that reference all those objects contained in that job manager. All the jobs contained in the job manager are accessible in its Jobs property, which is an array of job objects.
8-18
all_jobs = get(jm,'Jobs')
You can index through the array all_jobs to locate a specific job. Alternatively, you can use the findJob function to search in a job manager for particular job identified by any of its properties, such as its State.
finished_jobs = findJob(jm,'State','finished')
This command returns an array of job objects that reference all finished jobs on the job manager jm.
8-19
The destroy function permanently removes these jobs from the job manager. The clear function removes the object references from the local MATLAB workspace. Start a Job Manager from a Clean State. When a job manager starts, by default it starts so that it resumes its former session with all jobs intact. Alternatively, a job manager can start from a clean state with all its former history deleted. Starting from a clean state permanently removes all job and task data from the job manager of the specified name on a particular host. As a network administration feature, the -clean flag of the job manager startup script is described in Starting in a Clean State in the MATLAB Distributed Computing Server System Administrators Guide.
8-20
8-21
You set properties on the scheduler object to specify Where the job data is stored That the workers should access job data directly in a shared file system The MATLAB root for the workers to use
set(sched, 'DataLocation', '\\share\scratch\jobdata') set(sched, 'HasSharedFilesystem', true) set(sched, 'ClusterMatlabRoot', '\\apps\matlab\')
Alternatively, you can use a parallel configuration to find the scheduler and set the object properties with a single findResource statement. If DataLocation is not set, the default location for job data is the current working directory of the MATLAB client the first time you use findResource to create an object for this type of scheduler. All settable property values on a scheduler object are local to the MATLAB client, and are lost when you close the client session or when you remove the object from the client workspace with delete or clear all. Note In a shared file system, all nodes require access to the directory specified in the scheduler objects DataLocation directory. See the DataLocation reference page for information on setting this property for a mixed-platform environment.
8-22
You can look at all the property settings on the scheduler object. If no jobs are in the DataLocation directory, the Jobs property is a 0-by-1 array.
get(sched) Configuration: '' Type: 'lsf' DataLocation: '\\share\scratch\jobdata' HasSharedFilesystem: 1 Jobs: [0x1 double] ClusterMatlabRoot: '\\apps\matlab\' ClusterOsType: 'unix' UserData: [] ClusterSize: Inf ClusterName: 'CENTER_MATRIX_CLUSTER' MasterName: 'masterhost.clusternet.ourdomain.com' SubmitArguments: '' ParallelSubmissionWrapperScript: [1x92 char]
You set properties on the scheduler object to specify Where the job data is stored The MATLAB root for the workers to use The name of the scheduler host Cluster version, and whether to use SOA job submission (available only on Microsoft Windows HPC Server 2008).
set(sched, 'DataLocation', '\\share\scratch\jobdata'); set(sched, 'ClusterMatlabRoot', '\\apps\matlab\'); set(sched, 'SchedulerHostname', 'server04');
8-23
Alternatively, you can use a parallel configuration to find the scheduler and set the object properties with a single findResource statement. If DataLocation is not set, the default location for job data is the current working directory of the MATLAB client the first time you use findResource to create an object for this type of scheduler. All settable property values on a scheduler object are local to the MATLAB client, and are lost when you close the client session or when you remove the object from the client workspace with delete or clear all. Note Because Windows HPC Server requires a shared file system, all nodes require access to the directory specified in the scheduler objects DataLocation directory. You can look at all the property settings on the scheduler object. If no jobs are in the DataLocation directory, the Jobs property is a 0-by-1 array.
get(sched) Configuration: Type: DataLocation: HasSharedFilesystem: Jobs: ClusterMatlabRoot: ClusterOsType: UserData: ClusterSize: SchedulerHostname: UseSOAJobSubmission: JobTemplate: JobDescriptionFile: ClusterVersion: '' 'hpcserver' '\\share\scratch\jobdata' 1 [0x1 double] '\\apps\matlab\' 'pc' [] Inf 'server04' 0 '' '' 'HPCServer2008'
8-24
Create a Job
You create a job with the createJob function, which creates a job object in the client session. The job data is stored in the directory specified by the scheduler objects DataLocation property.
j = createJob(sched)
This statement creates the job object j in the client session. Use get to see the properties of this job object.
get(j) Configuration: Name: ID: UserName: Tag: State: CreateTime: SubmitTime: StartTime: FinishTime: Tasks: FileDependencies: PathDependencies: JobData: Parent: UserData: '' 'Job1' 1 'eng1' '' 'pending' 'Fri Jul 29 16:15:47 EDT 2005' '' '' '' [0x1 double] {0x1 cell} {0x1 cell} [] [1x1 distcomp.lsfscheduler] []
This output varies only slightly between jobs that use LSF and Windows HPC Server schedulers, but is quite different from a job that uses a job manager. For example, jobs on LSF or Windows HPC Server schedulers have no callback functions. The jobs State property is pending. This state means the job has not been queued for running yet. This new job has no tasks, so its Tasks property is a 0-by-1 array. The schedulers Jobs property is now a 1-by-1 array of distcomp.simplejob objects, indicating the existence of your job.
get(sched, 'Jobs')
8-25
You can transfer files to the worker by using the FileDependencies property of the job object. Workers can access shared files by using the PathDependencies property of the job object. For details, see the FileDependencies and PathDependencies reference pages and Share Code on page 8-29. Note In a shared file system, MATLAB clients on many computers can access the same job data on the network. Properties of a particular job or task should be set from only one computer at a time.
Create Tasks
After you have created your job, you can create tasks for the job. Tasks define the functions to be evaluated by the workers during the running of the job. Often, the tasks of a job are all identical except for different arguments or data. In this example, each task will generate a 3-by-3 matrix of random numbers.
createTask(j, createTask(j, createTask(j, createTask(j, createTask(j, @rand, @rand, @rand, @rand, @rand, 1, 1, 1, 1, 1, {3,3}); {3,3}); {3,3}); {3,3}); {3,3});
Alternatively, you can create the five tasks with one call to createTask by providing a cell array of five cell arrays defining the input arguments to each task.
T = createTask(job1, @rand, 1, {{3,3} {3,3} {3,3} {3,3} {3,3}});
8-26
The scheduler distributes the tasks of job j to MATLAB workers for evaluation. For each task, the scheduler starts a MATLAB worker session on a worker node; this MATLAB worker session runs for only as long as it takes to evaluate the one task. If the same node evaluates another task in the same job, it does so with a different MATLAB worker session. Each worker performs the following steps for task evaluation:
1 Receive FileDependencies and PathDependencies from the job. Place
FileDependencies or PathDependencies.
3 Run the taskStartup function. You can specify this function in
FileDependencies or PathDependencies.
If you have enabled UseSOAJobSubmission with HPC Server 2008, the scheduler can use a worker to evaluate multiple tasks in sequence. In this case, the worker runs taskStartup before evaluating each task, without rerunning jobStartup or receiving dependencies again.
4 If the worker is part of forming a new MATLAB pool, run the poolStartup function. (This occurs when executing matlabpool open or when running
property.
7 Run the taskFinish function.
8-27
The job runs asynchronously with the MATLAB client. If you need to wait for the job to complete before you continue in your MATLAB client session, you can use the waitForState function.
waitForState(j)
The default state to wait for is finished. This function causes MATLAB to pause until the State property of j is 'finished'. Note When you use an LSF scheduler in a nonshared file system, the scheduler might report that a job is in the finished state even though the LSF scheduler might not yet have completed transferring the jobs files.
8-28
Share Code
Because different machines evaluate the tasks of a job, each machine must have access to all the files needed to evaluate its tasks. The following sections explain the basic mechanisms for sharing data: Access Files Directly on page 8-29 Pass Data Between Sessions on page 8-30 Pass MATLAB Code for Startup and Finish on page 8-30
These files can be passed to the worker by the jobs FileDependencies or PathDependencies property. Otherwise, the version of each of these files that is used is the one highest on the workers path.
8-29
8-30
taskStartup.m automatically executes on a worker each time the worker begins evaluation of a task. poolStartup.m automatically executes on a worker each time the worker is included in a newly started MATLAB pool. taskFinish.m automatically executes on a worker each time the worker completes evaluation of a task. Empty versions of these files are provided in the directory
matlabroot/toolbox/distcomp/user
You can edit these files to include whatever MATLAB code you want the worker to execute at the indicated times. Alternatively, you can create your own versions of these files and pass them to the job as part of the FileDependencies property, or include the pathnames to their locations in the PathDependencies property. The worker gives precedence to the versions provided in the FileDependencies property, then to those pointed to in the PathDependencies property. If any of these files is not included in these properties, the worker uses the version of the file in the toolbox/distcomp/user directory of the workers MATLAB installation.
Manage Objects
Objects that the client session uses to interact with the scheduler are only references to data that is actually contained in the directory specified by the DataLocation property. After jobs and tasks are created, you can shut down your client session, restart it, and your job will still be stored in that remote location. You can find existing jobs using the Jobs property of the recreated scheduler object. The following sections describe how to access these objects and how to permanently remove them: What Happens When the Client Session Ends? on page 8-32 Recover Objects on page 8-32
8-31
Recover Objects
A client session of Parallel Computing Toolbox software can access any of the objects in the DataLocation, whether the current client session or another client session created these objects. You create scheduler objects in the client session by using the findResource function.
sched = findResource('scheduler', 'type', 'LSF'); set(sched, 'DataLocation', '/share/scratch/jobdata');
When you have access to the scheduler by the object sched, you can create objects that reference all the data contained in the specified location for that scheduler. All the job and task data contained in the scheduler data location are accessible in the scheduler objects Jobs property, which is an array of job objects.
all_jobs = get(sched, 'Jobs')
You can index through the array all_jobs to locate a specific job. Alternatively, you can use the findJob function to search in a scheduler object for a particular job identified by any of its properties, such as its State.
finished_jobs = findJob(sched, 'State', 'finished')
8-32
This command returns an array of job objects that reference all finished jobs on the scheduler sched, whose data is found in the specified DataLocation.
Destroy Jobs
Jobs in the scheduler continue to exist even after they are finished. From the command line in the MATLAB client session, you can call the destroy function for any job object. If you destroy a job, you destroy all tasks contained in that job. The job and task data is deleted from the DataLocation directory. For example, find and destroy all finished jobs in your scheduler whose data is stored in a specific directory.
sched = findResource('scheduler', 'name', 'LSF'); set(sched, 'DataLocation', '/share/scratch/jobdata'); finished_jobs = findJob(sched, 'State', 'finished'); destroy(finished_jobs); clear finished_jobs
The destroy function in this example permanently removes from the scheduler data those finished jobs whose data is in /apps/data/project_88. The clear function removes the object references from the local MATLAB client workspace.
8-33
Overview
Parallel Computing Toolbox software provides a generic interface that lets you interact with third-party schedulers, or use your own scripts for distributing tasks to other nodes on the cluster for evaluation. Because each job in your application is comprised of several tasks, the purpose of your scheduler is to allocate a cluster node for the evaluation of each task, or to distribute each task to a cluster node. The scheduler starts remote MATLAB worker sessions on the cluster nodes to evaluate individual tasks of the job. To evaluate its task, a MATLAB worker session needs access to certain information, such as where to find the job and task data. The generic scheduler interface provides a means of getting tasks from your Parallel Computing Toolbox client session to your scheduler and thereby to your cluster nodes. To evaluate a task, a worker requires five parameters that you must pass from the client to the worker. The parameters can be passed any way you want to transfer them, but because a particular one must be an environment variable, the examples in this section pass all parameters as environment variables.
8-34
Worker node
Environment variables
MATLAB worker
Decode function
Scheduler
Note Whereas a MathWorks job manager keeps MATLAB workers running between tasks, a third-party scheduler runs MATLAB workers for only as long as it takes each worker to evaluate its one task.
where sched is the scheduler object in the client session, created with the findResource function. In this case, the submit function gets called with its three default arguments: scheduler, job, and properties object, in that order. The function declaration line of the function might look like this:
function mysubmitfunc(scheduler, job, props)
Inside the function of this example, the three argument objects are known as scheduler, job, and props. You can write a submit function that accepts more than the three default arguments, and then pass those extra arguments by including them in the definition of the SubmitFcn property.
8-35
In this example, the submit function requires five arguments: the three defaults, along with the numeric value of time_limit and the string value of testlocation. The functions declaration line might look like this:
function mysubmitfunc(scheduler, job, props, localtimeout, plant)
The following discussion focuses primarily on the minimum requirements of the submit and decode functions. This submit function has three main purposes: To identify the decode function that MATLAB workers run when they start To make information about job and task data locations available to the workers via their decode function To instruct your scheduler how to start a MATLAB worker on the cluster for each task of your job
Scheduler
8-36
variable MDCE_DECODE_FUNCTION. The value of this variable is a string identifying the name of the decode function on the path of the MATLAB worker. Neither the decode function itself nor its name can be passed to the worker in a job or task property; the file must already exist before the worker starts. For more information on the decode function, see MATLAB Worker Decode Function on page 8-40. Standard decode functions for distributed and parallel jobs are provided with the product. If your submit functions make use of the definitions in these decode functions, you do not have to provide your own decode functions. For example, to use the standard decode function for distributed jobs, in your submit function set MDCE_DECODE_FUNCTION to 'parallel.cluster.generic.distributedDecodeFcn'.
Description String. Used internally to indicate that a file system is used to contain job and task data. String. Derived from the scheduler DataLocation property. String. Indicates where this jobs data is stored. Cell array. Indicates where each tasks data is stored. Each element of this array is passed to a separate worker. Double. Indicates the number of tasks in the job. You do not need to pass this value to the worker, but you can use it within your submit function.
NumberOfTasks
8-37
With these values passed into your submit function, the function can pass them to the worker nodes by any of several means. However, because the name of the decode function must be passed as an environment variable, the examples that follow pass all the other necessary property values also as environment variables. The submit function writes the values of these object properties out to environment variables with the setenv function.
Scheduler Condor
Scheduler Command
condor_submit
Your submit function might also use some of these properties and others when constructing and invoking your scheduler command. scheduler, job, and props (so named only for this example) refer to the first three arguments to the submit function.
8-38
Argument Object
scheduler scheduler job job props
Property
MatlabCommandToRun ClusterMatlabRoot MinimumNumberOfWorkers MaximumNumberOfWorkers NumberOfTasks
passed into the submit function as its first three input arguments: the scheduler object, the job object, and the props object.
function mysubmitfunc(scheduler, job, props)
This example function uses only the three default arguments. You can have additional arguments passed into your submit function, as discussed in MATLAB Client Submit Function on page 8-35.
2 Identify the values you want to send to your environment variables. For
The name of the decode function that must be available on the MATLAB worker path is mydecodefunc.
3 Set the environment variables, other than the task locations. All the
MATLAB workers use these values when evaluating tasks of the job.
8-39
Your submit function can use any names you choose for the environment variables, with the exception of MDCE_DECODE_FUNCTION; the MATLAB worker looks for its decode function identified by this variable. If you use alternative names for the other environment variables, be sure that the corresponding decode function also uses your alternative variable names. You can see the variable names used in the standard decode function by typing
edit parallel.cluster.generic.distributedDecodeFcn
4 Set the task-specific variables and scheduler commands. This is where you
The line constructSchedulerCommand represents the code you write to construct and execute your schedulers submit command. This command is typically a string that combines the scheduler command with necessary flags, arguments, and values derived from the values of your object properties. This command is inside the for-loop so that your scheduler gets a command to start a MATLAB worker on the cluster for each task. Note If you are not familiar with your network scheduler, ask your system administrator for help.
8-40
submit function illustrated only the usage of environment variables, so does this discussion of the decode function. When working with the decode function, you must be aware of the Name and location of the decode function itself Names of the environment variables this function must read
Worker node
Environment variables MDCE_DECODE_FUNCTION MDCE_STORAGE_CONSTRUCTOR MDCE_STORAGE_LOCATION MDCE_JOB_LOCATION MDCE_TASK_LOCATION getenv
MATLAB worker
Decode function
Scheduler
matlab...
Note Standard decode functions are now included in the product. If your submit functions make use of the definitions in these decode functions, you do not have to provide your own decode functions. For example, to use the standard decode function for distributed jobs, in your submit function set MDCE_DECODE_FUNCTION to 'parallel.cluster.generic.distributedDecodeFcn'. The remainder of this section is useful only if you use names and settings other than the standards used in the provided decode functions.
8-41
client node to the worker node. Your scheduler might perform this task for you automatically; if it does not, you must arrange for this copying. The value of the environment variable MDCE_DECODE_FUNCTION defines the filename of the decode function, but not its location. The file cannot be passed as part of the job PathDependencies or FileDependencies property, because the function runs in the MATLAB worker before that session has access to the job. Therefore, the file location must be available to the MATLAB worker as that worker starts. Note The decode function must be available on the MATLAB workers path. You can get the decode function on the workers path by either moving the file into a directory on the path (for example, matlabroot/toolbox/local), or by having the scheduler use cd in its command so that it starts the MATLAB worker from within the directory that contains the decode function. In practice, the decode function might be identical for all workers on the cluster. In this case, all workers can use the same decode function file if it is accessible on a shared drive. When a MATLAB worker starts, it automatically runs the file identified by the MDCE_DECODE_FUNCTION environment variable. This decode function runs before the worker does any processing of its task.
8-42
When the object is returned from the decode function to the MATLAB worker session, its values are used internally for managing job and task data.
8-43
Generic schedulers must use a shared file system for workers to access job and task data. Set the DataLocation and HasSharedFilesystem properties to specify where the job data is stored and that the workers should access job data directly in a shared file system.
set(sched, 'DataLocation', '\\share\scratch\jobdata') set(sched, 'HasSharedFilesystem', true)
Note All nodes require access to the directory specified in the scheduler objects DataLocation directory. See the DataLocation reference page for information on setting this property for a mixed-platform environment. If DataLocation is not set, the default location for job data is the current working directory of the MATLAB client the first time you use findResource to create an object for this type of scheduler, which might not be accessible to the worker nodes. If MATLAB is not on the workers system path, set the ClusterMatlabRoot property to specify where the workers are to find the MATLAB installation.
set(sched, 'ClusterMatlabRoot', '\\apps\matlab\')
You can look at all the property settings on the scheduler object. If no jobs are in the DataLocation directory, the Jobs property is a 0-by-1 array. All settable property values on a scheduler object are local to the MATLAB client, and are lost when you close the client session or when you remove the object from the client workspace with delete or clear all.
get(sched) Configuration: Type: DataLocation: HasSharedFilesystem: Jobs: ClusterMatlabRoot: ClusterOsType: UserData: '' 'generic' '\\share\scratch\jobdata' 1 [0x1 double] '\\apps\matlab\' 'pc' []
8-44
Inf 'worker' [] []
You must set the SubmitFcn property to specify the submit function for this scheduler.
set(sched, 'SubmitFcn', @mysubmitfunc)
With the scheduler object and the user-defined submit and decode functions defined, programming and running a job is now similar to doing so with a job manager or any other type of scheduler.
2. Create a Job
You create a job with the createJob function, which creates a job object in the client session. The job data is stored in the directory specified by the scheduler objects DataLocation property.
j = createJob(sched)
This statement creates the job object j in the client session. Use get to see the properties of this job object.
get(j) Configuration: Name: ID: UserName: Tag: State: CreateTime: SubmitTime: StartTime: FinishTime: Tasks: FileDependencies: PathDependencies: JobData: Parent: '' 'Job1' 1 'neo' '' 'pending' 'Fri Jan 20 16:15:47 EDT 2006' '' '' '' [0x1 double] {0x1 cell} {0x1 cell} [] [1x1 distcomp.genericscheduler]
8-45
UserData: []
Note Properties of a particular job or task should be set from only one computer at a time. This generic scheduler job has somewhat different properties than a job that uses a job manager. For example, this job has no callback functions. The jobs State property is pending. This state means the job has not been queued for running yet. This new job has no tasks, so its Tasks property is a 0-by-1 array. The schedulers Jobs property is now a 1-by-1 array of distcomp.simplejob objects, indicating the existence of your job.
get(sched) Configuration: Type: DataLocation: HasSharedFilesystem: Jobs: ClusterMatlabRoot: ClusterOsType: UserData: ClusterSize: MatlabCommandToRun: SubmitFcn: ParallelSubmitFcn: '' 'generic' '\\share\scratch\jobdata' 1 [1x1 distcomp.simplejob] '\\apps\matlab\' 'pc' [] Inf 'worker' @mysubmitfunc []
3. Create Tasks
After you have created your job, you can create tasks for the job. Tasks define the functions to be evaluated by the workers during the running of the job. Often, the tasks of a job are identical except for different arguments or data. In this example, each task generates a 3-by-3 matrix of random numbers.
createTask(j, @rand, 1, {3,3}); createTask(j, @rand, 1, {3,3}); createTask(j, @rand, 1, {3,3});
8-46
Alternatively, you can create the five tasks with one call to createTask by providing a cell array of five cell arrays defining the input arguments to each task.
T = createTask(job1, @rand, 1, {{3,3} {3,3} {3,3} {3,3} {3,3}});
The scheduler distributes the tasks of j to MATLAB workers for evaluation. The job runs asynchronously. If you need to wait for it to complete before you continue in your MATLAB client session, you can use the waitForState function.
waitForState(j)
The default state to wait for is finished or failed. This function pauses MATLAB until the State property of j is 'finished' or 'failed'.
8-47
In this folder are subdirectories for each of several types of scheduler. Depending on your network and cluster configuration, you might need to modify these files before they will work in your situation. Ask your system administrator for help. At the time of publication, there are folders for Condor (condor), PBS (pbs), and Platform LSF (lsf) schedulers, generic UNIX-based scripts (ssh), Sun Grid Engine (sge), and mpiexec on Microsoft Windows operating systems (winmpiexec). In addition, the pbs, lsf, and sge folders have subfolders called
8-48
particular cluster configurations. Each of these subfolders contains a file called README, which provides instruction on where and how to use its scripts. For each scheduler type, the folder (or configuration subfolder) contains wrappers, submit functions, and other job management scripts for for distributed and parallel jobs. For example, the directory
matlabroot/toolbox/distcomp/examples/integration/pbs/shared
contains the following files for use with a PBS scheduler: Filename
distributedSubmitFcn.m parallelSubmitFcn.m
Description Submit function for a distributed job Submit function for a parallel job workers that evaluate the tasks of a distributed job
Script that is submitted to PBS to start labs that evaluate the tasks of a parallel job Script to destroy a job from the scheduler Script to get the jobs ID from the scheduler Script to get the jobs state from the scheduler Script to get the submission string for the scheduler
These files are all programmed to use the standard decode functions provided with the product, so they do not have specialized decode functions. The folder for other scheduler types contain similar files. As more files or solutions for more schedulers might become available at any time, visit the support page for this product on the MathWorks Web site at
http://www.mathworks.com/support/product/product.html?product=DM.
This Web page also provides contact information in case you have any questions.
8-49
If your scheduler accepts only submissions of individual tasks, you might get return data pertaining to only each individual tasks. In this case, your submit function might have code structured like this:
for ii = 1:props.NumberOfTasks submit task to scheduler %Per-task settings: data_array(1,ii) = ... parse string returned from scheduler
8-50
data_array(2,ii) = ... save ID returned from scheduler etc end setJobSchedulerData(scheduler, job, data_array)
In a similar way, you can define what do to for destroying a job, and what to do for canceling and destroying tasks.
8-51
While it is running or queued, you can cancel or destroy the job or a task. This command cancels the task and moves it to the finished state, and triggers execution of myCancelTask, which sends the appropriate commands to the scheduler:
cancel(t(4))
This command deletes job data for j1, and triggers execution of myDestroyJob, which sends the appropriate commands to the scheduler:
destroy(j1)
8-52
Summary
The following list summarizes the sequence of events that occur when running a job that uses the generic scheduler interface:
1 Provide a submit function and a decode function. Be sure the decode
function is on all the MATLAB workers paths. The following steps occur in the MATLAB client session:
2 Define the SubmitFcn property of your scheduler object to point to the
submit function.
3 Send your job to the scheduler.
submit(job)
4 The client session runs the submit function. 5 The submit function sets environment variables with values derived from
its arguments.
8-53
6 The submit function makes calls to the scheduler generally, a call for
each task (with environment variables identified explicitly, if necessary). The following step occurs in your network:
7 For each task, the scheduler starts a MATLAB worker session on a cluster
on the path.
9 The decode function reads the pertinent environment variables. 10 The decode function sets the properties of its argument object with values
8-54
9
Program Parallel Jobs
Parallel jobs are those in which the workers (or labs) can communicate with each other during the evaluation of their tasks. The following sections describe how to program parallel jobs: Introduction on page 9-2 Use a Supported Scheduler on page 9-4 Use the Generic Scheduler Interface on page 9-8 Further Notes on Parallel Jobs on page 9-11
Introduction
A parallel job consists of only a single task that runs simultaneously on several workers, usually with different data. More specifically, the task is duplicated on each worker, so each worker can perform the task on a different set of data, or on a particular segment of a large data set. The workers can communicate with each other as each executes its task. In this configuration, workers are referred to as labs. In principle, creating and running parallel jobs is similar to programming distributed jobs:
1 Find a scheduler. 2 Create a parallel job. 3 Create a task. 4 Submit the job for running. For details about what each worker performs
for evaluating a task, see Submit a Job to the Job Queue on page 8-13.
5 Retrieve the results.
The differences between distributed jobs and parallel jobs are summarized in the following table. Distributed Job MATLAB sessions, called workers, perform the tasks but do not communicate with each other. You define any number of tasks in a job. Tasks need not run simultaneously. Tasks are distributed to workers as the workers become available, so a worker can perform several of the tasks in a job. Parallel Job MATLAB sessions, called labs, can communicate with each other during the running of their tasks. You define only one task in a job. Duplicates of that task run on all labs running the parallel job. Tasks run simultaneously, so you can run the job only on as many labs as are available at run time. The start of the job might be delayed until the required number of labs is available.
9-2
Introduction
A parallel job has only one task that runs simultaneously on every lab. The function that the task runs can take advantage of a labs awareness of how many labs are running the job, which lab this is among those running the job, and the features that allow labs to communicate with each other.
9-3
9-4
of these column sums are combined with the gplus function to calculate the total sum of the elements of the original magic square. The function for this example is shown below.
function total_sum = colsum if labindex == 1 % Send magic square to other labs A = labBroadcast(1,magic(numlabs)) else % Receive broadcast on other labs A = labBroadcast(1) end % Calculate sum of column identified by labindex for this lab column_sum = sum(A(:,labindex)) % Calculate total sum by combining column sum from all labs total_sum = gplus(column_sum)
This function is saved as the file colsum.m on the path of the MATLAB client. It will be sent to each lab by the jobs FileDependencies property. While this example has one lab create the magic square and broadcast it to the other labs, there are alternative methods of getting data to the labs. Each lab could create the matrix for itself. Alternatively, each lab could read its part of the data from a file on disk, the data could be passed in as an argument to the task function, or the data could be sent in a file contained in the jobs FileDependencies property. The solution to choose depends on your network configuration and the nature of the data.
9-5
where myconfig is the name of a user-defined configuration for the type of scheduler you are using. Any required differences for various scheduling options are controlled in the configuration. You can have one or more separate configurations for each type of scheduler. For complete details, see Parallel Configurations for Cluster Access on page 6-14. Create or modify configurations according to the instructions of your system administrator. When your scheduler object is defined, you create the job object with the createParallelJob function.
pjob = createParallelJob(sched);
The function file colsum.m (created in Code the Task Function on page 9-4) is on the MATLAB client path, but it has to be made available to the labs. One way to do this is with the jobs FileDependencies property, which can be set in the configuration you used, or by:
set(pjob, 'FileDependencies', {'colsum.m'})
Here you might also set other properties on the job, for example, setting the number of workers to use. Again, configurations might be useful in your particular situation, especially if most of your jobs require many of the same property settings. To run this example on four labs, you can established this in the configuration, or by the following client code:
set(pjob, 'MaximumNumberOfWorkers', 4) set(pjob, 'MinimumNumberOfWorkers', 4)
You create the jobs one task with the usual createTask function. In this example, the task returns only one argument from each lab, and there are no input arguments to the colsum function.
t = createTask(pjob, @colsum, 1, {})
9-6
Make the MATLAB client wait for the job to finish before collecting the results. The results consist of one value from each lab. The gplus function in the task shares data between the labs, so that each lab has the same result.
waitForState(pjob) results = getAllOutputArguments(pjob) results = [136] [136] [136] [136]
9-7
Introduction
This section discusses programming parallel jobs using the generic scheduler interface. This interface lets you execute jobs on your cluster with any scheduler you might have. The principles of using the generic scheduler interface for parallel jobs are the same as those for distributed jobs. The overview of the concepts and details of submit and decode functions for distributed jobs are discussed fully in Use the Generic Scheduler Interface on page 8-34 in the chapter on Programming Distributed Jobs.
defined in the configuration. Because the scheduler itself is often common to many users and applications, it is probably best to use a configuration for programming these properties. See Parallel Configurations for Cluster Access on page 6-14. Among the properties required for a parallel job is ParallelSubmitFcn. You can write your own parallel submit and decode functions, or use those come with the product for various schedulers and platforms; see the following section, Supplied Submit and Decode Functions on page 9-9.
9-8
3 Use createParallelJob to create a parallel job object for your scheduler. 4 Create a task, run the job, and retrieve the results as usual.
In this folder are subdirectories for each of several types of scheduler. Depending on your network and cluster configuration, you might need to modify these files before they will work in your situation. Ask your system administrator for help. At the time of publication, there are folders for Condor (condor), PBS (pbs), and Platform LSF (lsf) schedulers, generic UNIX-based scripts (ssh), Sun Grid Engine (sge), and mpiexec on Microsoft Windows operating systems (winmpiexec). In addition, the pbs, lsf, and sge folders have subfolders called shared, nonshared, and remoteSubmission, which contain scripts for use in particular cluster configurations. Each of these subfolders contains a file called README, which provides instruction on where and how to use its scripts. For each scheduler type, the folder (or configuration subfolder) contains wrappers, submit functions, and other job management scripts for for distributed and parallel jobs. For example, the directory
matlabroot/toolbox/distcomp/examples/integration/pbs/shared
contains the following files for use with a PBS scheduler: Filename
distributedSubmitFcn.m parallelSubmitFcn.m
Description Submit function for a distributed job Submit function for a parallel job workers that evaluate the tasks of a distributed job
9-9
Filename
parallelJobWrapper.sh destroyJobFcn.m extractJobId.m getJobStateFcn.m getSubmitString.m
Description Script that is submitted to PBS to start labs that evaluate the tasks of a parallel job Script to destroy a job from the scheduler Script to get the jobs ID from the scheduler Script to get the jobs state from the scheduler Script to get the submission string for the scheduler
These files are all programmed to use the standard decode functions provided with the product, so they do not have specialized decode functions. For parallel jobs, the standard decode function provided with the product is parallel.cluster.generic.parallelDecodeFcn. You can view the required variables in this file by typing
edit parallel.cluster.generic.parallelDecodeFcn
The folder for other scheduler types contain similar files. As more files or solutions for more schedulers might become available at any time, visit the support page for this product on the MathWorks Web site at
http://www.mathworks.com/support/product/product.html?product=DM.
This Web page also provides contact information in case you have any questions.
9-10
The reason this fails is because the gather function requires communication between all the labs across which the array is distributed. When the if statement limits execution to a single lab, the other labs required for execution of the function are not executing the statement. As an alternative, you can use gather itself to collect the data into the workspace of a single lab: assembled = gather(D, 1).
9-11
In another example, suppose you want to transfer data from every lab to the next lab on the right (defined as the next higher labindex). First you define for each lab what the labs on the left and right are.
from_lab_left = mod(labindex - 2, numlabs) + 1; to_lab_right = mod(labindex, numlabs) + 1;
The reason this code might fail is because, depending on the size of the data being transferred, the labSend function can block execution in a lab until the corresponding receiving lab executes its labReceive function. In this case, all the labs are attempting to send at the same time, and none are attempting to receive while labSend has them blocked. In other words, none of the labs get to their labReceive statements because they are all blocked at the labSend statement. To avoid this particular problem, you can use the labSendReceive function.
9-12
10
GPU Computing
When to Use a GPU for Matrix Operations on page 10-2 Using GPUArray on page 10-4 Execute MATLAB Code on a GPU on page 10-10 Identify and Select a GPU from Multiple GPUs on page 10-15 Executing CUDA or PTX Code on the GPU on page 10-17 GPU Characteristics and Limitations on page 10-25
10
GPU Computing
Capabilities
This chapter describes how to program MATLAB to use your computers graphics processing unit (GPU) for matrix operations. In many cases, execution in the GPU is faster than in the CPU, so the techniques described in this chapter might offer improved performance. Several options are available for using your GPU: Transferring data between the MATLAB workspace and the GPU Evaluating built-in functions on the GPU Running MATLAB code on the GPU Creating kernels from PTX files for execution on the GPU Choosing one of multiple GPU cards to use The particular workflows for these capabilities are described in the following sections of this chapter.
Requirements
The following are required for using the GPU with MATLAB: NVIDIA CUDA-enabled device with compute capability of 1.3 or greater The latest NVIDIA CUDA device driver Access from a MATLAB worker running on a Microsoft Windows operating system with a job manager as the scheduler, requires an NVIDIA Tesla Compute Cluster (TCC) driver with an NVIDIA Tesla card.
10-2
Demos
Demos showing the usage of the GPU are available in the Demos node under Parallel Computing Toolbox in the help browser. You can also access the product demos by entering the following command at the MATLAB prompt:
demo toolbox parallel
10-3
10
GPU Computing
Using GPUArray
In this section... Transfer Data Between Workspace and GPU on page 10-4 Create GPU Data Directly on page 10-5 Examine GPUArray Characteristics on page 10-7 Built-In Functions That Support GPUArray on page 10-7
10-4
Using GPUArray
Transfer Data of a Specified Precision. Create a matrix of double-precision random data in MATLAB, and then transfer the matrix as single-precision from MATLAB to the GPU:
X = rand(1000); G = gpuArray(single(X));
Construct an Array for Storing on the GPU. Construct a 100-by-100 matrix of uint32 ones and transfer it to the GPU. You can accomplish this with a single line of code:
G = gpuArray(ones(100, 'uint32'));
10-5
10
GPU Computing
The static constructors appear at the bottom of the output from this command. For help on any one of the constructors, type
help parallel.gpu.GPUArray/functionname
The default class of the data is double, so you do not have to specify it.
10-6
Using GPUArray
8192
Description Class of the underlying data in the array Indication if array data is real Length of vector or largest array dimension Number of dimensions in the array Size of array dimensions
10-7
10
GPU Computing
abs acos acosh acot acoth acsc acsch all any arrayfun asec asech asin asinh atan atan2 atanh bitand bitcmp bitor bitshift bitxor cast cat ceil chol classUnderlying colon complex conj
conv conv2 cos cosh cot coth csc csch ctranspose cumprod cumsum diag diff disp display dot double eig eps eq erf erfc erfcinv erfcx erfinv exp expm1 filter filter2 find
fft fft2 fix floor gamma gammaln gather ge gt horzcat hypot ifft ifft2 imag int16 int32 int64 int8 isempty isequal isequalwithequalnans isfinite isinf islogical isnan isreal ldivide le length
log repmat log10 reshape log1p round log2 sec logical sech lt sign lu sin max single meshgrid sinh min size minus sort mldivide sqrt mod subsasgn mrdivide subsindex mtimes subsref ndgrid sum ndims svd ne tan norm tanh not times numel transpose plot (and related) tril plus triu power uint16 prod uint32 rdivide uint64 real uint8 reallog uminus realpow uplus realsqrt vertcat rem
To get specific help on the overloaded functions, and to learn about any restrictions concerning their support for GPUArray objects, type:
help parallel.gpu.GPUArray/functionname
10-8
Using GPUArray
help parallel.gpu.GPUArray/lu
The following functions are not methods of the GPUArray class, but they do work with GPUArray data:
The whos command is instructive for showing where each variables data is stored.
whos Name G Ga Gb Gfft
Notice that all the arrays are stored on the GPU (GPUArray), except for G, which is the result of the gather function.
10-9
10
GPU Computing
Subsequent arguments provide inputs to the MATLAB function. These input arguments can be workspace data or GPUArray. If any of the input arguments is a GPUArray, the function executes on the GPU and returns a GPUArray. (If none of the inputs is GPUArray, then arrayfun executes in the CPU.) See the arrayfun reference page for descriptions of the available options.
10-10
The function performs only element-wise operations when applying a gain factor and offset to each element of the rawdata array. Create some nominal measurement:
meas = ones(1000)*3; % 1000-by-1000 matrix
The function allows the gain and offset to be arrays of the same size as rawdata, so that unique corrections can be applied to individual measurements. In a typical situation, you might keep the correction data on the GPU so that you do not have to transfer it for each application:
gn = gpuArray(rand(1000))/100 + 0.995; offs = gpuArray(rand(1000))/50 - 0.01;
This runs on the GPU because the input arguments gn and offs are already in GPU memory. Retrieve the corrected results from the GPU to the MATLAB workspace:
results = gather(corrected);
10-11
10
GPU Computing
abs acos acosh acot acoth acsc acsch asec asech asin asinh atan atan2 atanh bitand bitcmp bitor bitshift bitxor ceil complex conj cos cosh cot coth csc
csch double eps erf erfc erfcinv erfcx erfinv exp expm1 false fix floor gamma gammaln hypot imag Inf int32 isfinite isinf isnan log log2 log10 log1p logical
max min mod NaN pi rand randi randn real reallog realpow realsqrt rem round sec sech sign sin single sinh sqrt tan tanh true uint32 xor
Branching instructions:
break continue else elseif for if return while
10-12
generation:
randi randi() randi(IMAX, ...) randi([IMIN IMAX], ...) randi(..., 'single') randi(..., 'double') randi(..., 'int32') randi(..., 'uint32')
You do not specify the array size for random generation. Instead, the number of generated random values is determined by the sizes of the input variables to your function. In effect, there will be enough random number elements to satisfy the needs of any input or output variables. For example, suppose your function myfun.m contains the following code that includes generating and using the random matrix R:
function Y = myfun(X) R = rand(); Y = R.*X; end
If you use arrayfun to run this function with an input variable that is a GPUArray, the function runs on the GPU, where the number of random elements for R is determined by the size of X, so you do not need to specify it. The following code passes the GPUArray matrix G to myfun on the GPU.
G = 2*parallel.gpu.GPUArray.ones(4,4) H = arrayfun(@myfun, G)
10-13
10
GPU Computing
Because G is a 4-by-4 GPUArray, myfun generates 16 random value scalar elements for R, one for each calculation with an element of G.
10-14
Description The number of GPU cards in your computer Select which card to use, or see which card is selected and view its properties
gpuDeviceCount 2
2 With two devices, the first is the default. You can examine its properties
10-15
10
GPU Computing
SIMDWidth: TotalMemory: FreeMemory: MultiprocessorCount: ClockRateKHz: ComputeMode: GPUOverlapsTransfers: KernelExecutionTimeout: CanMapHostMemory: DeviceSupported: DeviceSelected:
view its properties to verify that it is the one you want. For example, this step chooses and views the second device (indexing is 1-based):
gpuDevice(2)
Note If you select a device that does not have sufficient compute capability, you get a warning and you will not be able to use that device.
10-16
10-17
10
GPU Computing
Even if the inputs are constants or variables for MATLAB workspace data, the output is GPUArray.
Because the output is a GPUArray, you can now perform other operations using this input or output data without further transfers between the MATLAB workspace and the GPU. When all your GPU computations are complete, gather your final result data into the MATLAB workspace:
o2 = feval(k, o1, i2); r1 = gather(o1); r2 = gather(o2);
the corresponding kernel object (k) in MATLAB has the following properties:
10-18
Therefore, to use the kernel object from this code with feval, you need to provide feval two input arguments (in addition to the kernel object), and you can use one output argument:
y = feval(k, x1, x2)
The input values x1 and x2 correspond to pInOut and c in the C function prototype. The output argument y corresponds to the value of pInOut in the C function prototype after the C kernel has executed. The following is a slightly more complicated example that shows a combination of const and non-const pointers:
void moreComplicated( const float * pIn, float * pInOut1, float * pInOut2 )
You can use feval on this codes kernel (k) with the syntax:
[y1, y2] = feval(k, x1, x2, x3)
The three input arguments x1, x2, and x3, correspond to the three arguments that are passed into the C function. The output arguments y1 and y2, correspond to the values of pInOut1 and pInOut2 after the C kernel has executed.
10-19
10
GPU Computing
parallel.gpu.CUDAKernel handle Package: parallel.gpu Properties: ThreadBlockSize: MaxThreadsPerBlock: GridSize: SharedMemorySize: EntryPoint: MaxNumLHSArguments: NumRHSArguments: ArgumentTypes:
The properties of a kernel object control some of its execution behavior. Use dot notation to alter those properties that can be changed. For a descriptions of the object properties, see the CUDAKernel object reference page.
A single PTX file can contain multiple entry points to different kernels. Each of these entry points has a unique name. These names are generally mangled (as in C++ mangling). However, when generated by nvcc the PTX name always contains the original function name from the CU. For example, if the CU file defines the kernel function as
__global__ void simplestKernelEver( float * x, float val )
then the PTX code contains an entry that might be called _Z18simplestKernelEverPff. When you have multiple entry points, specify the entry name for the particular kernel when calling CUDAKernel to generate your kernel.
10-20
Note The CUDAKernel function searches for your entry name in the PTX file, and matches on any substring occurrences. Therefore, you should not name any of your entries as substrings of any others.
In parsing C prototype, the supported C data types are listed in the following table. Float Types
double, double2 float, float2
Integer Types
short, unsigned short, short2, ushort2 int, unsigned int, int2, uint2 long, unsigned long, long2, ulong2 long long, unsigned long long, longlong2, ulonglong2
All inputs can be scalars or pointers, and can be labeled const. The C declaration of a kernel is always of the form:
__global__ void aKernel(inputs ...)
The kernel must return nothing, and operate only on its input arguments (scalars or pointers).
10-21
10
GPU Computing
A kernel is unable to allocate any form of memory, so all outputs must be pre-allocated before the kernel is executed. Therefore, the sizes of all outputs must be known before you run the kernel. In principle, all pointers passed into the kernel that are not const could contain output data, since the many threads of the kernel could modify that data. When translating the definition of a kernel in C into MATLAB: All scalar inputs in C (double, float, int, etc.) must be scalars in MATLAB, or scalar (i.e., single-element) GPUArray data. They are passed (after being cast into the requested type) directly to the kernel as scalars. All const pointer inputs in C (const double *, etc.) can be scalars or matrices in MATLAB. They are cast to the correct type, copied onto the card, and a pointer to the first element is passed to the kernel. No information about the original size is passed to the kernel. It is as though the kernel has directly received the result of mxGetData on an mxArray. All nonconstant pointer inputs in C are transferred to the kernel exactly as nonconstant pointers. However, because a nonconstant pointer could be changed by the kernel, this will be considered as an output from the kernel. These rules have some implications. The most notable is that every output from a kernel must necessarily also be an input to the kernel, since the input allows the user to define the size of the output (which follows from being unable to allocate memory on the GPU).
10-22
*pi += c; }
The directive __global__ indicates that this is an entry point to a kernel. The code uses a pointer to send out the result in pi, which is both an input and an output. Put this code in a file called test.cu in the current directory.
2 Compile the CU code at the shell command line to generate a PTX file
called test.ptx.
nvcc -ptx test.cu
3 Create the kernel in MATLAB. Currently this PTX file only has one entry
so you do not need to specify it. If you were to put more kernels in, you would specify add1 as the entry.
k = parallel.gpu.CUDAKernel('test.ptx', 'test.cu');
4 Run the kernel with two inputs of 1. By default, a kernel runs on one
thread.
>> o = feval(k, 1, 1); o = 2
pointers, and one is constant because you are not changing it. Each thread will simply add the elements at its thread index. The thread index must work out which element this thread should add. (Getting these thread- and block-specific values is a very common pattern in CUDA programming.)
__global__ void add2( double * v1, const double * v2 ) { int idx = threadIdx.x;
10-23
10
GPU Computing
v1[idx] += v2[idx]; }
10-24
Data Types
Code in a function passed to arrayfun for execution on the GPU can use only these GPU native data types: single, double, int32, uint32, and logical. The overloaded functions for GPUArrays support these types where appropriate. GPUArrays also support the storing of data types in addition to these. This allows a GPUArray to be used with kernels written for these alternative data types, such as int8, uint8, etc.
Complex Numbers
If the output of a function running on the GPU could potentially be complex, you must explicitly specify its input arguments as complex. This applies to gpuArray or to functions called in code run by arrayfun. For example, if creating a GPUArray which might have negative elements, use G = gpuArray(complex(p)), then you can successfully execute sqrt(G). Or, within a function passed to arrayfun, if x is a vector of real numbers, and some elements have negative values, sqrt(x) will generate an error; instead you should call sqrt(complex(x)). The following table lists the functions that might return complex data, along with the input range over which the output remains real. Function
acos(x) acosh(x) acoth(x)
10-25
10
GPU Computing
Function
acsc(x) asec(x) asech(x) asin(x) atanh log(x) log1p(x) log10(x) log2(x) power(x,y) reallog(x) realsqrt(x) sqrt(x)
10-26
11
Object Reference
Data (p. 11-2) Schedulers (p. 11-2) Generic Scheduler Interface Tools (p. 11-3) Jobs (p. 11-3) Tasks (p. 11-3) Workers (p. 11-4) Graphics Processing Unit (p. 11-4) Representing data on multiple resources Representing job manager, local scheduler, or third-party scheduler Access to remote clusters Representing different types of jobs Representing different types of tasks Representing MATLAB worker sessions Representing GPU devices, arrays, code
11
Object Reference
Data
codistributed codistributor1d codistributor2dbc Composite distributed GPUArray Access data of arrays distributed among workers in MATLAB pool 1-D distribution scheme for codistributed array 2-D block-cyclic distribution scheme for codistributed array Access nondistributed data on multiple labs from client Access data of distributed arrays from client Array of data stored on Graphics Processing Unit (GPU)
Schedulers
ccsscheduler genericscheduler jobmanager localscheduler lsfscheduler mpiexec pbsproscheduler torquescheduler Access Microsoft Windows HPC Server scheduler Access generic scheduler Control job queue and execution Access local scheduler on client machine Access Platform LSF scheduler Directly access mpiexec for job distribution Access PBS Pro scheduler Access TORQUE scheduler
11-2
Jobs
job matlabpooljob Define job behavior and properties when using job manager Define MATLAB pool job behavior and properties when using job manager Define parallel job behavior and properties when using job manager Define job behavior and properties when using local or third-party scheduler Define MATLAB pool job behavior and properties when using local or third-party scheduler Define parallel job behavior and properties when using local or third-party scheduler
paralleljob simplejob
simplematlabpooljob
simpleparalleljob
Tasks
simpletask Define task behavior and properties when using local or third-party scheduler Define task behavior and properties when using job manager
task
11-3
11
Object Reference
Workers
worker Access information about MATLAB worker session
11-4
12
Objects Alphabetical List
ccsscheduler
Parent Children
None
simplejob and simpleparalleljob objects
A ccsscheduler object provides access to your networks Windows HPC Server (including CCS) scheduler, which controls the job queue, and distributes job tasks to workers or labs for execution.
Methods
Create job object in scheduler and client Create MATLAB pool job Create parallel job object Find job objects stored in scheduler Read output messages from job run by supported third-party or local scheduler
Properties
ClusterMatlabRoot ClusterOsType
Specify MATLAB root for cluster Specify operating system of nodes on which scheduler will start workers Number of workers available to scheduler Version of HPC Server scheduler
ClusterSize ClusterVersion
12-2
ccsscheduler
Specify configuration to apply to object or toolbox function Specify directory where job data is stored Specify whether nodes share data location Name of XML job description file for Microsoft Windows HPC Server scheduler Jobs contained in job manager service or in schedulers data location Name of job template for HPC Server 2008 scheduler Name of host running Microsoft Windows HPC Server scheduler Type of scheduler object Specify data to associate with object Allow service-oriented architecture (SOA) submission on HPC Server 2008 cluster
Jobs
See Also
12-3
codistributed
Data of distributed arrays that exist on the labs are accessible from the other labs as codistributed array objects. Codistributed arrays on labs that you create inside spmd statements can be accessed via distributed arrays on the client.
Methods
Create codistributed cell array Distributed colon operation Create codistributed identity matrix Create codistributed false array Create codistributed array of Inf values Create codistributed array of Not-a-Number values Create codistributed array of ones Create codistributed array of uniformly distributed pseudo-random numbers Create codistributed array of normally distributed random values Allocate space for sparse codistributed matrix Create codistributed sparse identity matrix
codistributed.randn
codistributed.spalloc codistributed.speye
12-4
codistributed
codistributed.sprand
Create codistributed sparse array of uniformly distributed pseudo-random values Create codistributed sparse array of uniformly distributed pseudo-random values Create codistributed true array Create codistributed array of zeros Transfer distributed array data or GPUArray to local workspace Codistributor object for existing codistributed array Local portion of codistributed array Global indices for local part of codistributed array True if distributed arrays underlying elements are of specified class True for codistributed array Redistribute codistributed array with another distribution scheme Create sparse distributed or codistributed matrix
codistributed.sprandn
12-5
codistributor1d
A codistributor1d object defines the 1-D distribution scheme for a codistributed array. The 1-D codistributor distributes arrays along a single specified dimension, the distribution dimension, in a noncyclic, partitioned manner.
Methods
codistributor1d.defaultPartition
Global indices for local part of codistributed array True if codistributor object is complete
Properties
Dimension Partition
12-6
codistributor2dbc
A codistributor2dbc object defines the 2-D block-cyclic distribution scheme for a codistributed array. The 2-D block-cyclic codistributor can only distribute two-dimensional matrices. It distributes matrices along two subscripts over a rectangular computational grid of labs in a blocked, cyclic manner. The parallel matrix computation software library called ScaLAPACK uses the 2-D block-cyclic codistributor.
Methods
codistributor2dbc.defaultLabGrid
Global indices for local part of codistributed array True if codistributor object is complete
Properties
BlockSize
Block size of codistributor2dbc object Default block size for codistributor2dbc distribution scheme
codistributor2dbc.defaultBlockSize
LabGrid Orientation
12-7
Composite
Variables that exist on the labs running an spmd statement are accessible on the client as a Composite object. A Composite resembles a cell array with one element for each lab. So for Composite C:
C{1} represents value of C on lab1 C{2} represents value of C on lab2 etc. spmd statements create Composites automatically, which you can
access after the statement completes. You can also create a Composite explicitly with the Composite function.
Methods
Check whether Composite is defined on labs Subscripted assignment for Composite Subscripted reference for Composite
Other methods of a Composite object behave similarly to these MATLAB array functions:
disp, display end isempty length ndims
Display Composite Indicate last Composite index Determine whether Composite is empty Length of Composite Number of Composite dimensions
12-8
Composite
numel size
12-9
CUDAKernel
A CUDAKernel object represents a CUDA kernel, that can execute on a GPU. You create the kernel when you compile PTX or CU code, as described in Executing CUDA or PTX Code on the GPU on page 10-17.
Methods Properties
Property Name
ThreadBlockSize
feval
A CUDAKernel object has the following properties: Description Size of block of threads on the kernel. This can be an integer vector of length 1, 2, or 3 (since thread blocks can be up to 3-dimensional). The product of the elements of ThreadBlockSize must not exceed the MaxThreadsPerBlock for this kernel, and no element of ThreadBlockSize can exceed the corresponding element of the gpuDevice property MaxThreadBlockSize. Maximum number of threads permissible in a single block for this CUDA kernel. The product of the elements of ThreadBlockSize must not exceed this value. Size of grid (effectively the number of thread blocks that will be launched independently by the GPU). This is an integer vector of length 1 or 2. Neither element of this vector can exceed the corresponding element in the vector of the MaxGridSize property of the GPUDevice object.
MaxThreadsPerBlock
GridSize
12-10
CUDAKernel
Property Name
SharedMemorySize
Description The amount of dynamic shared memory (in bytes) that each thread block can use. Each thread block has an available shared memory region. The size of this region is limited in current cards to ~16 kB, and is shared with registers on the multiprocessors. As with all memory, this needs to be allocated before the kernel is launched. It is also common for the size of this shared memory region to be tied to the size of the thread block. Setting this value on the kernel ensures that each thread in a block can access this available shared memory region. (read-only) A string containing the actual entry point name in the PTX code that this kernel is going to call. An example might look like '_Z13returnPointerPKfPy'. (read-only) The maximum number of left hand side arguments that this kernel supports. It cannot be greater than the number of right hand side arguments, and if any inputs are constant or scalar it will be less. (read-only) The required number of right hand side arguments needed to call this kernel. All inputs need to define either the scalar value of an input, the data for a vector input/output, or the size of an output argument. (read-only) Cell array of strings, the same length as NumRHSArguments. Each of the strings indicates what the expected MATLAB type for that input is (a numeric type such as uint8, single, or double followed by the word scalar or vector to indicate if we are passing by reference or value). In addition, if that argument is only an input to the kernel, it is prefixed by in; and if it is an input/output, it is prefixed by inout. This allows you to decide how to efficiently call the kernel with both MATLAB data and GPUArray, and to see which of the kernel inputs are being treated as outputs.
GPUArray, GPUDevice
EntryPoint
MaxNumLHSArguments
NumRHSArguments
ArgumentTypes
See Also
12-11
distributed
Data of distributed arrays that exist on the labs are accessible on the client as a distributed array. A distributed array resembles a normal array in the way you access and manipulate its elements, but none of its data exists on the client. Codistributed arrays that you create inside spmd statements are accessible via distributed arrays on the client. You can also create a distributed array explicitly on the client with the distributed function.
Methods
Create distributed cell array Create distributed identity matrix Create distributed false array Create distributed array of Inf values Create distributed array of Not-a-Number values Create distributed array of ones Create distributed array of uniformly distributed pseudo-random numbers Create distributed array of normally distributed random values Allocate space for sparse distributed matrix
distributed.randn
distributed.spalloc
12-12
distributed
distributed.speye distributed.sprand
Create distributed sparse identity matrix Create distributed sparse array of uniformly distributed pseudo-random values Create distributed sparse array of normally distributed pseudo-random values Create distributed true array Create distributed array of zeros Transfer distributed array data or GPUArray to local workspace True if distributed arrays underlying elements are of specified class True for distributed array Create sparse distributed or codistributed matrix
distributed.sprandn
isdistributed sparse
12-13
genericscheduler
Parent Children
None
simplejob and simpleparalleljob objects
A genericscheduler object provides access to your networks scheduler, which distributes job tasks to workers or labs for execution. The generic scheduler interface requires use of the MATLAB code submit function on the client and the MATLAB code decode function on the worker node.
Methods
Create job object in scheduler and client Create MATLAB pool job Create parallel job object Find job objects stored in scheduler Get specific user data for job on generic scheduler Set specific user data for job on generic scheduler
Properties
CancelJobFcn CancelTaskFcn
Specify function to run when canceling job on generic scheduler Specify function to run when canceling task on generic scheduler
12-14
genericscheduler
ClusterMatlabRoot ClusterOsType
Specify MATLAB root for cluster Specify operating system of nodes on which scheduler will start workers Number of workers available to scheduler Specify configuration to apply to object or toolbox function Specify directory where job data is stored Specify function to run when destroying job on generic scheduler Specify function to run when destroying task on generic scheduler Specify function to run when querying job state on generic scheduler Specify whether nodes share data location Jobs contained in job manager service or in schedulers data location MATLAB command that generic scheduler runs to start lab Specify function to run when parallel job submitted to generic scheduler Specify function to run when job submitted to generic scheduler
DestroyTaskFcn
GetJobStateFcn
HasSharedFilesystem Jobs
MatlabCommandToRun ParallelSubmitFcn
SubmitFcn
12-15
genericscheduler
Type UserData
See Also
12-16
GPUArray
Purpose Constructor
GPUArray with data stored on the GPU device. Also, the following static methods create GPUArray data:
You can get help on any of these methods with the command
help parallel.gpu.GPUArray.methodname
where methodname is the name of the method. For example, to get help on ones, type
help parallel.gpu.GPUArray.ones
Description
A GPUArray object represents an array of data stored on the GPU. You can use the data for direct calculations, or in CUDA kernels that execute on the GPU. You can return data to the MATLAB workspace with the gather function. The methods for a GPUArray object are too numerous to list here. Most resemble and behave the same as built-in MATLAB functions. See Using GPUArray on page 10-4. For the complete list, use the methods function on the GPUArray class:
Methods
12-17
GPUArray
methods('parallel.gpu.GPUArray')
See Also
CUDAKernel, GPUDevice
12-18
GPUDevice
A GPUDevice object represents a graphic processing unit (GPU) in your computer. You can use the GPU to execute CUDA kernels or MATLAB code. The following convenience functions let you identify and select a GPU device:
gpuDevice gpuDeviceCount
Methods
Description True if the GPU specified by index idx is supported and capable of being selected. idx can be an integer or a vector of integers; the default index is the current device. Returns a GPUDevice object without selecting it.
parallel.gpu.GPUDevice.getDevice(idx)
For the complete list, use the methods function on the GPUDevice class:
methods('parallel.gpu.GPUDevice')
You can get help on any of the class methods with the command
help parallel.gpu.GPUDevice.methodname
where methodname is the name of the method. For example, to get help on isAvailable, type
12-19
GPUDevice
help parallel.gpu.GPUDevice.isAvailable
Properties
Property Name
Name Index
A GPUDevice object has the following read-only properties: Description Name of the CUDA device. Index by which you can select the device. Computational capability of the CUDA device. Must meet required specification. Indicates if this device can support double precision operations. The CUDA device driver version currently in use. Must meet required specification. Maximum supported number of threads per block during CUDAKernel execution. Maximum supported amount of shared memory that can be used by a thread block during CUDAKernel execution. Maximum size in each dimension for thread block. Each dimension of a thread block must not exceed these dimensions. Also, the product of the thread block size must not exceed MaxThreadsPerBlock. Maximum size of grid of thread blocks. Number of simultaneously executing threads. Total available memory (in bytes) on the device. Free memory (in bytes) on the device. This property is available only for the currently selected device, and has the value NaN for unselected devices. The number of vector processors present on the device. The total core count of the device is 8 times this property.
MaxThreadBlockSize
MultiprocessorCount
12-20
GPUDevice
Property Name
ClockRateKHz ComputeMode
Description Peak clock rate of the GPU in kHz. The compute mode of the device, according to the following values:
'Default' The device is not restricted and can support multiple CUDA sessions simultaneously. 'Exclusive' The device can be used by only one CUDA session. 'Prohibited' The device cannot be used.
GPUOverlapsTransfers KernelExecutionTimeout
Indicates if the device supports overlapped transfers. Indicates if the device can abort long-running kernels. If true, the operating system places an upper bound on the time allowed for the CUDA kernel to execute, after which the CUDA driver times out the kernel and returns an error. Indicates if the device supports mapping host memory into the CUDA address space. Indicates if toolbox can use this this device. Not all devices are supported; for example, if their ComputeCapability is insufficient, the toolbox cannot use them. Indicates if this is the currently selected device.
CanMapHostMemory DeviceSupported
DeviceSelected
See Also
CUDAKernel, GPUArray
12-21
job
Parent Children
A job object contains all the tasks that define what each worker does as part of the complete job execution. A job object is used only with a job manager as scheduler.
Methods
Cancel job or task Create new task in job Remove job or task object from parent and memory Display or save Command Window text of batch job Task objects belonging to job object Output arguments from evaluation of all tasks in job object Load workspace variables from batch job Queue job in scheduler Wait for job to finish or change state Wait for object to change state
12-22
job
Properties
AuthorizedUsers Configuration CreateTime FileDependencies FinishedFcn FinishTime ID JobData MaximumNumberOfWorkers MinimumNumberOfWorkers Name Parent PathDependencies QueuedFcn
Specify users authorized to access job Specify configuration to apply to object or toolbox function When task or job was created Directories and files that worker can access Specify callback to execute after task or job runs When task or job finished Object identifier Data made available to all workers for jobs tasks Specify maximum number of workers to perform job tasks Specify minimum number of workers to perform job tasks Name of job manager, job, or worker object Parent object of job or task Specify directories to add to MATLAB worker path Specify function file to execute when job is submitted to job manager queue Specify whether to restart MATLAB workers before evaluating job tasks
RestartWorker
12-23
job
Specify function file to execute when job or task starts running When job or task started Current state of task, job, job manager, or worker When job was submitted to queue Specify label to associate with job object Tasks contained in job object Specify time limit to complete task or job Specify data to associate with object User who created job or job manager object
See Also
12-24
jobmanager
Parent Children
None
job, paralleljob, and worker objects
A jobmanager object provides access to the job manager, which controls the job queue, distributes job tasks to workers or labs for execution, and maintains job results. The job manager is provided with the MATLAB Distributed Computing Server product, and its use as a scheduler is optional.
Methods
changePassword clearLocalPassword createJob createMatlabPoolJob createParallelJob demote findJob pause promote resume
Prompt user to change job manager password Delete local store of users job manager password Create job object in scheduler and client Create MATLAB pool job Create parallel job object Demote job in job manager queue Find job objects stored in scheduler Pause job manager queue Promote job in job manager queue Resume processing queue in job manager
12-25
jobmanager
Properties
BusyWorkers ClusterOsType
Workers currently running tasks Specify operating system of nodes on which scheduler will start workers Number of workers available to scheduler Specify configuration to apply to object or toolbox function IP address of host running job manager or worker session Name of host running job manager or worker session Idle workers available to run tasks True if job manager and workers use secure communication Jobs contained in job manager service or in schedulers data location Name of job manager, job, or worker object Number of workers currently running tasks Number of idle workers available to run tasks Specify if system should prompt for password when authenticating user Security level controlling access to job manager and its jobs
SecurityLevel
12-26
jobmanager
Current state of task, job, job manager, or worker Type of scheduler object Specify data to associate with object User who created job or job manager object
See Also
12-27
localscheduler
Parent Children
None
simplejob and simpleparalleljob objects
A localscheduler object provides access to your client machines local scheduler, which controls the job queue, and distributes job tasks to workers or labs for execution on the client machine.
Methods
Create job object in scheduler and client Create MATLAB pool job Create parallel job object Find job objects stored in scheduler Read output messages from job run by supported third-party or local scheduler
Properties
ClusterMatlabRoot ClusterOsType
Specify MATLAB root for cluster Specify operating system of nodes on which scheduler will start workers Number of workers available to scheduler
ClusterSize
12-28
localscheduler
Specify configuration to apply to object or toolbox function Specify directory where job data is stored Specify whether nodes share data location Jobs contained in job manager service or in schedulers data location Type of scheduler object Specify data to associate with object
Type UserData
See Also
jobmanager
12-29
lsfscheduler
Parent Children
None
simplejob and simpleparalleljob objects
An lsfscheduler object provides access to your networks Platform LSF scheduler, which controls the job queue, and distributes job tasks to workers or labs for execution.
Methods
Create job object in scheduler and client Create MATLAB pool job Create parallel job object Find job objects stored in scheduler Read output messages from job run by supported third-party or local scheduler Set options for submitting parallel jobs to scheduler
setupForParallelExecution
Properties
Specify MATLAB root for cluster Name of Platform LSF cluster Specify operating system of nodes on which scheduler will start workers
12-30
lsfscheduler
Number of workers available to scheduler Specify configuration to apply to object or toolbox function Specify directory where job data is stored Specify whether nodes share data location Jobs contained in job manager service or in schedulers data location Name of Platform LSF master node start labs
MasterName
Specify additional arguments to use when submitting job to Platform LSF, PBS Pro, TORQUE, or mpiexec scheduler Type of scheduler object Specify data to associate with object
Type UserData
See Also
12-31
matlabpooljob
Define MATLAB pool job behavior and properties when using job manager
createMatlabPoolJob
Parent Children
A matlabpooljob object contains all the information needed to define what each lab does as part of the complete job execution. A MATLAB pool job uses one worker in a MATLAB pool to run a parallel job on the other labs of the pool. A matlabpooljob object is used only with a job manager as scheduler.
Methods
Cancel job or task Create new task in job Remove job or task object from parent and memory Display or save Command Window text of batch job Task objects belonging to job object Output arguments from evaluation of all tasks in job object Load workspace variables from batch job Queue job in scheduler
load submit
12-32
matlabpooljob
wait waitForState
Wait for job to finish or change state Wait for object to change state
Properties
AuthorizedUsers Configuration CreateTime FileDependencies FinishedFcn FinishTime ID JobData MaximumNumberOfWorkers MinimumNumberOfWorkers Name Parent PathDependencies
Specify users authorized to access job Specify configuration to apply to object or toolbox function When task or job was created Directories and files that worker can access Specify callback to execute after task or job runs When task or job finished Object identifier Data made available to all workers for jobs tasks Specify maximum number of workers to perform job tasks Specify minimum number of workers to perform job tasks Name of job manager, job, or worker object Parent object of job or task Specify directories to add to MATLAB worker path
12-33
matlabpooljob
QueuedFcn
Specify function file to execute when job is submitted to job manager queue Specify whether to restart MATLAB workers before evaluating job tasks Specify function file to execute when job or task starts running When job or task started Current state of task, job, job manager, or worker When job was submitted to queue Specify label to associate with job object First task contained in MATLAB pool job object Tasks contained in job object Specify time limit to complete task or job Specify data to associate with object User who created job or job manager object
RestartWorker
RunningFcn StartTime State SubmitTime Tag Task Tasks Timeout UserData UserName
See Also
12-34
mpiexec
Parent Children
None
simplejob and simpleparalleljob objects
An mpiexec object provides direct access to the mpiexec executable for distribution of a jobs tasks to workers or labs for execution.
Methods
Create job object in scheduler and client Create MATLAB pool job Create parallel job object Find job objects stored in scheduler Read output messages from job run by supported third-party or local scheduler
Properties
ClusterMatlabRoot ClusterOsType
Specify MATLAB root for cluster Specify operating system of nodes on which scheduler will start workers Number of workers available to scheduler Specify configuration to apply to object or toolbox function
ClusterSize Configuration
12-35
mpiexec
DataLocation EnvironmentSetMethod
Specify directory where job data is stored Specify means of setting environment variables for mpiexec scheduler Specify whether nodes share data location Jobs contained in job manager service or in schedulers data location Specify pathname of executable mpiexec command Specify additional arguments to use when submitting job to Platform LSF, PBS Pro, TORQUE, or mpiexec scheduler Type of scheduler object Specify data to associate with object Specify operating system of nodes on which mpiexec scheduler will start labs
HasSharedFilesystem Jobs
MpiexecFileName SubmitArguments
See Also
12-36
paralleljob
Define parallel job behavior and properties when using job manager
createParallelJob
Parent Children
A paralleljob object contains all the tasks that define what each lab does as part of the complete job execution. A parallel job runs simultaneously on all labs and uses communication among the labs during task evaluation. A paralleljob object is used only with a job manager as scheduler.
Methods
Cancel job or task Create new task in job Remove job or task object from parent and memory Display or save Command Window text of batch job Task objects belonging to job object Output arguments from evaluation of all tasks in job object Load workspace variables from batch job Queue job in scheduler
load submit
12-37
paralleljob
wait waitForState
Wait for job to finish or change state Wait for object to change state
Properties
AuthorizedUsers Configuration CreateTime FileDependencies FinishedFcn FinishTime ID JobData MaximumNumberOfWorkers MinimumNumberOfWorkers Name Parent PathDependencies
Specify users authorized to access job Specify configuration to apply to object or toolbox function When task or job was created Directories and files that worker can access Specify callback to execute after task or job runs When task or job finished Object identifier Data made available to all workers for jobs tasks Specify maximum number of workers to perform job tasks Specify minimum number of workers to perform job tasks Name of job manager, job, or worker object Parent object of job or task Specify directories to add to MATLAB worker path
12-38
paralleljob
QueuedFcn
Specify function file to execute when job is submitted to job manager queue Specify whether to restart MATLAB workers before evaluating job tasks Specify function file to execute when job or task starts running When job or task started Current state of task, job, job manager, or worker When job was submitted to queue Specify label to associate with job object Tasks contained in job object Specify time limit to complete task or job Specify data to associate with object User who created job or job manager object
RestartWorker
See Also
12-39
pbsproscheduler
Parent Children
None
simplejob and simpleparalleljob objects
A pbsproscheduler object provides access to your networks PBS Pro scheduler, which controls the job queue, and distributes job tasks to workers or labs for execution.
Methods
Create job object in scheduler and client Create MATLAB pool job Create parallel job object Find job objects stored in scheduler Read output messages from job run by supported third-party or local scheduler Set options for submitting parallel jobs to scheduler
setupForParallelExecution
Properties
ClusterMatlabRoot ClusterOsType
Specify MATLAB root for cluster Specify operating system of nodes on which scheduler will start workers
12-40
pbsproscheduler
Number of workers available to scheduler Specify configuration to apply to object or toolbox function Specify directory where job data is stored Specify whether nodes share data location Jobs contained in job manager service or in schedulers data location start labs
Command to copy files from client Resource definition for PBS Pro or TORQUE scheduler Remote execution command used on worker nodes during parallel job Name of current PBS Pro or TORQUE server machine Specify additional arguments to use when submitting job to Platform LSF, PBS Pro, TORQUE, or mpiexec scheduler Type of scheduler object Specify data to associate with object
ServerName SubmitArguments
Type UserData
See Also
12-41
RemoteClusterAccess
Purpose Constructor
Description
behavior of the connection. The accepted parameters are: 'IdentityFilename' A string containing the full path to the identity file to use when connecting to a remote host. If 'IdentityFilename' is not specified, you are prompted for a password when establishing the connection. 'IdentityFileHasPassphrase' A boolean indicating whether or not the identity file requires a passphrase. If true, you are prompted for a password when establishing a connection. If an identity file is not supplied, this property is ignored. This value is false by default. For more information and detailed examples, see the integration scripts provided in matlabroot/toolbox/distcomp/examples/integration. For example, the scripts for PBS in a nonshared file system are in
matlabroot/toolbox/distcomp/examples/integration/pbs/nonshared
12-42
RemoteClusterAccess
Methods
Method Name
connect
Description
RemoteClusterAccess.connect(clusterHost)
establishes a connection to the specified host using the user credential options supplied in the constructor. File mirroring is not supported.
RemoteClusterAccess.connect(clusterHost, remoteDataLocation) establishes a connection to the
specified host using the user credential options supplied in the constructor. remoteDataLocation identifies a folder on the clusterHost that is used for file mirroring. The user credentials supplied in the constructor must have write access to this folder.
disconnect RemoteClusterAccess.disconnect() disconnects the existing remote connection. The connect method must
performs a final copy of changed files from the remote DataLocation to the local DataLocation for the supplied job. Any running mirrors for the job also stop and the job files are removed from the remote DataLocation. The startMirrorForJob or resumeMirrorForJob method must have already been called.
getRemoteJobLocation RemoteClusterAccess.getRemoteJobLocation(jobID, remoteOS) returns the full path to the remote job location for the supplied jobID. Valid values for remoteOS are 'pc' and 'unix'. RemoteClusterAccess.isJobUsingConnection(jobID)
isJobUsingConnection
12-43
RemoteClusterAccess
Method Name
resumeMirrorForJob
Description
RemoteClusterAccess.resumeMirrorForJob(job)
resumes the mirroring of files from the remote DataLocation to the local DataLocation for the supplied job. This is similar to the startMirrorForJob method, but does not first copy the files from the local DataLocation to the remote DataLocation. The connect method must have already been called. This is useful if the original client MATLAB session has ended, and you are accessing the same files from a new client session.
runCommand [status, result] = RemoteClusterAccess.runCommand(command) runs the
supplied command on the remote host and returns the resulting status and standard output. The connect method must have already been called.
startMirrorForJob RemoteClusterAccess.startMirrorForJob(job)
copies all the job files from the local DataLocation to the remote DataLocation, and starts mirroring files so that any changes to the files in the remote DataLocation are copied back to the local DataLocation. The connect method must have already been called.
stopMirrorForJob RemoteClusterAccess.stopMirrorForJob(job)
immediately stops the mirroring of files from the remote DataLocation to the local DataLocation for the specified job. The startMirrorForJob or resumeMirrorForJob method must have already been called. This cancels the running mirror and removes the files for the job from the remote location. This is similar to doLastMirrorForJob, except that stopMirrorForJob makes no attempt to ensure that the local job files are up to date. For normal mirror stoppage, use doLastMirrorForJob.
12-44
RemoteClusterAccess
Properties
A RemoteClusterAccess object has the following read-only properties. Their values are set when you construct the object or call its connect method. Description Location on the remote host for files that are being mirrored. Name of the remote host to access. Indicates if the identity file requires a passphrase. Full path to the identity file used when connecting to the remote host. Indicates if there is an active connection to the remote host. Indicates if file mirroring is supported for this connection. This is false if no remote DataLocation is supplied to the connect() method. Indicates if an identity file should be used when connecting to the remote host. User name for connecting to the remote host. Mirror files from the remote data location. Assume the object job represents a job on your generic scheduler.
remoteConnection = parallel.cluster.RemoteClusterAccess('testname'); remoteConnection.connect('headnode1','/tmp/filemirror'); remoteConnection.startMirrorForJob(job); submit(job) % Wait for the job to finish job.wait(); % Ensure that all the local files are up to date, and remove the % remote files remoteConnection.doLastMirrorForJob(job);
Property Name
DataLocation Hostname IdentityFileHasPassphrase IdentityFilename IsConnected IsFileMirrorSupported
UseIdentityFile Username
Examples
12-45
RemoteClusterAccess
For more detailed examples, see the integration scripts provided in matlabroot/toolbox/distcomp/examples/integration. For example, the scripts for PBS in a nonshared file system are in
matlabroot/toolbox/distcomp/examples/integration/pbs/nonshared
12-46
simplejob
Define job behavior and properties when using local or third-party scheduler
createJob
Parent
ccsscheduler, genericscheduler, localscheduler, lsfscheduler, mpiexec, pbsproscheduler, or torquescheduler object simpletask objects
Children
Description
A simplejob object contains all the tasks that define what each worker does as part of the complete job execution. A simplejob object is used only with a local or third-party scheduler.
Methods
Cancel job or task Create new task in job Remove job or task object from parent and memory Display or save Command Window text of batch job Task objects belonging to job object Output arguments from evaluation of all tasks in job object Load workspace variables from batch job Queue job in scheduler
load submit
12-47
simplejob
wait waitForState
Wait for job to finish or change state Wait for object to change state
Properties
Configuration CreateTime FileDependencies FinishTime ID JobData Name Parent PathDependencies StartTime State SubmitTime Tag Tasks
Specify configuration to apply to object or toolbox function When task or job was created Directories and files that worker can access When task or job finished Object identifier Data made available to all workers for jobs tasks Name of job manager, job, or worker object Parent object of job or task Specify directories to add to MATLAB worker path When job or task started Current state of task, job, job manager, or worker When job was submitted to queue Specify label to associate with job object Tasks contained in job object
12-48
simplejob
UserData UserName
Specify data to associate with object User who created job or job manager object
See Also
12-49
simplematlabpooljob
Define MATLAB pool job behavior and properties when using local or third-party scheduler
createMatlabPoolJob
Parent
ccsscheduler, genericscheduler, localscheduler, lsfscheduler, mpiexec, pbsproscheduler, or torquescheduler object simpletask object
Children
Description
A simplematlabpooljob object contains all the information needed to define what each lab does as part of the complete job execution. A MATLAB pool job uses one worker in a MATLAB pool to run a parallel job on the other labs of the pool. A simplematlabpooljob object is used only with a local or third-party scheduler.
Methods
Cancel job or task Create new task in job Remove job or task object from parent and memory Display or save Command Window text of batch job Task objects belonging to job object Output arguments from evaluation of all tasks in job object Load workspace variables from batch job Queue job in scheduler
load submit
12-50
simplematlabpooljob
wait waitForState
Wait for job to finish or change state Wait for object to change state
Properties
Configuration CreateTime FileDependencies FinishTime ID JobData MaximumNumberOfWorkers MinimumNumberOfWorkers Name Parent PathDependencies StartTime State SubmitTime Tag
Specify configuration to apply to object or toolbox function When task or job was created Directories and files that worker can access When task or job finished Object identifier Data made available to all workers for jobs tasks Specify maximum number of workers to perform job tasks Specify minimum number of workers to perform job tasks Name of job manager, job, or worker object Parent object of job or task Specify directories to add to MATLAB worker path When job or task started Current state of task, job, job manager, or worker When job was submitted to queue Specify label to associate with job object
12-51
simplematlabpooljob
First task contained in MATLAB pool job object Tasks contained in job object Specify data to associate with object User who created job or job manager object
See Also
12-52
simpleparalleljob
Define parallel job behavior and properties when using local or third-party scheduler
createParallelJob
Parent
ccsscheduler, genericscheduler, localscheduler, lsfscheduler, mpiexec, pbsproscheduler, or torquescheduler object simpletask objects
Children
Description
A simpleparalleljob object contains all the tasks that define what each lab does as part of the complete job execution. A parallel job runs simultaneously on all labs and uses communication among the labs during task evaluation. A simpleparalleljob object is used only with a local or third-party scheduler.
Methods
Cancel job or task Create new task in job Remove job or task object from parent and memory Display or save Command Window text of batch job Task objects belonging to job object Output arguments from evaluation of all tasks in job object Load workspace variables from batch job Queue job in scheduler
load submit
12-53
simpleparalleljob
wait waitForState
Wait for job to finish or change state Wait for object to change state
Properties
Configuration CreateTime FileDependencies FinishTime ID JobData MaximumNumberOfWorkers MinimumNumberOfWorkers Name Parent PathDependencies StartTime State SubmitTime Tag
Specify configuration to apply to object or toolbox function When task or job was created Directories and files that worker can access When task or job finished Object identifier Data made available to all workers for jobs tasks Specify maximum number of workers to perform job tasks Specify minimum number of workers to perform job tasks Name of job manager, job, or worker object Parent object of job or task Specify directories to add to MATLAB worker path When job or task started Current state of task, job, job manager, or worker When job was submitted to queue Specify label to associate with job object
12-54
simpleparalleljob
Tasks contained in job object Specify data to associate with object User who created job or job manager object
See Also
12-55
simpletask
Define task behavior and properties when using local or third-party scheduler
createTask
Parent Children
None
Description
A simpletask object defines what each lab or worker does as part of the complete job execution. A simpletask object is used only with a local or third-party scheduler.
Methods
Cancel job or task Remove job or task object from parent and memory Wait for object to change state
Properties
Specify whether to return Command Window output Text produced by execution of task objects function Specify configuration to apply to object or toolbox function When task or job was created Task error information Task error identifier Message from task error
12-56
simpletask
FinishTime Function ID InputArguments Name NumberOfOutputArguments OutputArguments Parent StartTime State UserData
When task or job finished Function called when evaluating task Object identifier Input arguments to task object Name of job manager, job, or worker object Number of arguments returned by task function Data returned from execution of task Parent object of job or task When job or task started Current state of task, job, job manager, or worker Specify data to associate with object
See Also
task
12-57
task
Parent Children
None
A task object defines what each lab or worker does as part of the complete job execution. A task object is used only with a job manager as scheduler.
Methods
Cancel job or task Remove job or task object from parent and memory Wait for object to change state
Properties
Number of times failed task was rerun Specify whether to return Command Window output Text produced by execution of task objects function Specify configuration to apply to object or toolbox function When task or job was created Task error information Task error identifier
12-58
task
ErrorMessage FailedAttemptInformation FinishedFcn FinishTime Function ID InputArguments MaximumNumberOfRetries NumberOfOutputArguments OutputArguments Parent RunningFcn StartTime State Timeout UserData Worker
Message from task error Information returned from failed task Specify callback to execute after task or job runs When task or job finished Function called when evaluating task Object identifier Input arguments to task object Specify maximum number of times to rerun failed task Number of arguments returned by task function Data returned from execution of task Parent object of job or task Specify function file to execute when job or task starts running When job or task started Current state of task, job, job manager, or worker Specify time limit to complete task or job Specify data to associate with object Worker session that performed task
12-59
task
See Also
simpletask
12-60
torquescheduler
Parent Children
None
simplejob and simpleparalleljob objects
A torquescheduler object provides access to your networks TORQUE scheduler, which controls the job queue, and distributes job tasks to workers or labs for execution.
Methods
Create job object in scheduler and client Create MATLAB pool job Create parallel job object Find job objects stored in scheduler Read output messages from job run by supported third-party or local scheduler Set options for submitting parallel jobs to scheduler
setupForParallelExecution
Properties
ClusterMatlabRoot ClusterOsType
Specify MATLAB root for cluster Specify operating system of nodes on which scheduler will start workers
12-61
torquescheduler
Number of workers available to scheduler Specify configuration to apply to object or toolbox function Specify directory where job data is stored Specify whether nodes share data location Jobs contained in job manager service or in schedulers data location start labs
Command to copy files from client Resource definition for PBS Pro or TORQUE scheduler Remote execution command used on worker nodes during parallel job Name of current PBS Pro or TORQUE server machine Specify additional arguments to use when submitting job to Platform LSF, PBS Pro, TORQUE, or mpiexec scheduler Type of scheduler object Specify data to associate with object
ServerName SubmitArguments
Type UserData
See Also
12-62
worker
Parent Children
jobmanager object
None
A worker object represents the MATLAB worker session that evaluates tasks in a job scheduled by a job manager. Only worker sessions started with the startworker script can be represented by a worker object. None
Methods Properties
Information about computer on which worker is running Job whose task this worker session is currently evaluating Task that worker is currently running IP address of host running job manager or worker session Name of host running job manager or worker session Job manager that this worker is registered with Name of job manager, job, or worker object Job whose task this worker previously ran
12-63
worker
PreviousTask State
Task that this worker previously ran Current state of task, job, job manager, or worker
See Also
12-64
13
Function Reference
Parallel Code Execution (p. 13-2) Distributed and Codistributed Arrays (p. 13-3) Jobs and Tasks (p. 13-6) Interlab Communication Within a Parallel Job (p. 13-9) Graphics Processing Unit (p. 13-11) Utilities (p. 13-11) Constructs for automatically running code in parallel Data partitioned across multiple MATLAB sessions Parallel computation through individual tasks Communications between labs during job execution Transferring data and running code on the GPU Utilities for using Parallel Computing Toolbox
13
Function Reference
Run MATLAB script or function as batch job Create Composite object Create distributed array from data in client workspace Open or close pool of MATLAB sessions for parallel computation Execute code loop in parallel Execute code in parallel on MATLAB pool
Default parallel computing configuration Display or save Command Window text of batch job Check whether Composite is defined on labs Import parallel configuration .mat file
13-2
Load workspace variables from batch job Run command on client and all workers in matlabpool Subscripted assignment for Composite Subscripted reference for Composite
Interactive Functions
mpiprofile pmode
Profile parallel communication and execution times Interactive Parallel Command Window
Toolbox Functions
codistributed codistributed.build codistributed.colon codistributor
Create codistributed array from replicated local data Create codistributed array from distributed data Distributed colon operation Create codistributor object for codistributed arrays
13-3
13
Function Reference
codistributor1d
Create 2-D block-cyclic codistributor object for codistributed arrays block-cyclic distributed arrays
codistributor2dbc.defaultLabGrid Default computational grid for 2-D for gather getCodistributor getLocalPart globalIndices isaUnderlying for-loop over distributed range
Transfer distributed array data or GPUArray to local workspace Codistributor object for existing codistributed array Local portion of codistributed array Global indices for local part of codistributed array True if distributed arrays underlying elements are of specified class True if codistributor object is complete True for replicated array Redistribute codistributed array with another distribution scheme
Create codistributed cell array Create codistributed identity matrix Create codistributed false array
13-4
Create codistributed array of Inf values Create codistributed array of Not-a-Number values Create codistributed array of ones Create codistributed array of uniformly distributed pseudo-random numbers Create codistributed array of normally distributed random values Allocate space for sparse codistributed matrix Create codistributed sparse identity matrix Create codistributed sparse array of uniformly distributed pseudo-random values Create codistributed sparse array of uniformly distributed pseudo-random values Create codistributed true array Create codistributed array of zeros Create distributed cell array Create distributed identity matrix Create distributed false array Create distributed array of Inf values Create distributed array of Not-a-Number values Create distributed array of ones Create distributed array of uniformly distributed pseudo-random numbers
codistributed.sprandn
13-5
13
Function Reference
Create distributed array of normally distributed random values Allocate space for sparse distributed matrix Create distributed sparse identity matrix Create distributed sparse array of uniformly distributed pseudo-random values Create distributed sparse array of normally distributed pseudo-random values Create distributed true array Create distributed array of zeros Create sparse distributed or codistributed matrix
distributed.sprandn
Job Creation
createJob createMatlabPoolJob
Create job object in scheduler and client Create MATLAB pool job
13-6
createParallelJob createTask dfeval dfevalasync findResource jobStartup mpiLibConf mpiSettings pctconfig poolStartup
Create parallel job object Create new task in job Evaluate function using cluster Evaluate function asynchronously using cluster Find available parallel computing resources File for user-defined options to run when job starts Location of MPI implementation Configure options for MPI communication Configure settings for Parallel Computing Toolbox client session File for user-defined options to run on each worker when MATLAB pool starts Set options for submitting parallel jobs to scheduler File for user-defined options to run when task finishes File for user-defined options to run when task starts
Job Management
cancel changePassword clearLocalPassword
Cancel job or task Prompt user to change job manager password Delete local store of users job manager password
13-7
13
Function Reference
Demote job in job manager queue Remove job or task object from parent and memory Find job objects stored in scheduler Task objects belonging to job object Output arguments from evaluation of all tasks in job object Read output messages from job run by supported third-party or local scheduler Get specific user data for job on generic scheduler Pause job manager queue Promote job in job manager queue Resume processing queue in job manager Set specific user data for job on generic scheduler Queue job in scheduler Wait for job to finish or change state Wait for object to change state
Job object whose task is currently being evaluated Job manager object that scheduled current task Task object currently being evaluated in this worker session
13-8
getCurrentWorker getFileDependencyDir
Worker object currently running this session Directory where FileDependencies are written on worker machine
Object Control
clear get inspect length methods set size
Remove objects from MATLAB workspace Object properties Open Property Inspector Length of object array List functions of object class Configure or display object properties Size of object array
Global concatenation Global operation across all labs Global addition Block execution until all labs reach this call Send data to all labs or receive data sent to all labs Index of this lab Test to see if messages are ready to be received from other lab Receive data from another lab
13-9
13
Function Reference
Send data to another lab Simultaneously send data to and receive data from another lab Total number of labs operating in parallel on current job Load file into parallel session Save data from parallel job session
13-10
Apply function to each element of array on GPU Evaluate kernel on GPU Transfer distributed array data or GPUArray to local workspace Create array on GPU Query or select GPU device Number of GPU devices present Create GPU CUDA kernel object from PTX and CU code
Utilities
help pctRunDeployedCleanup
Help for toolbox functions in Command Window Clean up after deployed parallel applications
13-11
13
Function Reference
13-12
14
Functions Alphabetical List
arrayfun
Purpose Syntax
Description
This method of a GPUArray object is very similar in behavior to the MATLAB function arrayfun, except that the actual evaluation of the function happens on the GPU, not on the CPU. Thus, any required data not already on the GPU is moved to GPU memory, the MATLAB function passed in for evaluation is compiled for the GPU, and then executed on the GPU. All the output arguments return as GPUArray objects, whose data can be retrieved with the gather method.
A = arrayfun(FUN, B) applies the function specified by FUN to each element of the GPUArray B, and returns the results in GPUArray A. A is the same size as B, and A(i,j,...) is equal to FUN(B(i,j,...)). FUN is a function handle to a function that takes one input argument and returns a scalar value. FUN must return values of the same class each time it is called. The input data must be a arrays of one of the following types: single, double, int32, uint32, logical, or GPUArray. The order in which arrayfun computes elements of A is not specified and should not be relied on. FUN must be a handle to a function that is written in the MATLAB
language (i.e., not a built-in function or a mex function); it must not be a nested, anonymous, or sub-function; and the MATLAB file that defines the function must contain exactly one function definition. The subset of the MATLAB language that is currently supported for execution on the GPU can be found in Execute MATLAB Code on a GPU on page 10-10.
A = arrayfun(FUN, B, C, ...) evaluates FUN using elements of arrays B, C, ... as input arguments. The resulting GPUArray element A(i,j,...) is equal to FUN(B(i,j,...), C(i,j,...), ...). The inputs B, C, ... must all have the same size or be scalar. Any scalar inputs are scalar expanded before being input to the function FUN.
14-2
arrayfun
One or more of the inputs B, C, ... must be a GPUArray; any of the others can reside in CPU memory. Each array that is held in CPU memory is converted to a GPUArray before calling the function on the GPU. If you plan to use an array in several different arrayfun calls, it is more efficient to convert that array to a GPUArray before making the series of calls to arrayfun.
[A, B, ...] = arrayfun(FUN, C, ...), where FUN is a function handle to a function that returns multiple outputs, returns GPUArrays A, B, ..., each corresponding to one of the output arguments of FUN. arrayfun calls FUN each time with as many outputs as there are in the call to arrayfun. FUN can return output arguments having different classes, but the class of each output must be the same each time FUN is called. This means that all elements of A must be the same class; B can be a different class from A, but all elements of B must be of the same class, etc.
Examples
Attr
14-3
arrayfun
Use gather to retrieve the data from the GPU to the MATLAB workspace.
d = gather(o2);
See Also
gather | gpuArray
14-4
batch
Purpose Syntax
Arguments
The batch job object. The script of MATLAB code to be evaluated by the MATLAB pool job. Scheduler object for allocating cluster compute resources. Function handle or string of function name to be evaluated by the MATLAB pool job. The number of output arguments from the evaluated function. Cell array of input arguments to the function. Object properties or other arguments to control job behavior. Initial values for corresponding object properties or arguments.
Description
j = batch('aScript') runs the script aScript.m on a worker according to the scheduler defined in the default parallel configuration. The function returns j, a handle to the job object that runs the script. The script file aScript.m is added to the FileDependencies and copied to the worker.
14-5
batch
except that the script runs on a worker according to the scheduler identified by schedobj.
j = batch(fcn, N, {x1, ..., xn}) runs the function specified by a function handle or function name, fcn, on a worker according to the
scheduler defined in the default parallel configuration. The function returns j, a handle to the job object that runs the function. The function is evaluated with the given arguments, x1, ...,xn, returning N output arguments. The function file for fcn is added to the FileDependencies and copied to the worker.
j = batch(schedobj, fcn, N, {x1, ..., xn}) is identical to batch(fcn, N, {x1, ..., xn}) except that the function runs on a worker according to the scheduler identified by schedobj. j = batch(..., 'p1', v1, 'p2', v2, ...) allows additional parameter-value pairs that modify the behavior of the job. These parameters support batch for functions and scripts, unless otherwise indicated. The accepted parameters are:
'Workspace' A 1-by-1 struct to define the workspace on the worker just before the script is called. The field names of the struct define the names of the variables, and the field values are assigned to the workspace variables. By default this parameter has a field for every variable in the current workspace where batch is executed. This parameter supports only the running of scripts. 'Configuration' A single string that is the name of a parallel configuration to use to find the correct cluster. By default it is the string returned from defaultParallelConfig. The configuration is applied to the properties of the job and tasks that are created for the batch. 'PathDependencies' A string or cell array of strings that defines paths to be added to the workers MATLAB path before the script or function is executed. See the PathDependencies reference page.
14-6
batch
'FileDependencies' A string or cell array of strings. Each string in the list identifies either a file or a directory, which is transferred to the worker. See the FileDependencies reference page. 'CurrentDirectory' A string indicating in what folder the script executes. There is no guarantee that this folder exists on the worker. The default value for this property is the cwd of MATLAB when the batch command is executed, unless any FileDependencies are defined. If the string for this argument is '.', there is no change in folder before batch execution. 'CaptureDiary' A boolean flag to indicate that the toolbox should collect the diary from the function call. See the diary function for information about the collected data. The default is true. 'Matlabpool' A nonnegative scalar integer that defines the number of labs to make into a MATLAB pool for the job to run on in addition to the worker running the batch. The script or function uses the pool for execution of statements such as parfor and spmd. A value of N for the property Matlabpool is effectively the same as adding a call to matlabpool N into the code. Because the MATLAB pool requires N workers in addition to the worker running the batch, there must be at least N+1 workers available on the cluster. (See Run a Batch Parallel Loop on page 1-8.) The default value is 0, which causes the script or function to run on only the single worker without a MATLAB pool.
Tips
As a matter of good programming practice, when you no longer need it, you should destroy the job created by the batch function so that it does not continue to consume cluster storage resources. Run a batch script on a worker without using a MATLAB pool:
j = batch('script1', 'matlabpool', 0);
Examples
Run a batch MATLAB pool job on a remote cluster, using eight workers for the MATLAB pool in addition to the worker running the batch script.
14-7
batch
Capture the diary, and load the results of the job into the workspace. This job requires a total of nine workers:
j = batch('script1', 'matlabpool', 8, 'CaptureDiary', true); wait(j); % Wait for the job to finish diary(j) % Display the diary load(j) % Load job workspace data into client workspace
Run a batch MATLAB pool job on a local worker, which employs two other local workers for the pool. Note, this runs a total of three worker sessions, in addition to the client, all on the local machine:
j = batch('script1', 'configuration', 'local', ... 'matlabpool', 2);
Clean up a batch jobs data after you are finished with it:
destroy(j)
Run a batch function on a remote cluster that generates a 10-by-10 random matrix:
sch = findResource('scheduler', ... 'configuration', defaultParallelConfig); j = batch(sch, @rand, 1, {10, 10}); wait(j) diary(j) % Wait for the job to finish % Display the diary
See Also
14-8
cancel
Purpose Syntax
Arguments
t j
Description
cancel(t) stops the task object, t, that is currently in the pending or running state. The tasks State property is set to finished, and no output arguments are returned. An error message stating that the task was canceled is placed in the task objects ErrorMessage property, and the worker session running the task is restarted. cancel(j) stops the job object, j, that is pending, queued, or running. The jobs State property is set to finished, and a cancel is executed on all tasks in the job that are not in the finished state. A job object that has been canceled cannot be started again.
If the job is running in a job manager, any worker sessions that are evaluating tasks belonging to the job object will be restarted.
Examples
Cancel a task. Note afterward the tasks State, ErrorMessage, and OutputArguments properties.
job1 = createJob(jm); t = createTask(job1, @rand, 1, {3,3}); cancel(t) get(t) ID: 1 Function: @rand NumberOfOutputArguments: 1 InputArguments: {[3] CaptureCommandWindowOutput: 0 CommandWindowOutput: '' [3]} OutputArguments: {1x0 cell}
14-9
cancel
State: 'finished' ErrorMessage: 'Task cancelled by user' ErrorIdentifier: 'distcomp:task:Cancelled' Timeout: Inf CreateTime: 'Fri Oct 22 11:38:39 EDT 2004' StartTime: 'Fri Oct 22 11:38:46 EDT 2004' FinishTime: 'Fri Oct 22 11:38:46 EDT 2004' Worker: [] Parent: [1x1 distcomp.job] UserData: [] RunningFcn: [] FinishedFcn: []
See Also
destroy | submit
14-10
changePassword
Purpose Syntax
Arguments
jm username
Job manager object on which password is changing User whose password is changed
Description
changePassword(jm) prompts the user to change the password for the current user. The users current password must be entered as well as the new password. changePassword(jm, username) prompts the job managers admin user to change the password for the specified user. The admin users password must be entered as well as the users new password. This enables the admin user to reset a password if the user has forgotten it.
For more information on job manager security, see Setting Job Manager Security.
See Also
clearLocalPassword
14-11
clear
obj
If obj references an object in the job manager, it is cleared from the workspace, but it remains in the job manager. You can restore obj to the workspace with the findResource, findJob, or findTask function; or with the Jobs or Tasks property. This example creates two job objects on the job manager jm. The variables for these job objects in the MATLAB workspace are job1 and job2. job1 is copied to a new variable, job1copy; then job1 and job2 are cleared from the MATLAB workspace. The job objects are then restored to the workspace from the job objects Jobs property as j1 and j2, and the first job in the job manager is shown to be identical to job1copy, while the second job is not.
job1 = createJob(jm); job2 = createJob(jm); job1copy = job1; clear job1 job2; j1 = jm.Jobs(1); j2 = jm.Jobs(2); isequal (job1copy, j1) ans = 1 isequal (job1copy, j2) ans = 0
Examples
See Also
14-12
clearLocalPassword
jm
local computer. When you call a privileged action (for example, with findResource, createJob, or submit) on a job manager running with a higher security level, a popup dialog requires you to enter your user name and password for authentication. To prevent having to enter it for every subsequent privileged action, you can choose to have the password remembered in a local password store. The clearLocalPassword function removes your password from this local store. This means any subsequent call to a privileged action on this computer requires you to re-authenticate with a valid password. Clearing your password might be useful after you have finished working on a shared machine. For more information on job manager security, see Setting Job Manager Security.
See Also
changePassword
14-13
codistributed
Purpose Syntax
Description
C = codistributed(X) distributes a replicated X using the default codistributor. X must be a replicated array, that is, it must have the same value on all labs. size(C) is the same as size(X). C = codistributed(X, codist) distributes a replicated X using the codistributor codist. X must be a replicated array, namely it must have the same value on all labs. size(C) is the same as size(X). For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. C = codistributed(X, codist, lab) distributes a local array X that resides on the lab identified by lab, using the codistributor codist. Local array X must be defined on all labs, but only the value from lab is used to construct C. size(C) is the same as size(X). C = codistributed(C1, codist) where the input array C1 is already a codistributed array, redistributes the array C1 according to the distribution scheme defined by codistributor codist. This is the same as calling C = redistribute(C1, codist). If the specified distribution scheme is that same as that already in effect, then the result is the same as the input.
Tips Examples
14-14
codistributed
end
Create a 1000-by-1000 codistributed array C2, distributed by rows (over its first dimension).
spmd N = 1000; X = magic(N); C2 = codistributed(X, codistributor1d(1)); end
See Also
14-15
codistributed.build
the local parts conform with the specified distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc.
D = codistributed.build(L, codist, 'noCommunication')
builds a codistributed array, without performing any interworker communications for error checking.
codist must be complete, which you can check by calling codist.isComplete(). The requirements on the size and structure of the local part L depend on the class of codist. For the 1-D and 2-D block-cyclic codistributors, L must have the same class and sparsity
on all labs. Furthermore, the local part L must represent the region described by the globalIndices method on codist.
Examples
Create a codistributed array of size 1001-by-1001 such that column ii contains the value ii.
spmd N = 1001; globalSize = [N, N]; % Distribute the matrix over the second dimension (columns), % and let the codistributor derive the partition from the % global size. codistr = codistributor1d(2, ... codistributor1d.unsetPartition, globalSize) % On 4 labs, codistr.Partition equals [251, 250, 250, 250]. % Allocate storage for the local part.
14-16
codistributed.build
localSize = [N, codistr.Partition(labindex)]; L = zeros(localSize); % Use globalIndices to map the indices of the columns % of the local part into the global column indices. globalInd = codistr.globalIndices(2); % On 4 labs, globalInd has the values: % 1:251 on lab 1 % 252:501 on lab 2 % 502:751 on lab 3 % 752:1001 on lab 4 % Initialize the columns of the local part to % the correct value. for localCol = 1:length(globalInd) globalCol = globalInd(localCol); L(:, localCol) = globalCol; end D = codistributed.build(L, codistr) end
See Also
14-17
codistributed.cell
Purpose Syntax
Description
codistributed array of underlying class cell, using a default scheme of distributing along the last nonsingleton dimension. Optional arguments to codistributed.cell must be specified after the required arguments, and in the following order: codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no communication is to be performed when constructing the array, skipping some error checking steps.
C = cell(n, codist) is the same as C = codistributed.cell(n, codist). You can also use the 'noCommunication' object with this
syntax. To use the default distribution scheme, specify a codistributor constructor without arguments. For example:
spmd C = cell(8, codistributor1d()); end
14-18
codistributed.cell
C = cell(m, n, p, ..., codist) and C = cell([m, n, p, ...], codist) are the same as C = codistributed.cell(m, n, p, ...) and C = codistributed.cell([m, n, p, ...]), respectively. You can also use the optional 'noCommunication' argument with this
syntax.
Examples
creates a 1000-by-1000 distributed cell array C, distributed by its second dimension (columns). Each lab contains a 1000-by-250 local piece of C.
spmd(4) codist = codistributor1d(2, 1:numlabs); C = cell(10, 10, codist); end
creates a 10-by-10 codistributed cell array C, distributed by its columns. Each lab contains a 10-by-labindex local piece of C.
See Also
cell | distributed.cell
14-19
codistributed.colon
contiguous subvectors of equal, or nearly equal length, and creates a codistributed array whose local portion on each lab is the labindex-th subvector.
codistributed.colon(a,b) uses d = 1.
Optional arguments to codistributed.colon must be specified after the required arguments, and in the following order: codist A codistributor object specifying the distribution scheme of the resulting vector. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no communication is to be performed when constructing the vector, skipping some error checking steps.
Examples
Partition the vector 1:10 into four subvectors among four labs.
spmd(4); C = codistributed.colon(1,10), end Lab 1: This lab stores C(1:3). LocalPart: [1 2 3] Codistributor: [1x1 codistributor1d] Lab 2: This lab stores C(4:6). LocalPart: [4 5 6] Codistributor: [1x1 codistributor1d] Lab 3: This lab stores C(7:8).
14-20
codistributed.colon
LocalPart: [7 8] Codistributor: [1x1 codistributor1d] Lab 4: This lab stores C(9:10). LocalPart: [9 10] Codistributor: [1x1 codistributor1d]
See Also
14-21
codistributed.eye
Purpose Syntax
Description
ones on the diagonal and zeros elsewhere. Optional arguments to codistributed.eye must be specified after the required arguments, and in the following order: classname Specifies the class of the codistributed array C. Valid choices are the same as for the regular eye function: 'double' (the default), 'single', 'int8', 'uint8', 'int16', 'uint16', 'int32', 'uint32', 'int64', and 'uint64'. codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no interworker communication is to be performed when constructing the array, skipping some error checking steps.
C = eye(n, codist) is the same as C = codistributed.eye(n, codist). You can also use the optional arguments with this syntax. To
use the default distribution scheme, specify a codistributor constructor without arguments. For example:
14-22
codistributed.eye
spmd C = eye(8, codistributor1d()); end C = eye(m, n, codist) and C = eye([m, n], codist) are the same as C = codistributed.eye(m, n) and C = codistributed.eye([m, n]), respectively. You can also use the optional arguments with this
syntax.
Examples
creates a 1000-by-1000 codistributed double array C, distributed by its second dimension (columns). Each lab contains a 1000-by-250 local piece of C.
spmd(4) codist = codistributor('1d', 2, 1:numlabs); C = eye(10, 10, 'uint16', codist); end
creates a 10-by-10 codistributed uint16 array D, distributed by its columns. Each lab contains a 10-by-labindex local piece of D.
See Also
14-23
codistributed.false
Purpose Syntax
Description
of logical zeros.
F = codistributed.false(m, n, ...) or F = codistributed.false([m, n, ...]) creates an m-by-n-by-...
codistributed array of logical zeros. Optional arguments to codistributed.false must be specified after the required arguments, and in the following order: codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no interworker communication is to be performed when constructing the array, skipping some error checking steps.
F = false(n, codist) is the same as F = codistributed.false(n, codist). You can also use the optional arguments with this syntax. To
use the default distribution scheme, specify a codistributor constructor without arguments. For example:
spmd F = false(8, codistributor1d()); end
14-24
codistributed.false
F = false(m, n, ..., codist) and F = false([m, n, ...], codist) are the same as F = codistributed.false(m, n, ...) and F = codistributed.false([m, n, ...]), respectively. You can also
Examples
creates a 1000-by-1000 codistributed false array F, distributed by its second dimension (columns). Each lab contains a 1000-by-250 local piece of F.
spmd codist = codistributor('1d', 2, 1:numlabs); F = false(10, 10, codist); end
creates a 10-by-10 codistributed false array F, distributed by its columns. Each lab contains a 10-by-labindex local piece of F.
See Also
14-25
codistributed.Inf
Purpose Syntax
Description
C = codistributed.Inf(n) creates an n-by-n codistributed matrix of Inf values. C = codistributed.Inf(m, n, ...) or C = codistributed.Inf([m, n, ...]) creates an m-by-n-by-... codistributed array of Inf values.
Optional arguments to codistributed.Inf must be specified after the required arguments, and in the following order: classname Specifies the class of the codistributed array C. Valid choices are the same as for the regular Inf function: 'double' (the default), or 'single'. codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no interworker communication is to be performed when constructing the array, skipping some error checking steps.
C = Inf(n, codist) is the same as C = codistributed.Inf(n, codist). You can also use the optional arguments with this syntax. To
use the default distribution scheme, specify a codistributor constructor without arguments. For example:
spmd C = Inf(8, codistributor1d());
14-26
codistributed.Inf
end C = Inf(m, n, ..., codist) and C = Inf([m, n, ...], codist) are the same as C = codistributed.Inf(m, n, ...) and C = codistributed.Inf([m, n, ...]), respectively. You can also use the
Examples
creates a 1000-by-1000 codistributed double matrix C, distributed by its second dimension (columns). Each lab contains a 1000-by-250 local piece of C.
spmd(4) codist = codistributor('1d', 2, 1:numlabs); C = Inf(10, 10, 'single', codist); end
creates a 10-by-10 codistributed single array C, distributed by its columns. Each lab contains a 10-by-labindex local piece of C.
See Also
14-27
codistributed.NaN
Purpose Syntax
Description
C = codistributed.NaN(n) creates an n-by-n codistributed matrix of NaN values. C = codistributed.NaN(m, n, ...) or C = codistributed.NaN([m, n, ...]) creates an m-by-n-by-... codistributed array of NaN values.
Optional arguments to codistributed.NaN must be specified after the required arguments, and in the following order: classname Specifies the class of the codistributed array C. Valid choices are the same as for the regular NaN function: 'double' (the default), or 'single'. codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no interworker communication is to be performed when constructing the array, skipping some error checking steps.
C = NaN(n, codist) is the same as C = codistributed.NaN(n, codist). You can also use the optional arguments with this syntax. To
use the default distribution scheme, specify a codistributor constructor without arguments. For example:
spmd C = NaN(8, codistributor1d());
14-28
codistributed.NaN
end C = NaN(m, n, ..., codist) and C = NaN([m, n, ...], codist) are the same as C = codistributed.NaN(m, n, ...) and C = codistributed.NaN([m, n, ...]), respectively. You can also use the
Examples
creates a 1000-by-1000 codistributed double matrix C of NaN values, distributed by its second dimension (columns). Each lab contains a 1000-by-250 local piece of C.
spmd(4) codist = codistributor('1d', 2, 1:numlabs); C = NaN(10, 10, 'single', codist); end
creates a 10-by-10 codistributed single array C, distributed by its columns. Each lab contains a 10-by-labindex local piece of C.
See Also
14-29
codistributed.ones
Purpose Syntax
Description
codistributed array of ones. Optional arguments to codistributed.ones must be specified after the required arguments, and in the following order: classname Specifies the class of the codistributed array C. Valid choices are the same as for the regular ones function: 'double' (the default), 'single', 'int8', 'uint8', 'int16', 'uint16', 'int32', 'uint32', 'int64', and 'uint64'. codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no interworker communication is to be performed when constructing the array, skipping some error checking steps.
C = ones(n, codist) is the same as C = codistributed.ones(n, codist). You can also use the optional arguments with this syntax. To
use the default distribution scheme, specify a codistributor constructor without arguments. For example:
14-30
codistributed.ones
spmd C = ones(8, codistributor1d()); end C = ones(m, n, codist) and C = ones([m, n], codist) are the same as C = codistributed.ones(m, n, ...) and C = codistributed.ones([m, n, ...]), respectively. You can also use
Examples
creates a 1000-by-1000 codistributed double array of ones, C, distributed by its second dimension (columns). Each lab contains a 1000-by-250 local piece of C.
spmd(4) codist = codistributor('1d', 2, 1:numlabs); C = ones(10, 10, 'uint16', codist); end
creates a 10-by-10 codistributed uint16 array of ones, C, distributed by its columns. Each lab contains a 10-by-labindex local piece of C.
See Also
14-31
codistributed.rand
Purpose Syntax
Description
codistributed array of underlying class double. Optional arguments to codistributed.rand must be specified after the required arguments, and in the following order: classname Specifies the class of the codistributed array C. Valid choices are the same as for the regular rand function: 'double' (the default), 'single', 'int8', 'uint8', 'int16', 'uint16', 'int32', 'uint32', 'int64', and 'uint64'. codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no interworker communication is to be performed when constructing the array, skipping some error checking steps.
R = rand(n, codist) is the same as R = codistributed.rand(n, codist). You can also use the optional arguments with this syntax. To
use the default distribution scheme, specify a codistributor constructor without arguments. For example:
14-32
codistributed.rand
spmd R = codistributed.rand(8, codistributor1d()); end R = rand(m, n, codist) and R = rand([m, n], codist) are the same as R = codistributed.rand(m, n, ...) and R = codistributed.rand([m, n, ...]), respectively. You can also use
Tips
When you use rand on the workers in the MATLAB pool, or in a distributed or parallel job (including pmode), each worker or lab sets its random generator seed to a value that depends only on the lab index or task ID. Therefore, the array on each lab is unique for that job. However, if you repeat the job, you get the same random data. With four labs,
spmd(4) R = codistributed.rand(1000, codistributor()) end
Examples
creates a 1000-by-1000 codistributed double array R, distributed by its second dimension (columns). Each lab contains a 1000-by-250 local piece of R.
spmd(4) codist = codistributor('1d', 2, 1:numlabs); R = codistributed.rand(10, 10, 'uint16', codist); end
creates a 10-by-10 codistributed uint16 array R, distributed by its columns. Each lab contains a 10-by-labindex local piece of R.
See Also
14-33
codistributed.randn
Purpose Syntax
Description
codistributed array of normally distributed random values. Optional arguments to codistributed.randn must be specified after the required arguments, and in the following order: classname Specifies the class of the codistributed array C. Valid choices are the same as for the regular rand function: 'double' (the default), 'single', 'int8', 'uint8', 'int16', 'uint16', 'int32', 'uint32', 'int64', and 'uint64'. codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no interworker communication is to be performed when constructing the array, skipping some error checking steps.
RN = randn(n, codist) is the same as RN = codistributed.randn(n, codist). You can also use the optional
arguments with this syntax. To use the default distribution scheme, specify a codistributor constructor without arguments. For example:
14-34
codistributed.randn
spmd RN = codistributed.randn(8, codistributor1d()); end RN = randn(m, n, codist) and RN = randn([m, n], codist) are the same as RN = codistributed.randn(m, n, ...) and RN = codistributed.randn([m, n, ...]), respectively. You can also use
Tips
When you use randn on the workers in the MATLAB pool, or in a distributed or parallel job (including pmode), each worker or lab sets its random generator seed to a value that depends only on the lab index or task ID. Therefore, the array on each lab is unique for that job. However, if you repeat the job, you get the same random data. With four labs,
spmd(4) RN = codistributed.randn(1000); end
Examples
creates a 1000-by-1000 codistributed double array RN, distributed by its second dimension (columns). Each lab contains a 1000-by-250 local piece of RN.
spmd(4) codist = codistributor('1d', 2, 1:numlabs); RN = randn(10, 10, 'uint16', codist); end
creates a 10-by-10 codistributed uint16 array RN, distributed by its columns. Each lab contains a 10-by-labindex local piece of RN.
See Also
14-35
codistributed.spalloc
Optional arguments to codistributed.spalloc must be specified after the required arguments, and in the following order: codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. The allocated space for nonzero elements is consistent with the distribution of the matrix among the labs according to the Partition of the codistributor. 'noCommunication' Specifies that no communication is to be performed when constructing the array, skipping some error checking steps. You can also use this argument with SD = spalloc(M, N, nzmax, codistr).
SD = spalloc(M, N, nzmax, codist) is the same as SD = codistributed.spalloc(M, N, nzmax, codist). You can also use
Examples
Allocate space for a 1000-by-1000 sparse codistributed matrix with room for up to 2000 nonzero elements. Use the default codistributor. Define several elements of the matrix.
spmd % codistributed array created inside spmd statement N = 1000; SD = codistributed.spalloc(N, N, 2*N); for ii=1:N-1 SD(ii,ii:ii+1) = [ii ii]; end
end
14-36
codistributed.spalloc
See Also
14-37
codistributed.speye
Purpose Syntax
Description
array of underlying class double. Optional arguments to codistributed.speye must be specified after the required arguments, and in the following order: codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no interworker communication is to be performed when constructing the array, skipping some error checking steps.
CS = speye(n, codist) is the same as CS = codistributed.speye(n, codist). You can also use the optional
arguments with this syntax. To use the default distribution scheme, specify a codistributor constructor without arguments. For example:
spmd CS = codistributed.speye(8, codistributor1d()); end
14-38
codistributed.speye
CS = speye(m, n, codist) and CS = speye([m, n], codist) are the same as CS = codistributed.speye(m, n) and CS = codistributed.speye([m, n]), respectively. You can also use the
optional arguments with this syntax. Note To create a sparse codistributed array of underlying class logical, first create an array of underlying class double and then cast it using the logical function:
CLS = logical(speye(m, n, codistributor1d()))
Examples
creates a 1000-by-1000 sparse codistributed double array CS, distributed by its second dimension (columns). Each lab contains a 1000-by-250 local piece of CS.
spmd(4) codist = codistributor1d(2, 1:numlabs); CS = speye(10, 10, codist); end
creates a 10-by-10 sparse codistributed double array CS, distributed by its columns. Each lab contains a 10-by-labindex local piece of CS.
See Also
14-39
codistributed.sprand
distributed nonzero double entries. Optional arguments to codistributed.sprand must be specified after the required arguments, and in the following order: codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no interworker communication is to be performed when constructing the array, skipping some error checking steps.
CS = sprand(n, codist) is the same as CS = codistributed.sprand(n, codist). You can also use the optional
arguments with this syntax. To use the default distribution scheme, specify a codistributor constructor without arguments. For example:
spmd CS = codistributed.sprand(8, 8, 0.2, codistributor1d()); end
Tips
When you use sprand on the workers in the MATLAB pool, or in a distributed or parallel job (including pmode), each worker or lab sets its random generator seed to a value that depends only on the lab index or task ID. Therefore, the array on each lab is unique for that job. However, if you repeat the job, you get the same random data.
14-40
codistributed.sprand
Examples
creates a 1000-by-1000 sparse codistributed double array CS with approximately 1000 nonzeros. CS is distributed by its second dimension (columns), and each lab contains a 1000-by-250 local piece of CS.
spmd(4) codist = codistributor1d(2, 1:numlabs); CS = sprand(10, 10, .1, codist); end
creates a 10-by-10 codistributed double array CS with approximately 10 nonzeros. CS is distributed by its columns, and each lab contains a 10-by-labindex local piece of CS.
See Also
14-41
codistributed.sprandn
distributed nonzero double entries. Optional arguments to codistributed.sprandn must be specified after the required arguments, and in the following order: codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no interworker communication is to be performed when constructing the array, skipping some error checking steps.
CS = sprandn(n, codist) is the same as CS = codistributed.sprandn(n, codist). You can also use the optional
arguments with this syntax. To use the default distribution scheme, specify a codistributor constructor without arguments. For example:
spmd CS = codistributed.sprandn(8, 8, 0.2, codistributor1d()); end
Tips
When you use sprandn on the workers in the MATLAB pool, or in a distributed or parallel job (including pmode), each worker or lab sets its random generator seed to a value that depends only on the lab index or task ID. Therefore, the array on each lab is unique for that job. However, if you repeat the job, you get the same random data.
14-42
codistributed.sprandn
Examples
creates a 1000-by-1000 sparse codistributed double array CS with approximately 1000 nonzeros. CS is distributed by its second dimension (columns), and each lab contains a 1000-by-250 local piece of CS.
spmd(4) codist = codistributor1d(2, 1:numlabs); CS = sprandn(10, 10, .1, codist); end
creates a 10-by-10 codistributed double array CS with approximately 10 nonzeros. CS is distributed by its columns, and each lab contains a 10-by-labindex local piece of CS.
See Also
14-43
codistributed.true
Purpose Syntax
Description
of logical ones.
T = codistributed.true(m, n, ...) or T = codistributed.true([m, n, ...]) creates an m-by-n-by-...
codistributed array of logical ones. Optional arguments to codistributed.true must be specified after the required arguments, and in the following order: codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no interworker communication is to be performed when constructing the array, skipping some error checking steps.
T = true(n, codist) is the same as T = codistributed.true(n, codist). You can also use the optional arguments with this syntax. To
use the default distribution scheme, specify a codistributor constructor without arguments. For example:
spmd T = true(8, codistributor1d()); end
14-44
codistributed.true
T = true(m, n, ..., codist) and T = true([m, n, ...], codist) are the same as T = codistributed.true(m, n, ...) and T = codistributed.true([m, n, ...]), respectively. You can also
Examples
creates a 1000-by-1000 codistributed true array T, distributed by its second dimension (columns). Each lab contains a 1000-by-250 local piece of T.
spmd(4) codist = codistributor('1d', 2, 1:numlabs); T = true(10, 10, codist); end
creates a 10-by-10 codistributed true array T, distributed by its columns. Each lab contains a 10-by-labindex local piece of T.
See Also
14-45
codistributed.zeros
Purpose Syntax
Description
codistributed array of zeros. Optional arguments to codistributed.zeros must be specified after the required arguments, and in the following order: classname Specifies the class of the codistributed array C. Valid choices are the same as for the regular zeros function: 'double' (the default), 'single', 'int8', 'uint8', 'int16', 'uint16', 'int32', 'uint32', 'int64', and 'uint64'. codist A codistributor object specifying the distribution scheme of the resulting array. If omitted, the array is distributed using the default distribution scheme. For information on constructing codistributor objects, see the reference pages for codistributor1d and codistributor2dbc. 'noCommunication' Specifies that no interworker communication is to be performed when constructing the array, skipping some error checking steps.
C = zeros(n, codist) is the same as C = codistributed.zeros(n, codist). You can also use the optional arguments with this syntax. To
use the default distribution scheme, specify a codistributor constructor without arguments. For example:
14-46
codistributed.zeros
spmd C = zeros(8, codistributor1d()); end C = zeros(m, n, codist) and C = zeros([m, n], codist) are the same as C = codistributed.zeros(m, n, ...) and C = codistributed.zeros([m, n, ...]), respectively. You can also use
Examples
creates a 1000-by-1000 codistributed double array of zeros, C, distributed by its second dimension (columns). Each lab contains a 1000-by-250 local piece of C.
spmd(4) codist = codistributor('1d', 2, 1:numlabs) C = zeros(10, 10, 'uint16', codist); end
creates a 10-by-10 codistributed uint16 array of zeros, C, distributed by its columns. Each lab contains a 10-by-labindex local piece of C.
See Also
14-47
codistributor
Purpose Syntax
Description
There are two schemes for distributing arrays. The scheme denoted by the string '1d' distributes an array along a single specified subscript, the distribution dimension, in a noncyclic, partitioned manner. The scheme denoted by '2dbc', employed by the parallel matrix computation software ScaLAPACK, applies only to two-dimensional arrays, and varies both subscripts over a rectangular computational grid of labs in a blocked, cyclic manner.
codist = codistributor(), with no arguments, returns a default
codistributor object with zero-valued or empty parameters, which can then be used as an argument to other functions to indicate that the function is to create a codistributed array if possible with default distribution. For example,
Z = zeros(..., codistributor()) R = randn(..., codistributor()) codist = codistributor('1d') is the same as codist = codistributor(). codist = codistributor('1d', dim) also forms a codistributor object with codist.Dimension = dim and default partition. codist = codistributor('1d', dim, part) also forms a codistributor object with codist.Dimension = dim and codist.Partition = part.
14-48
codistributor
codist = codistributor('2dbc') forms a 2-D block-cyclic codistributor object. For more information about '2dbc' distribution, see 2-Dimensional Distribution on page 5-17. codist = codistributor('2dbc', lbgrid) forms a 2-D block-cyclic codistributor object with the lab grid defined by lbgrid and with default block size. codist = codistributor('2dbc', lbgrid, blksize) forms a 2-D block-cyclic codistributor object with the lab grid defined by lbgrid and with a block size defined by blksize. codist = getCodistributor(D) returns the codistributor object of
codistributed array D.
Examples
On four labs, create a 3-dimensional, 2-by-6-by-4 array with distribution along the second dimension, and partition scheme [1 2 1 2]. In other words, lab 1 contains a 2-by-1-by-4 segment, lab 2 a 2-by-2-by-4 segment, etc.
spmd dim = 2; % distribution dimension codist = codistributor('1d', dim, [1 2 1 2], [2 6 4]); if mod(labindex, 2) L = rand(2,1,4); else L = rand(2,2,4); end A = codistributed.build(L, codist) end A
On four labs, create a 20-by-5 codistributed array A, distributed by rows (over its first dimension) with a uniform partition scheme.
spmd dim = 1; % distribution dimension partn = codistributor1d.defaultPartition(20); codist = codistributor('1d', dim, partn, [20 5]);
14-49
codistributor
See Also
14-50
codistributor1d
Purpose Syntax
Description
The 1-D codistributor distributes arrays along a single, specified distribution dimension, in a noncyclic, partitioned manner.
codist = codistributor1d() forms a 1-D codistributor object using default dimension and partition. The default dimension is the last nonsingleton dimension of the codistributed array. The default partition distributes the array along the default dimension as evenly as possible. codist = codistributor1d(dim) forms a 1-D codistributor object for distribution along the specified dimension: 1 distributes along rows, 2 along columns, etc. codist = codistributor1d(dim, part) forms a 1-D codistributor object for distribution according to the partition vector part. For example C1 = codistributor1d(1, [1, 2, 3, 4]) describes the
distribution scheme for an array of ten rows to be codistributed by its first dimension (rows), to four labs, with 1 row to the first, 2 rows to the second, etc. The resulting codistributor of any of the above syntax is incomplete because its global size is not specified. A codistributor constructed in this manner can be used as an argument to other functions as a template codistributor when creating codistributed arrays.
codist = codistributor1d(dim, part, gsize) forms a codistributor object with distribution dimension dim, distribution partition part, and global size of its codistributed arrays gsize. The resulting codistributor
object is complete and can be used to build a codistributed array from its local parts with codistributed.build. To use a default dimension, specify codistributor1d.unsetDimension for that argument; the distribution dimension is derived from gsize and is set to the last non-singleton dimension. Similarly, to use a default partition, specify
14-51
codistributor1d
is then derived from the default for that global size and distribution dimension. The local part on lab labidx of a codistributed array using such a codistributor is of size gsize in all dimensions except dim, where the size is part(labidx). The local part has the same class and attributes as the overall codistributed array. Conceptually, the overall global array could be reconstructed by concatenating the various local parts along dimension dim.
Examples
Use a fully specified codistributor1d object to create a trivial N-by-N codistributed matrix from its local parts. Then visualize which elements are stored on lab 2.
N = 1000; spmd codistr = codistributor1d( ... codistributor1d.unsetDimension, ... codistributor1d.unsetPartition, ... [N, N]); myLocalSize = [N, N]; % start with full size on each lab % then set myLocalSize to default part of whole array: myLocalSize(codistr.Dimension) = codistr.Partition(labindex); myLocalPart = labindex*ones(myLocalSize); % arbitrary values D = codistributed.build(myLocalPart, codistr); end spy(D == 2);
14-52
codistributor1d
See Also
14-53
codistributor1d.defaultPartition
Examples
See Also
14-54
codistributor2dbc
Purpose Syntax
Description
The 2-D block-cyclic codistributor can be used only for two-dimensional arrays. It distributes arrays along two subscripts over a rectangular computational grid of labs in a block-cyclic manner. For a complete description of 2-D block-cyclic distribution, default parameters, and the relationship between block size and lab grid, see 2-Dimensional Distribution on page 5-17. The 2-D block-cyclic codistributor is used by the ScaLAPACK parallel matrix computation software library.
codist = codistributor2dbc() forms a 2-D block-cyclic codistributor object using default lab grid and block size. codist = codistributor2dbc(lbgrid) forms a 2-D block-cyclic codistributor object using the specified lab grid and default block size. lbgrid must be a two-element vector defining the rows and columns of the lab grid, and the rows times columns must equal the number of labs for the codistributed array. codist = codistributor2dbc(lbgrid, blksize) forms a 2-D block-cyclic codistributor object using the specified lab grid and block size. codist = codistributor2dbc(lbgrid, blksize, orient) allows an orientation argument. Valid values for the orientation argument are 'row' for row orientation, and 'col' for column orientation of the lab grid. The default is row orientation.
The resulting codistributor of any of the above syntax is incomplete because its global size is not specified. A codistributor constructed this way can be used as an argument to other functions as a template codistributor when creating codistributed arrays.
14-55
codistributor2dbc
codist = codistributor2dbc(lbgrid, blksize, orient, gsize) forms a codistributor object that distributes arrays with the global size gsize. The resulting codistributor object
is complete and can therefore be used to build a codistributed array from its local parts with codistributed.build. To use the default values for lab grid, block size, and orientation, specify them using codistributor2dbc.defaultLabGrid, codistributor2dbc.defaultBlockSize, and codistributor2dbc.defaultOrientation, respectively.
Examples
Use a fully specified codistributor2dbc object to create a trivial N-by-N codistributed matrix from its local parts. Then visualize which elements are stored on lab 2.
N = 1000; spmd codistr = codistributor2dbc(... codistributor2dbc.defaultLabGrid, ... codistributor2dbc.defaultBlockSize, ... 'row', [N, N]); myLocalSize = [length(codistr.globalIndices(1)), ... length(codistr.globalIndices(2))]; myLocalPart = labindex*ones(myLocalSize); D = codistributed.build(myLocalPart, codistr); end spy(D == 2);
See Also
14-56
codistributor2dbc.defaultLabGrid
The grid defined by codistributor2dbc.defaultLabGrid is as close to a square as possible. The following rules define nrow and ncol: If numlabs is a perfect square, nrow = ncol = sqrt(numlabs). If numlabs is an odd power of 2, then nrow = ncol/2 = sqrt(numlabs/2). nrow <= ncol. If numlabs is a prime, nrow = 1, ncol = numlabs. nrow is the greatest integer less than or equal to sqrt(numlabs) for which ncol = numlabs/nrow is also an integer.
Examples
View the computational grid layout of the default distribution scheme for the open MATLAB pool.
spmd grid = codistributor2dbc.defaultLabGrid end
See Also
14-57
Composite
from the MATLAB pool. The actual number of labs referenced by this Composite object depends on the size of the MATLAB pool and any existing Composite objects. Generally, you should construct Composite objects outside any spmd statement.
C = Composite(nlabs) creates a Composite object on the parallel resource set that matches the specified constraint. nlabs must be a vector of length 1 or 2, containing integers or Inf. If nlabs is of length 1, it specifies the exact number of labs to use. If nlabs is of size 2, it specifies the minimum and maximum number of labs to use. The actual number of labs used is the maximum number of labs compatible with the size of the MATLAB pool, and with other existing Composite objects. An error is thrown if the constraints on the number of labs cannot be met.
A Composite object has one entry for each lab; initially each entry contains no data. Use either indexing or an spmd block to define values for the entries.
Examples
Create a Composite object with no defined entries, then assign its values:
c = Composite(); % One element per lab in the pool for ii = 1:length(c) % Set the entry for each lab to zero c{ii} = 0; % Value stored on each lab end
See Also
matlabpool | spmd
14-58
createJob
Purpose Syntax
Arguments
The job object. The scheduler object created by findResource. Object properties configured at object creation. Initial values for corresponding object properties.
Description
obj = createJob() creates a job using the scheduler identified by the default parallel configuration and sets the property values of the job as specified in the default configuration. obj = createJob(scheduler) creates a job object at the data location
for the identified scheduler, or in the job manager. When you specify a
scheduler without using the 'configuration' option, no configuration
object with the specified property values. For a listing of the valid properties of the created object, see the job object reference page (if using a job manager) or simplejob object reference page (if using a third-party scheduler). If an invalid property name or property value is specified, the object will not be created. Note that the property value pairs can be in any format supported by the set function, i.e., param-value string pairs, structures, and param-value cell array pairs. If a structure is used, the structure field names are job object property names and the field values specify the property values.
14-59
createJob
If you are using a third-party scheduler instead of a job manager, the jobs data is stored in the location specified by the schedulers DataLocation property.
obj = createJob(..., 'configuration', 'ConfigurationName', ...) creates a job object using the scheduler identified by the
configuration and sets the property values of the job as specified in that configuration. For details about defining and applying configurations, see Parallel Configurations for Cluster Access on page 6-14.
Examples
Wait for the job to finish running, and retrieve the job results.
waitForState(obj); out = getAllOutputArguments(obj);
See Also
14-60
createMatlabPoolJob
Purpose Syntax
Arguments
The job object. Object properties configured at object creation. Initial values for corresponding object properties.
Description
MATLAB pool job with the specified property values. For a listing of the valid properties of the created object, see the matlabpooljob object reference page (if using a job manager) or simplematlabpooljob object reference page (if using a third-party scheduler). If an invalid property name or property value is specified, the object is not created. These values override any values in the default configuration.
job = createMatlabPoolJob(..., 'configuration', 'ConfigurationName',...) creates a MATLAB pool job using the
scheduler identified by the configuration and sets the property values of the job as specified in that configuration. For details about defining and applying configurations, see Parallel Configurations for Cluster Access on page 6-14.
Examples
14-61
createMatlabPoolJob
See Also
14-62
createParallelJob
Purpose Syntax
Arguments
The parallel job object. The scheduler object created by findResource. Object properties configured at object creation. Initial values for corresponding object properties.
Description
scheduler identified by the default parallel configuration and sets the property values of the job as specified in the default configuration.
pjob = createParallelJob(scheduler) creates a parallel job object at the data location for the identified scheduler, or in the job manager. When you specify a scheduler without using the 'configuration' option, no configuration is used, so no configuration properties are applied to the job object. pjob = createParallelJob(..., 'p1', v1, 'p2', v2, ...)
creates a parallel job object with the specified property values. For a listing of the valid properties of the created object, see the paralleljob object reference page (if using a job manager) or simpleparalleljob object reference page (if using a third-party scheduler). If an invalid property name or property value is specified, the object will not be created. Property value pairs can be in any format supported by the set function, i.e., param-value string pairs, structures, and param-value cell array pairs. Future modifications to the job object result in a remote call to the job manager or modification to data at the schedulers data location.
14-63
createParallelJob
using the scheduler identified by the configuration and sets the property values of the job as specified in that configuration. For details about defining and applying configurations, see Parallel Configurations for Cluster Access on page 6-14.
Examples
Wait for the job to finish running, and retrieve the job results.
waitForState(pjob); out = getAllOutputArguments(pjob);
14-64
createParallelJob
See Also
14-65
createTask
Purpose Syntax
Arguments
t j F
Task object or vector of task objects. The job that the task object is created in. A handle to the function that is called when the task is evaluated, or an array of function handles. The number of output arguments to be returned from execution of the task function. This is a double or array of doubles. A row cell array specifying the input arguments to be passed to the function F. Each element in the cell array will be passed as a separate input argument. If this is a cell array of cell arrays, a task is created for each cell array. Cell array of cell arrays defining input arguments to each of m tasks. Task object properties configured at object creation. Initial values for corresponding task object properties.
{inputargs}
Description
t = createTask(j, F, N, {inputargs}) creates a new task object in job j, and returns a reference, t, to the added task object. This task evaluates the function specified by a function handle or function
14-66
createTask
name F, with the given input arguments {inputargs}, returning N output arguments.
t = createTask(j, F, N, {C1,...,Cm}) uses a cell array of m cell arrays to create m task objects in job j, and returns a vector, t, of references to the new task objects. Each task evaluates the function specified by a function handle or function name F. The cell array C1 provides the input arguments to the first task, C2 to the second task, and so on, so that there is one task per cell array. Each task returns N output arguments. If F is a cell array, each element of F specifies a function for each task in the vector; it must have m elements. If N is an array of doubles, each element specifies the number of output arguments for each task in the vector. Multidimensional matrices of inputs F, N and {C1,...,Cm} are supported; if a cell array is used for F, or a double array for N, its dimensions must match those of the input arguments cell array of cell arrays. The output t will be a vector with the same number of elements as {C1,...,Cm}. Note that because a parallel job has only one task, this form of vectorized task creation is not appropriate for parallel jobs. t = createTask(..., 'p1',v1,'p2',v2,...) adds a task object with the specified property values. For a listing of the valid properties of the created object, see the task object reference page (if using a job manager) or simpletask object reference page (if using a third-party scheduler). If an invalid property name or property value is specified, the object will not be created.
Note that the property value pairs can be in any format supported by the set function, i.e., param-value string pairs, structures, and param-value cell array pairs. If a structure is used, the structure field names are task object property names and the field values specify the property values.
t = createTask(...,'configuration', 'ConfigurationName',...)
creates a task job object with the property values specified in the configuration ConfigurationName. For details about defining and applying configurations, see Parallel Configurations for Cluster Access on page 6-14.
14-67
createTask
Examples
Wait for the job to finish running, and get the output from the task evaluation.
waitForState(j); taskoutput = get(obj, 'OutputArguments');
Create a job with three tasks, each of which generates a 10-by-10 random matrix.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); t = createTask(j, @rand, 1, {{10,10} {10,10} {10,10}});
See Also
14-68
defaultParallelConfig
Purpose Syntax
Arguments
String indicating name of current default configuration Cell array of strings indicating names of all available configurations String indicating name of previous default configuration String specifying name of new default configuration
Description
The defaultParallelConfig function allows you to programmatically get or set the default parallel configuration and obtain a list of all valid configurations.
[config, allconfigs] = defaultParallelConfig returns the name of the default parallel computing configuration, as well as a cell array containing the names of all available configurations. [oldconfig, allconfigs] = defaultParallelConfig(newconfig) sets the default parallel computing configuration to newconfig and
returns the previous default configuration and a cell array containing the names of all available configurations. The previous configuration is provided so that you can reset the default configuration to its original setting at the end of your session. The settings specified for defaultParallelConfig are saved as a part of your MATLAB preferences. The cell array allconfigs always contains a configuration called 'local' for the local scheduler. The default configuration returned by defaultParallelConfig is guaranteed to be found in allconfigs.
14-69
defaultParallelConfig
If the default configuration has been deleted, or if it has never been set, defaultParallelConfig returns 'local' as the default configuration.
Examples
Read the name of the default parallel configuration that is currently in effect, and get a listing of all available configurations.
[ConfigNow ConfigList] = defaultParallelConfig
See Also
14-70
demote
jm job
The job manager object that contains the job. Job object demoted in the job queue.
Description
demote(jm, job) demotes the job object job that is queued in the job manager jm.
If job is not the last job in the queue, demote exchanges the position of job and the job that follows it in the queue.
Tips
After a call to demote or promote, there is no change in the order of job objects contained in the Jobs property of the job manager object. To see the scheduled order of execution for jobs in the queue, use the findJob function in the form [pending queued running finished] = findJob(jm). Create and submit multiple jobs to the job manager identified by the default parallel configuration:
jm = findResource(); j1 = createJob('name','Job A'); j2 = createJob('name','Job B'); j3 = createJob('name','Job C'); submit(j1);submit(j2);submit(j3);
Examples
14-71
demote
See Also
14-72
destroy
obj
destroy(obj) removes the job object reference or task object reference obj from the local session, and removes the object from the job manager memory. When obj is destroyed, it becomes an invalid object. You can remove an invalid object from the workspace with the clear command.
If multiple references to an object exist in the workspace, destroying one reference to that object invalidates all the remaining references to it. You should remove these remaining references from the workspace with the clear command. The task objects contained in a job will also be destroyed when a job object is destroyed. This means that any references to those task objects will also be invalid.
Tips Examples
Because its data is lost when you destroy an object, destroy should be used after output data has been retrieved from a job object. Destroy a job and its tasks.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm, 'Name', 'myjob'); t = createTask(j, @rand, 1, {10}); destroy(j); clear t clear j
See Also
createJob | createTask
14-73
dfeval
Purpose Syntax
Arguments
Function name, function handle, or cell array of function names or handles. Cell arrays of input arguments to the functions. Cell arrays of output arguments; each element of a cell array corresponds to each task of the job. Property name/property value pairs for the created job object; can be name/value pairs or structures.
Description
[y1,...,ym] = dfeval(F, x1,...,xn) performs the equivalent of an feval in a cluster of machines using Parallel Computing Toolbox software. dfeval evaluates the function F, with arguments provided in the cell arrays x1,...,xn. F can be a function handle, a function name, or a cell array of function handles/function names where the length of the cell array is equal to the number of tasks to be executed. x1,...,xn are the inputs to the function F, specified as cell arrays in which the number of elements in the cell array equals the number of tasks to be executed. The first task evaluates function F using the first element of each cell array as input arguments; the second task uses the second element of each cell array, and so on. The sizes of x1,...,xn must all be the same.
The results are returned to y1,...,ym, which are column-based cell arrays, each of whose elements corresponds to each task that was created. The number of cell arrays (m) is equal to the number of output arguments returned from each task. For example, if the job has 10
14-74
dfeval
tasks that each generate three output arguments, the results of dfeval are three cell arrays of 10 elements each. When evaluation is complete, dfeval destroys the job.
y = dfeval( ..., 'P1',V1,'P2',V2,...) accepts additional
arguments for configuring different properties associated with the job. Valid properties and property values are Job object property value pairs, specified as name/value pairs or structures. (Properties of other object types, such as scheduler, task, or worker objects are not permitted. Use a configuration to set scheduler and task properties.) 'JobManager','JobManagerName'. This specifies the job manager on which to run the job. If you do not use this property to specify a job manager, the default is to run the job on the first job manager returned by findResource. 'LookupURL','host:port'. This makes a unicast call to the job manager lookup service at the specified host and port. The job managers available for this job are those accessible from this lookup service. For more detail, see the description of this option on the findResource reference page. 'StopOnError',true|{false}. You may also set the value to logical 1 (true) or 0 (false). If true (1), any error that occurs during execution in the cluster will cause the job to stop executing. The default value is 0 (false), which means that any errors that occur will produce a warning but will not stop function execution.
[y1,...,ym] = dfeval(F, x1,...,xn, ... 'configuration', 'ConfigurationName',...) evaluates the function F in a
initialize a scheduler, create a job, and create tasks. For details about defining and applying configurations, see Parallel Configurations for Cluster Access on page 6-14. Note that configurations enable you to use dfeval with any type of scheduler.
14-75
dfeval
Note that dfeval runs synchronously (sync); that is, it does not return the MATLAB prompt until the job is completed. For further discussion of the usage of dfeval, see Evaluate Functions Synchronously on page 7-2.
Examples
Create three tasks that return a 1-by-1, a 2-by-2, and a 3-by-3 random matrix.
y = dfeval(@rand,{1 2 3}) y = [ 0.9501] [2x2 double] [3x3 double]
Create two tasks that return random matrices of size 2-by-3 and 1-by-4.
y = dfeval(@rand,{2 1},{3 4}); y{1} ans = 0.8132 0.1389 0.1987 0.0099 0.2028 0.6038 y{2} ans = 0.6154 0.9218 0.1763
0.9355
Create two tasks, where the first task creates a 1-by-2 random array and the second task creates a 3-by-4 array of zeros.
y = dfeval({@rand @zeros},{1 3},{2 4}); y{1} ans = 0.0579 0.3529 y{2} ans = 0 0 0 0 0 0 0 0 0 0 0 0
14-76
dfeval
Create five random 2-by-4 matrices using MyJobManager to execute tasks, where the tasks time out after 10 seconds, and the function will stop if an error occurs while any of the tasks are executing.
y = dfeval(@rand,{2 2 2 2 2},{4 4 4 4 4}, ... 'JobManager','MyJobManager','Timeout',10,'StopOnError',true);
Evaluate the user function myFun using the cluster as defined in the configuration myConfig.
y = dfeval(@myFun, {task1args, task2args, task3args}, ... 'configuration', 'myConfig', ... 'FileDependencies', {'myFun.m'});
See Also
14-77
dfevalasync
Purpose Syntax
Arguments
Job object created to evaluation the function. Function name, function handle, or cell array of function names or handles. Number of output arguments from each tasks execution of function F. Cell arrays of input arguments to the functions. Property name/property value pairs for the created job object; can be name/value pairs or structures.
Description
it runs asynchronously (async), returning to the prompt immediately with a single output argument containing the job object that it has created and sent to the cluster. You have immediate access to the job object before the job is completed. You can use waitForState to determine when the job is completed, and getAllOutputArguments to retrieve your results.
Job = dfevalasync(F, numArgOut, x1,...,xn, ... 'configuration', 'ConfigurationName',...) evaluates the function F in a cluster by using all the properties defined in the configuration ConfigurationName. The configuration settings are used to find and
initialize a scheduler, create a job, and create tasks. For details about defining and applying configurations, see Parallel Configurations
14-78
dfevalasync
for Cluster Access on page 6-14. Configurations enable you to use dfevalasync with any type of scheduler. For further discussion on the usage of dfevalasync, see Evaluate Functions Asynchronously on page 7-8.
Examples
When the job is finished, you can obtain the results associated with the job.
waitForState(job); data = getAllOutputArguments(job) data = [ 3] [ 7] [11] data is an M-by-numArgOut cell array, where M is the number of tasks.
See Also
14-79
diary
Purpose Syntax
Arguments
job 'filename'
Job from which to view Command Window output text. File to append with Command Window output text from batch job
Description
diary(job) displays the Command Window output from the batch job
in the MATLAB Command Window. The Command Window output will be captured only if the batch command included the 'CaptureDiary' argument with a value of true.
diary(job, 'filename') causes the Command Window output from
See Also
14-80
distributed
parts on the workers of the open MATLAB pool. Constructing a distributed array from local data this way is appropriate only if the MATLAB client can store the entirety of X in its memory. To construct large distributed arrays, use one of the static constructor methods such as distributed.ones, distributed.zeros, etc. If the input argument is already a distributed array, the result is the same as the input.
Examples
14-81
distributed.cell
Purpose Syntax
Description
Examples
See Also
cell | codistributed.cell
14-82
distributed.eye
Purpose Syntax
Description
Examples
See Also
14-83
distributed.false
Purpose Syntax
Description
logical zeros.
F = distributed.false(m, n, ...) or F = distributed.false([m, n, ...]) creates an m-by-n-by-... distributed array of logical zeros.
Examples
See Also
14-84
distributed.Inf
Purpose Syntax
Description
values.
D = distributed.Inf(m, n, ...) or D = distributed.Inf([m, n, ...]) creates an m-by-n-by-... distributed array of Inf values. D = distributed.Inf(..., classname) specifies the class of the distributed array D. Valid choices are the same as for the regular Inf function: 'double' (the default), or 'single'.
Examples
See Also
14-85
distributed.NaN
Purpose Syntax
Description
values.
D = distributed.NaN(m, n, ...) or D = distributed.NaN([m, n, ...]) creates an m-by-n-by-... distributed array of NaN values. D = distributed.NaN(..., classname) specifies the class of the distributed array D. Valid choices are the same as for the regular NaN function: 'double' (the default), or 'single'.
Examples
See Also
14-86
distributed.ones
Purpose Syntax
Description
of class double.
D = distributed.ones(m, n, ...) or D = distributed.ones([m, n, ...]) creates an m-by-n-by-... distributed array of ones. D = distributed.ones(..., classname) specifies the class of the distributed array D. Valid choices are the same as for the regular ones function: 'double' (the default), 'single', 'int8', 'uint8', 'int16', 'uint16', 'int32', 'uint32', 'int64', and 'uint64'.
Examples
See Also
14-87
distributed.rand
Purpose Syntax
Description
double.
R = distributed.rand(..., classname) specifies the class of the distributed array R. Valid choices are the same as for the regular rand function: 'double' (the default), 'single', 'int8', 'uint8', 'int16', 'uint16', 'int32', 'uint32', 'int64', and 'uint64'.
Tips
When you use rand on the workers in the MATLAB pool, or in a distributed or parallel job (including pmode), each worker or lab sets its random generator seed to a value that depends only on the lab index or task ID. Therefore, the array on each lab is unique for that job. However, if you repeat the job, you get the same random data. Create a 1000-by-1000 distributed matrix of random values of class double:
R = distributed.rand(1000);
Examples
See Also
14-88
distributed.randn
Purpose Syntax
Description
Tips
When you use randn on the workers in the MATLAB pool, or in a distributed or parallel job (including pmode), each worker or lab sets its random generator seed to a value that depends only on the lab index or task ID. Therefore, the array on each lab is unique for that job. However, if you repeat the job, you get the same random data. Create a 1000-by-1000 distributed matrix of normally distributed random values of class double:
RN = distributed.randn(1000);
Examples
See Also
14-89
distributed.spalloc
Allocate space for a 1000-by-1000 sparse distributed matrix with room for up to 2000 nonzero elements, then define several elements:
N = 1000; SD = distributed.spalloc(N, N, 2*N); for ii=1:N-1 SD(ii,ii:ii+1) = [ii ii]; end
See Also
14-90
distributed.speye
Purpose Syntax
Description
double.
Examples
See Also
14-91
distributed.sprand
Tips
When you use sprand on the workers in the MATLAB pool, or in a distributed or parallel job (including pmode), each worker or lab sets its random generator seed to a value that depends only on the lab index or task ID. Therefore, the array on each lab is unique for that job. However, if you repeat the job, you get the same random data. Create a 1000-by-1000 sparse distributed double array DS with approximately 1000 nonzeros.
DS = distributed.sprand(1000, 1000, .001);
Examples
See Also
14-92
distributed.sprandn
Tips
When you use sprandn on the workers in the MATLAB pool, or in a distributed or parallel job (including pmode), each worker or lab sets its random generator seed to a value that depends only on the lab index or task ID. Therefore, the array on each lab is unique for that job. However, if you repeat the job, you get the same random data. Create a 1000-by-1000 sparse distributed double array DS with approximately 1000 nonzeros.
DS = distributed.sprandn(1000, 1000, .001);
Examples
See Also
14-93
distributed.true
Purpose Syntax
Description
logical ones.
T = distributed.true(m, n, ...) or T = distributed.true([m, n, ...]) creates an m-by-n-by-... distributed array of logical ones.
Examples
See Also
14-94
distributed.zeros
Purpose Syntax
Description
Examples
See Also
14-95
dload
Purpose Syntax
Description
dload without any arguments retrieves all variables from the binary file named matlab.mat. If matlab.mat is not available, the command
generates an error.
dload filename retrieves all variables from a file given a full pathname or a relative partial pathname. If filename has no extension, dload looks for filename.mat. dload loads the contents of distributed arrays and Composite objects onto MATLAB pool workers, other data types are loaded directly into the workspace of the MATLAB client. dload filename X loads only variable X from the file. dload filename X Y Z ... loads only the specified variables. dload does not support wildcards, nor the -regexp option. If any requested variable is not
returns the specified variables as separate output arguments (rather than a structure, which the load function returns). If any requested variable is not present in the file, an error occurs. When loading distributed arrays, the data is distributed over the available MATLAB pool workers using the default distribution scheme. It is not necessary to have the same size MATLAB pool open when loading as when saving using dsave. When loading Composite objects, the data is sent to the available MATLAB pool workers. If the Composite is too large to fit on the current
14-96
dload
MATLAB pool, the data is not loaded. If the Composite is smaller than the current MATLAB pool, a warning is issued.
Examples
Use the function form of dload to load distributed arrays P and Q from file fname.mat:
[P, Q] = dload('fname.mat', 'P', 'Q');
See Also
14-97
dsave
Purpose Syntax
Description
and writes to the file all workspace variables, including distributed arrays and Composite objects. You can retrieve the variable data using dload.
dsave filename saves all workspace variables to the binary file named filename.mat. If you do not specify an extension for filename, it assumes the extension .mat. dsave filename X saves only variable X to the file. dsave filename X Y Z saves X, Y, and Z. dsave does not support wildcards, nor the -regexp option. dsave does not support saving sparse distributed arrays.
Examples
With a MATLAB pool open, create and save several variables to mydatafile.mat:
D = distributed.rand(1000); C = Composite(); C{1} = magic(20); X = rand(40); dsave mydatafile D C X % % % % % Distributed array Data on lab 1 only Client workspace only Save all three variables
See Also
14-98
exist
If exist(C, labidx) returns true, C(labidx) does not throw an error, provided that the values of C on those labs are serializable. The function throws an error if the lab indices are invalid.
Examples
Define a variable on a random number of labs. Check on which labs the Composite entries are defined, and get all those values:
spmd if rand() > 0.5 c = labindex; end end ind = exist(c); cvals = c(ind);
See Also
Composite
14-99
feval
output arguments from the evaluation of the kernel. Each output argument corresponds to the value of the non-const pointer inputs to the CUDA kernel after it has executed. The output from feval running a kernel on the GPU is always GPUArray type, even if all the inputs are data from the MATLAB workspace. The number of output arguments, m, must not exceed the value of the MaxNumLHSArguments property of KERN.
Examples
You can use feval on this codes kernel (KERN) with the syntax:
[y1, y2] = feval(KERN, x1, x2, x3)
The three input arguments, x1, x2, and x3, correspond to the three arguments that are passed into the CUDA function. The output
14-100
feval
arguments, y1 and y2, are GPUArray types, and correspond to the values of pInOut1 and pInOut2 after the CUDA kernel has executed.
See Also
14-101
findJob
Purpose Syntax
Arguments
Scheduler object in which to find the job. Array of jobs whose State is pending in scheduler sched. Array of jobs whose State is queued in scheduler sched. Array of jobs whose State is running in scheduler sched. Array of jobs that have completed running, i.e., whose State is finished or failed in scheduler sched. Array of jobs found in scheduler sched. Job object properties to match. Values for corresponding object properties.
Description
out = findJob(sched) returns an array, out, of all job objects stored in the scheduler sched. Jobs in the array are ordered by the ID property of the jobs, indicating the sequence in which they were created. [pending queued running completed] = findJob(sched) returns arrays of all job objects stored in the scheduler sched, by state. Within pending, running, and completed, the jobs are returned in sequence of creation. Jobs in the array queued are in the order in which they are queued, with the job at queued(1) being the next to execute. The completed jobs include those that failed. Jobs that are destroyed or
14-102
findJob
of job objects whose property names and property values match those passed as parameter-value pairs, p1, v1, p2, v2. Note that the property value pairs can be in any format supported by the set function, i.e., param-value string pairs, structures, and param-value cell array pairs. If a structure is used, the structure field names are job object property names and the field values are the appropriate property values to match. When a property value is specified, it must use the same exact value that the get function returns, including letter case. For example, if get returns the Name property value as MyJob, then findJob will not find that object while searching for a Name property value of myjob.
See Also
14-103
findResource
Purpose Syntax
Arguments
out 'configuration'
Object or array of objects returned. Literal string to indicate usage of a configuration. Literal string specifying that you are finding a scheduler, which can be a job manager or a third-party scheduler. Specifies the type of scheduler: 'jobmanager', 'local', 'hpcserver', 'LSF', 'pbspro', 'torque', 'mpiexec', or any string that starts with 'generic'. Literal string specifying that you are finding a worker. Literal string to indicate usage of a remote lookup service. Host name and (optionally) port of remote lookup service to use.
'SchedType'
14-104
findResource
argument identifies the head node for the HPC Server scheduler. Ignored if scheduler type is not 'hpcserver'.
'headNode' p1, p2 v1, v2
String specifying the HPC Server head node name. Object properties to match. Values for corresponding object properties.
Description
out = findResource() returns a scheduler object , out, representing the scheduler identified by the default parallel configuration, with the scheduler object properties set to the values defined in that configuration. out = findResource('scheduler', ... 'configuration', 'ConfigurationName') returns a scheduler object , out,
representing the scheduler identified by the parallel configuration ConfigurationName, with the scheduler object properties set to the values defined in that configuration. For details about defining and applying parallel configurations, see Parallel Configurations for Cluster Access on page 6-14. Note If you specify the 'scheduler' option without the 'configuration' option, no configuration is used, so no configuration properties are applied to the object.
out = findResource('scheduler', 'type', 'SchedType') returns an array, out, containing objects representing all available parallel computing schedulers of the given type. SchedType can be 'jobmanager', 'local', 'hpcserver', 'LSF', 'pbspro', 'torque', 'mpiexec', or any string starting with 'generic'. A 'local' scheduler
queues jobs for running on workers that it will start on your local client machine. You can use different scheduler types starting with 'generic'
14-105
findResource
to identify one generic scheduler or configuration from another. You can have multiple scheduler objects to simultaneously support several job managers or generic schedulers, but you cannot create more than one object for each type of fully supported third-party scheduler or the local scheduler. For third-party and generic schedulers, job data is stored in the location specified by the scheduler objects DataLocation property.
out = findResource('scheduler', 'type', 'jobmanager', 'LookupURL', 'host:port') uses the lookup process of the job
manager running at a specific location. The lookup process is part of a job manager. By default, findResource uses all the lookup processes that are available to the local machine via multicast. If you specify 'LookupURL' with a host, findResource uses the job manager lookup process running at that location. The port is optional, and is necessary only if the lookup process was configured to use a port other than the default BASEPORT setting of the mdce_def file. This URL is where the lookup is performed from, it is not necessarily the host running the job manager or worker. This unicast call is useful when you want to find resources that might not be available via multicast or in a network that does not support multicast. Notes Although Version 5 of the Parallel Computing Toolbox and MATLAB Distributed Computing Server products continue to support multicast communications between their processes, multicast is not recommended and might not be supported in future releases.
findResource ignores LookupURL when finding third-party schedulers.
out = findResource('scheduler', 'type', 'hpcserver') searches your environment variables and Active Directory to find the HPC Server cluster head node. If more than one HPC Server cluster head node is found in the Active Directory, an error is thrown listing the names of all the head nodes found. out = findResource('scheduler', 'type', 'hpcserver', 'SchedulerHostname', 'headNode') uses the HPC Server scheduler
14-106
findResource
with the specified head node. SchedulerHostname is ignored for all scheduler types other than 'hpcserver'.
out = findResource('worker') or out = findResource('worker', 'LookupURL', 'host:port') returns an array of worker objects
representing all found workers or those that match the search criteria.
out = findResource(... ,'p1', v1, 'p2', v2,...) returns an array, out, of resources whose property names and property values match those passed as parameter-value pairs, p1, v1, p2, v2.
Note that the property value pairs can be in any format supported by the set function. When a property value is specified, it must use the same exact value that the get function returns, including letter case. For example, if get returns the Name property value as 'MyJobManager', then findResource will not find that object if searching for a Name property value of 'myjobmanager'.
Tips
You are allowed to use parameter-value string pairs, structures, parameter-value cell array pairs, and configurations in the same call to findResource. Find a particular job manager by its name.
jm1 = findResource('scheduler','type','jobmanager', ... 'Name', 'ClusterQueue1');
Examples
14-107
findResource
Find all job managers accessible from the lookup service on a particular host.
jms = findResource('scheduler','type','jobmanager', ... 'LookupURL','host234');
Find a particular job manager accessible from the lookup service on a particular host. In this example, subnet2.hostalpha port 6789 is where the lookup is performed, but the job manager named SN2Jmgr might be running on another machine.
jm = findResource('scheduler','type','jobmanager', ... 'LookupURL', 'subnet2.hostalpha:6789', 'Name', 'SN2JMgr');
Create a local scheduler that will start workers on the client machine for running your job.
local_sched = findResource('scheduler','type','local')
Find the scheduler identified by the default parallel configuration, with the scheduler object properties set to the values defined in that configuration.
sched = findResource();
See Also
findJob | findTask
14-108
findTask
Purpose Syntax
Arguments
Job object. Returned task objects. Array of tasks in job obj whose State is pending. Array of tasks in job obj whose State is running. Array of completed tasks in job obj, i.e., those whose State is finished or failed. Task object properties to match. Values for corresponding object properties.
Description
tasks = findTask(obj) gets a 1-by-N array of task objects belonging to a job object obj Tasks in the array are ordered by the ID property of the tasks, indicating the sequence in which they were created. [pending running completed] = findTask(obj) returns arrays of all task objects stored in the job object obj, sorted by state. Within each array (pending, running, and completed), the tasks are returned in sequence of creation. tasks = findTask(obj,'p1',v1,'p2',v2,...) gets a 1-by-N array of task objects belonging to a job object obj. The returned task objects will be only those having the specified property-value pairs.
Note that the property value pairs can be in any format supported by the set function, i.e., param-value string pairs, structures, and param-value cell array pairs. If a structure is used, the structure
14-109
findTask
field names are object property names and the field values are the appropriate property values to match. When a property value is specified, it must use the same exact value that the get function returns, including letter case. For example, if get returns the Name property value as MyTask, then findTask will not find that object while searching for a Name property value of mytask.
Tips
If obj is contained in a remote service, findTask will result in a call to the remote service. This could result in findTask taking a long time to complete, depending on the number of tasks retrieved and the network speed. Also, if the remote service is no longer available, an error will be thrown. Create a job object.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); obj = createJob(jm);
Examples
Create the task object t, which refers to the task we just added to obj.
t = findTask(obj)
See Also
14-110
for
Purpose Syntax
for-loop over distributed range FOR variable = drange(colonop) statement ... statement end
Description
The colonop is an expression of the form start:increment:finish or start:finish. The default value of increment is 1. The colonop is partitioned by codistributed.colon into numlabs contiguous segments of nearly equal length. Each segment becomes the iterator for a conventional for-loop on an individual lab. The most important property of the loop body is that each iteration must be independent of the other iterations. Logically, the iterations can be done in any order. No communication with other labs is allowed within the loop body. The functions that perform communication are gop, gcat, gplus, codistributor, codistributed, gather, and redistribute. It is possible to access portions of codistributed arrays that are local to each lab, but it is not possible to access other portions of codistributed arrays. The break statement can be used to terminate the loop prematurely.
14-111
for
Examples
Find the rank of magic squares. Access only the local portion of a codistributed array.
r = zeros(1, 40, codistributor()); for n = drange(1:40) r(n) = rank(magic(n)); end r = gather(r);
Perform Monte Carlo approximation of pi. Each lab is initialized to a different random number state.
m = 10000; for p = drange(1:numlabs) z = rand(m, 1) + i*rand(m, 1); c = sum(abs(z) < 1) end k = gplus(c) p = 4*k/(m*numlabs);
Attempt to compute Fibonacci numbers. This will not work, because the loop bodies are dependent.
f = zeros(1, 50, codistributor()); f(1) = 1; f(2) = 2; for n = drange(3:50) f(n) = f(n - 1) + f(n - 2) end
See Also
14-112
gather
parallel job to gather together the data of a codistributed array, or outside an spmd statement to gather the data of a distributed array. If you execute this inside an spmd statement, pmode, or parallel job, X is replicated array with all the data of the array on every lab. If you execute this outside an spmd statement, X is an array in the local workspace, with the data transferred from the multiple labs.
X = gather(distributed(X)) or X = gather(codistributed(X)) returns the original array X. X = gather(C, lab) converts a codistributed array C to a variant array X, such that all of the data is contained on lab lab, and X is a 0-by-0 empty double on all other labs.
For a GPUArray input, X = gather(A) transfers the data from the GPU to the local workspace. If the input argument to gather is not a distributed, a codistributed, or a GPUArray, the output is the same as the input.
Tips
Note that gather assembles the codistributed or distributed array in the workspaces of all the labs on which it executes, or on the MATLAB client, respectively, but not both. If you are using gather within an spmd statement, the gathered array is accessible on the client via its corresponding Composite object; see Accessing Data with Composites on page 3-7. If you are running gather in a parallel job, you can return the gathered array to the client as an output argument from the task. As the gather function requires communication between all the labs, you cannot gather data from all the labs onto a single lab by placing the function inside a conditional statement such as if labindex == 1.
14-113
gather
Examples
Distribute a magic square across your labs, then gather the whole matrix onto every lab and then onto the client. This code results in the equivalent of M = magic(n) on all labs and the client.
n = 10; spmd C = codistributed(magic(n)); M = gather(C) % Gather data on all labs end S = gather(C) % Gather data on client
Gather all of the data in C onto lab 1, for operations that cannot be performed across distributed arrays.
n = 10; spmd C = codistributed(magic(n)); out = gather(C, 1); if labindex == 1 % Characteristic sum for this magic square: characteristicSum = sum(1:n^2)/n; % Ensure that the diagonal sums are equal to the % characteristic sum: areDiagonalsEqual = isequal ... (trace(out), trace(flipud(out)), characteristicSum) end end Lab 1: areDiagonalsEqual = 1
Gather all of the data from a distributed array into D on the client.
n = 10; D = distributed(magic(n)); % Distribute data to labs M = gather(D) % Return data to client
14-114
gather
G = gpuArray(rand(1024,1)); F = sqrt(G); %input and output both GPUArray W = gather(G); % Return data to client whos Name Size Bytes Class F G W 1024x1 1024x1 1024x1 108 108 8192 parallel.gpu.GPUArray parallel.gpu.GPUArray double
See Also
14-115
gcat
Purpose Syntax
Global concatenation
Xs = gcat(X) Xs = gcat(X, dim) Xs = gcat(X, dim, targetlab) Xs = gcat(X) concatenates the variant array X from each lab in the second dimension. The result is replicated on all labs. Xs = gcat(X, dim) concatenates the variant array X from each lab in the dimension indicated by dim. Xs = gcat(X, dim, targetlab) performs the reduction, and places the result into res only on the lab indicated by targetlab. res is set to [] on all other labs.
Description
Examples
See Also
14-116
get
Purpose Syntax
Object properties
get(obj) out = get(obj) out = get(obj,'PropertyName')
Arguments
An object or an array of objects. A property name or a cell array of property names. A single property value, a structure of property values, or a cell array of property values.
Description
get(obj) returns all property names and their current values to the command line for obj. out = get(obj) returns the structure out where each field name is the name of a property of obj, and each field contains the value of that
property.
out = get(obj,'PropertyName') returns the value out of the property specified by PropertyName for obj. If PropertyName is replaced by a 1-by-n or n-by-1 cell array of strings containing property names, then get returns a 1-by-n cell array of values to out. If obj is an array of objects, then out will be an m-by-n cell array of property values where m is equal to the length of obj and n is equal to the number of properties specified.
Tips
When specifying a property name, you can do so without regard to case, and you can make use of property name completion. For example, if jm is a job manager object, then these commands are all valid and return the same result.
out = get(jm,'HostAddress'); out = get(jm,'hostaddress'); out = get(jm,'HostAddr');
14-117
get
Examples
This example illustrates some of the ways you can use get to return property values for the job object j1.
get(j1,'State') ans = pending get(j1,'Name') ans = MyJobManager_job out = get(j1); out.State ans = pending out.Name ans = MyJobManager_job two_props = {'State' 'Name'}; get(j1, two_props) ans = 'pending' 'MyJobManager_job'
See Also
inspect | set
14-118
getAllOutputArguments
obj data
Description
data = getAllOutputArguments(obj) returns data, the output data contained in the tasks of a finished job. If the job has M tasks, each row of the M-by-N cell array data contains the output arguments for the corresponding task in the job. Each row has N columns, where N is the
greatest number of output arguments from any one task in the job. The N elements of a row are arrays containing the output arguments from that task. If a task has less than N output arguments, the excess arrays in the row for that task are empty. The order of the rows in data will be the same as the order of the tasks contained in the job.
Tips
If you are using a job manager, getAllOutputArguments results in a call to a remote service, which could take a long time to complete, depending on the amount of data being retrieved and the network speed. Also, if the remote service is no longer available, an error will be thrown. Note that issuing a call to getAllOutputArguments will not remove the output data from the location where it is stored. To remove the output data, use the destroy function to remove the individual task or their parent job object. The same information returned by getAllOutputArguments can be obtained by accessing the OutputArguments property of each task in the job.
Examples
14-119
getAllOutputArguments
See Also
submit
14-120
getCodistributor
Examples
Get the codistributor object for a 1-D codistributed array that uses default distribution on 4 labs:
spmd (4) I1 = codistributed.eye(64, codistributor1d()); codist1 = getCodistributor(I1) dim = codist1.Dimension partn = codist1.Partition end
Get the codistributor object for a 2-D block cyclic codistributed array that uses default distribution on 4 labs:
spmd (4) I2 = codistributed.eye(128, codistributor2dbc()); codist2 = getCodistributor(I2) blocksz = codist2.BlockSize partn = codist2.LabGrid ornt = codist2.Orientation end
14-121
getCodistributor
isComplete(codist2) end
See Also
14-122
getCurrentJob
job
The job object that contains the task currently being evaluated by the worker session.
job = getCurrentJob returns the job object that is the Parent of the
task currently being evaluated by the worker session. If the function is executed in a MATLAB session that is not a worker, you get an empty result.
getCurrentJobmanager | getCurrentTask | getCurrentWorker | getFileDependencyDir
14-123
getCurrentJobmanager
jm
The job manager object that scheduled the task currently being evaluated by the worker session.
Description
jm = getCurrentJobmanager returns the job manager object that has sent the task currently being evaluated by the worker session. jm is the Parent of the tasks parent job.
Tips
If the function is executed in a MATLAB session that is not a worker, you get an empty result. If your tasks are scheduled by a third-party scheduler instead of a job manager, getCurrentJobmanager returns a distcomp.taskrunner object.
See Also
14-124
getCurrentTask
task
evaluated by the worker session. If the function is executed in a MATLAB session that is not a worker, you get an empty result.
getCurrentJob | getCurrentJobmanager | getCurrentWorker | getFileDependencyDir
14-125
getCurrentWorker
worker
The worker object that is currently evaluating the task that contains this function.
Description Tips
worker = getCurrentWorker returns the worker object representing the session that is currently evaluating the task that calls this function.
If the function is executed in a MATLAB session that is not a worker or if you are using a third-party scheduler instead of a job manager, you get an empty result. Create a job with one task, and have the task return the name of the worker that evaluates it.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); t = createTask(j, @() get(getCurrentWorker,'Name'), 1, {}); submit(j) waitForState(j) get(t,'OutputArgument') ans = 'c5_worker_43'
Examples
The function of the task t is an anonymous function that first executes getCurrentWorker to get an object representing the worker that is evaluating the task. Then the task function uses get to examine the Name property value of that object. The result is placed in the OutputArgument property of the task.
See Also
14-126
getDebugLog
Read output messages from job run by supported third-party or local scheduler
str = getDebugLog(sched, job_or_task)
str sched
Variable to which messages are returned as a string expression. Scheduler object referring to mpiexec, Microsoft Windows HPC Server (or CCS), Platform LSF, PBS Pro, or TORQUE scheduler, created by findResource. Object identifying job, parallel job, or task whose messages you want.
job_or_task
Description
str = getDebugLog(sched, job_or_task) returns any output written to the standard output or standard error stream by the job or task identified by job_or_task, being run by the scheduler identified by sched. You cannot use this function to retrieve messages from a task if the scheduler is mpiexec.
Examples
Construct a scheduler object so you can create a parallel job. Assume that you have already defined a configuration called mpiexec to define the properties of the scheduler object.
mpiexecObj = findResource('scheduler', 'Configuration', 'mpiexec');
14-127
getDebugLog
See Also
14-128
getFileDependencyDir
depdir
Description
depdir = getFileDependencyDir returns a string, which is the path to the local directory into which FileDependencies are written. This function will return an empty array if it is not called on a MATLAB worker.
Examples
See Also
14-129
getJobSchedulerData
Information that was previously stored for this job. Scheduler object identifying the generic third-party scheduler running the job. Job object identifying the job for which to retrieve data.
Description
userdata = getJobSchedulerData(sched, job) returns data stored for the job job that was derived from the generic scheduler sched. The information was originally stored with the function setJobSchedulerData. For example, it might be useful to store the
third-party schedulers external ID for this job, so that the function specified in GetJobStateFcn can later query the scheduler about the state of the job. To use this feature, you should call the function setJobSchedulerData in the submit function (identified by the SubmitFcn property) and call getJobSchedulerData in any of the functions identified by the properties GetJobStateFcn, DestroyJobFcn, DestroyTaskFcn, CancelJobFcn, or CancelTaskFcn. For more information and examples on using these functions and properties, see Manage Jobs with Generic Scheduler on page 8-50.
See Also
setJobSchedulerData
14-130
getLocalPart
returns
Lab Lab Lab Lab 1: 2: 3: 4: L L L L = = = = [16 2 3 13] [ 5 11 10 8] [ 9 7 6 12] [ 4 14 15 1]
See Also
codistributed | codistributor
14-131
globalIndices
Purpose Syntax
Description
indices for a codistributed array without having to create the array itself.
[E,F] = codist.globalIndices(dim, lab) is the same as [E,F] = globalIndices(R, dim, lab), where codist is the codistributor for R, or codist = getCodistributor(R). This allows you to get the global
indices for a codistributed array without having to create the array itself.
Examples
Create a 2-by-22 codistributed array among four labs, and view the global indices on each lab:
14-132
globalIndices
spmd C = codistributed.zeros(2, 22, codistributor1d(2,[6 6 5 5])); if labindex == 1 K = globalIndices(C, 2); elseif labindex == 2 [E,F] = globalIndices(C, 2); % returns E = 7, F = 12. end K = globalIndices(C, 2, 3); end % returns K = 13:17. [E,F] = globalIndices(C, 2, 4); % returns E = 18, F = 22. % returns K = 1:6.
Use globalIndices to load data from a file and construct a codistributed array distributed along its columns, i.e., dimension 2. Notice how globalIndices makes the code not specific to the number of labs and alleviates you from calculating offsets or partitions.
spmd siz = [1000, 1000]; codistr = codistributor1d(2, [], siz); % Use globalIndices to figure out which columns % each lab should load. [firstCol, lastCol] = codistr.globalIndices(2); % Call user-defined function readRectangleFromFile to % load all the values that should go into % the local part for this lab. labLocalPart = readRectangleFromFile(fileName, ... 1, siz(1), firstCol, lastCol); % With the local part and codistributor, % construct the corresponding codistributed array. C = codistributed.build(labLocalPart, codistr); end
See Also
getLocalPart | labindex
14-133
gop
Purpose Syntax
Arguments
F x res
Function to operate across labs. Argument to function F, should be same variable on all labs, but can have different values. Variable to hold reduction result.
Description
res = gop(@F, x) is the reduction via the function F of the quantities x from each lab. The result is duplicated on all labs.
The function F(x,y) should accept two arguments of the same type and produce one result of that type, so it can be used iteratively, that is,
F(F(x1,x2),F(x3,x4))
Examples
14-134
gop
See Also
labBarrier | numlabs
14-135
gplus
Global addition
S = gplus(X) S = gplus(X, targetlab) S = gplus(X) returns the addition of the variant array X from each lab. The result S is replicated on all labs. S = gplus(X, targetlab) performs the addition, and places the result into S only on the lab indicated by targetlab. S is set to [] on all other labs.
Examples
See Also
gop | labindex
14-136
gpuArray
defined for GPUArray objects in Using GPUArray on page 10-4. The MATLAB data X must be numeric (for example: single, double, int8, etc.) or logical, and the GPU device must have sufficient free memory to store the data. X must be a full matrix, not sparse. If the input argument is already a GPUArray, the output is the same as the input.
Examples
Transfer a 10-by-10 matrix of random single-precision values to the GPU, then use the GPU to square each element.
X = rand(10, 'single'); G = gpuArray(X); isequal(gather(G), X) % classUnderlying(G) % G2 = G .* G % %
Returns true Returns 'single' Uses times method defined for GPUArray objects
See Also
14-137
gpuDevice
GPU device.
D = gpuDevice(IDX) selects the GPU device specified by index IDX. IDX must be in the range of 1 to gpuDeviceCount. A warning or error
Examples
See Also
14-138
gpuDeviceCount
your computer. Determine how many GPU devices you have available in your computer and examine the properties of each.
n = gpuDeviceCount; for ii = 1:n gpuDevice(ii) end
See Also
14-139
help
class
A Parallel Computing Toolbox object class: distcomp.jobmanager, distcomp.job, or distcomp.task. A function for the specified class. To see what functions are available for a class, see the methods reference page.
function
Description
help class/function returns command-line help for the specified function of the given class.
If you do not know the class for the function, use class(obj), where function is of the same class as the object obj.
Examples
Get help on functions from each of the Parallel Computing Toolbox object classes.
help distcomp.jobmanager/createJob help distcomp.job/cancel help distcomp.task/waitForState class(j1) ans = distcomp.job help distcomp.job/createTask
See Also
methods
14-140
importParallelConfig
The importParallelConfig function allows you to import a configuration that was stored in a .mat file.
configname = importParallelConfig(filename) imports the configuration stored in the specified file and returns the name of the imported configuration as a string assigned to configname. If a configuration with the same name already exists in your MATLAB session, an extension is added to the name of the imported configuration. If filename has no extension, .mat is assumed. Each configuration .mat file contains only one configuration.
You can use the imported configuration with any functions that support configurations. importParallelConfig does not set the imported configuration as the default; you can set it as the default configuration with the defaultParallelConfig function. To export a configuration, use the Configurations Manager, which you can open by selecting Parallel > Manage Configurations. Configurations exported from earlier versions of the product are upgraded during the import. Configurations that you import with importParallelConfig are saved as a part of your MATLAB preferences, so these configurations are available in your subsequent MATLAB sessions without importing them again.
Examples
Import a configuration from the file Config01.mat and use it to open a pool of MATLAB workers:
conf_1 = importParallelConfig('Config01') matlabpool('open', conf_1)
Import a configuration from the file ConfigMaster.mat and set it as the default parallel configuration:
14-141
importParallelConfig
See Also
defaultParallelConfig
14-142
inspect
obj
inspect(obj) opens the Property Inspector and allows you to inspect and set properties for the object obj.
You can also open the Property Inspector via the Workspace browser by double-clicking an object. The Property Inspector does not automatically update its display. To refresh the Property Inspector, open it again. Note that properties that are arrays of objects are expandable. In the figure of the example below, the Tasks property is expanded to enumerate the individual task objects that make up this property. These individual task objects can also be expanded to display their own properties.
14-143
inspect
Examples
See Also
get | set
14-144
isaUnderlying
Examples
distributed.ones(1, N, 'uint8'); distributed.cell(1, N); isaUnderlying(D_uint8, 'uint8') % returns true isaUnderlying(D_cell, 'double') % returns false
See Also
14-145
iscodistributed
Examples
See Also
isdistributed
14-146
isComplete
See Also
14-147
isdistributed
otherwise. For a description of a distributed array, see Nondistributed Versus Distributed Arrays on page 5-2. With an open MATLAB pool,
L = ones(100, 1); D = distributed.ones(100, 1); isdistributed(L) % returns false isdistributed(D) % returns true
Examples
See Also
iscodistributed
14-148
isreplicated
otherwise. For a description of a replicated array, see Nondistributed Versus Distributed Arrays on page 5-2. isreplicated also returns true for a Composite X if all its elements are identical.
isreplicated(X) requires checking for equality of the array X across
Tips
Examples
See Also
iscodistributed | isdistributed
14-149
jobStartup
job
worker evaluates a task for a particular job. You do not call this function from the client session, nor explicitly as part of a task function. You add MATLAB code to the jobStartup.m file to define job initialization actions on the worker. The worker looks for jobStartup.m in the following order, executing the one it finds first:
1 Included in the jobs FileDependencies property. 2 In a folder included in the jobs PathDependencies property. 3 In the workers MATLAB installation at the location
matlabroot/toolbox/distcomp/user/jobStartup.m
To create a version of jobStartup.m for FileDependencies or PathDependencies, copy the provided file and modify it as required. For further details on jobStartup and its implementation, see the text in the installed jobStartup.m file.
See Also
14-150
labBarrier
shared resources such as file I/O. For a demonstration that uses labSend, labReceive, labBarrier, and labSendReceive, see the demo Profiling Explicit Parallel Communication.
Examples
Lab 1 writes some data to the file, which all other labs will read.
if labindex == 1 data = randn(100, 1); save(fname, 'data'); pause(5) %allow time for file to become available to other labs end
All labs wait until all have reached the barrier; this ensures that no lab attempts to load the file until lab 1 writes to it.
labBarrier; load(fname);
See Also
14-151
labBroadcast
Purpose Syntax
Arguments
senderlab data
The labindex of the lab sending the broadcast. The data being broadcast. This argument is required only for the lab that is broadcasting. The absence of this argument indicates that a lab is receiving. The broadcast data as it is received on all other labs.
shared_data
Description
shared_data = labBroadcast(senderlab, data) sends the specified data to all executing labs. The data is broadcast from the lab with labindex == senderlab, and received by all other labs. shared_data = labBroadcast(senderlab) receives on each executing lab the specified shared_data that was sent from the lab whose labindex is senderlab.
If labindex is not senderlab, then you do not include the data argument. This indicates that the function is to receive data, not broadcast it. The received data, shared_data, is identical on all labs. This function blocks execution until the labs involvement in the collective broadcast operation is complete. Because some labs may complete their call to labBroadcast before others have started, use labBarrier to guarantee that all labs are at the same point in a program.
Examples
14-152
labBroadcast
See Also
14-153
labindex
See Also
14-154
labProbe
Purpose Syntax
Arguments
Tag defined by the sending labs labSend function to identify particular data. String to indicate that all labs should be tested for a message. Boolean indicating if a message is ready to be received.
Description
is_data_available = labProbe returns a logical value indicating whether any data is available for this lab to receive with the labReceive function. is_data_available = labProbe(source) tests for a message only
See Also
14-155
labReceive
Purpose Syntax
Arguments
receive data. Tag defined by the sending labs labSend function to identify particular data. String to indicate that data can come from any lab. Data sent by the sending labs labSend function.
Description
data = labReceive receives data from any lab with any tag. data = labReceive(source) receives data from the specified lab with
any tag
data = labReceive('any',tag) receives data from any lab with the
specified tag.
data = labReceive(source,tag) receives data from only the specified
the data.
Tips
This function blocks execution in the lab until the corresponding call to labSend occurs in the sending lab. For a demonstration that uses labSend, labReceive, labBarrier, and labSendReceive, see the demo Profiling Explicit Parallel Communication.
14-156
labReceive
See Also
14-157
labSend
Purpose Syntax
Arguments
Description
labSend(data, destination) sends the data to the specified destination, with a tag of 0. labSend(data, destination, tag) sends the data to the specified destination with the specified tag. data can be any MATLAB data type. destination identifies the labindex of the receiving lab, and must be either a scalar or a vector of integers between 1 and numlabs; it cannot be labindex (i.e., the current lab). tag can be any integer from 0 to 32767.
Tips
This function might return before the corresponding labReceive completes in the receiving lab. For a demonstration that uses labSend, labReceive, labBarrier, and labSendReceive, see the demo Profiling Explicit Parallel Communication.
See Also
14-158
labSendReceive
Purpose Syntax
Arguments
Data on the sending lab that is sent to the receiving lab; any MATLAB data type. Data accepted on the receiving lab.
labindex of the lab to which data is sent. labindex of the lab from which data is received.
Description
received = labSendReceive(labTo, labFrom, data) sends data to the lab whose labindex is labTo, and receives received from the lab whose labindex is labFrom. labTo and labFrom must be scalars. This
with the important exception that both the sending and receiving of data happens concurrently. This can eliminate deadlocks that might otherwise occur if the equivalent call to labSend would block. If labTo is an empty array, labSendReceive does not send data, but only receives. If labFrom is an empty array, labSendReceive does not receive data, but only sends.
received = labSendReceive(labTo, labFrom, data, tag) uses the specified tag for the communication. tag can be any integer from 0 to 32767.
For a demonstration that uses labSend, labReceive, labBarrier, and labSendReceive, see the demo Profiling Explicit Parallel Communication.
14-159
labSendReceive
Examples
Create a unique set of data on each lab, and transfer each labs data one lab to the right (to the next higher labindex). First use magic to create a unique value for the variant array mydata on each lab.
mydata = Lab 1: mydata 1 Lab 2: mydata 1 4 Lab 3: mydata 8 3 4 magic(labindex) =
= 3 2 = 1 5 9 6 7 2
Define the lab on either side, so that each lab will receive data from the lab on the left while sending data to the lab on the right, cycling data from the end lab back to the beginning lab.
labTo = mod(labindex, numlabs) + 1; % one lab to the right labFrom = mod(labindex - 2, numlabs) + 1; % one lab to the left
Transfer the data, sending each labs mydata into the next labs otherdata variable, wrapping the third labs data back to the first lab.
otherdata = labSendReceive(labTo, labFrom, mydata) Lab 1: otherdata = 8 1 6 3 5 7 4 9 2 Lab 2: otherdata =
14-160
labSendReceive
1 Lab 3: otherdata = 1 3 4 2
Transfer data to the next lab without wrapping data from the last lab to the first lab.
if labindex < numlabs; labTo = labindex + 1; else labTo = []; end; if labindex > 1; labFrom = labindex - 1; else labFrom = []; end; otherdata = labSendReceive(labTo, labFrom, mydata) Lab 1: otherdata = [] Lab 2: otherdata = 1 Lab 3: otherdata = 1 4 3 2
See Also
14-161
length
obj
See Also
size
14-162
load
Purpose Syntax
Arguments
Job from which to load workspace variables. Variables to load from the job. Wildcards allow pattern matching in MAT-file style. Indication to use regular expression pattern matching. Struct containing the variables after loading.
Description
load(job) retrieves all variables from a batch job and assigns them into the current workspace. load throws an error if the batch runs a function (instead of a script), the job is not finished, or the job encountered an error while running, . load(job, 'X') loads only the variable named X from the job. load(job, 'X', 'Y', 'Z*') loads only the specified variables. The wildcard '*' loads variables that match a pattern (MAT-file only). load(job, '-regexp', 'PAT1', 'PAT2') can be used to load all
variables matching the specified patterns using regular expressions. For more information on using regular expressions, type doc regexp at the command prompt.
S = load(job ...) returns the contents of job into variable S, which
14-163
load
Examples
Run a batch job and load its results into your client workspace.
j = batch('myScript'); wait(j) load(j)
See Also
batch | getAllOutputArguments
14-164
matlabpool
Purpose Syntax
Description
parallel configuration, with the pool size specified by that configuration. (For information about setting up and selecting parallel configurations, see Parallel Configurations for Cluster Access on page 6-14.) You can also specify the pool size using matlabpool open poolsize, but most schedulers have a maximum number of processes that they can start (12 for a local scheduler). If the configuration specifies a job manager as the scheduler, matlabpool reserves its workers from among those already
14-165
matlabpool
running and available under that job manager. If the configuration specifies a third-party scheduler, matlabpool instructs the scheduler to start the workers.
matlabpool open configname or matlabpool open configname poolsize starts a worker pool using the Parallel Computing Toolbox user configuration identified by configname rather than the default
configuration to locate a scheduler. If the pool size is specified, it overrides the maximum and minimum number of workers specified in the configuration, and starts a pool of exactly that number of workers, even if it has to wait for them to be available. Without specifying open or close, the command default is open. So, matlabpool poolsize, matlabpool(poolsize), matlabpool configname, and matlabpool configname poolsize operate as matlabpool open ..., and are provided for convenience.
matlabpool(schedobj) or matlabpool(schedobj, 'open') is the same as matlabpool open, except that the worker pool is started on the scheduler specified by the object schedobj. matlabpool(schedobj, 'open', ...) is the same as matlabpool('open', ...) except that the worker pool is started on the scheduler specified by the object schedobj. matlabpool(schedobj, poolsize) is the same as matlabpool poolsize except that the worker pool is started on the scheduler specified by the object schedobj. matlabpool close stops the worker pool, destroys the parallel job, and
makes all parallel language features revert to using the MATLAB client for computing their results.
matlabpool close force destroys all parallel jobs created by matlabpool for the current user under the scheduler specified by the
14-166
matlabpool
matlabpool size returns the size of the worker pool if it is open, or 0 if the pool is closed. matlabpool('open', ...) and matlabpool('close', ...) can be
invoked as functions with optional arguments, such as configuration name and pool size. The default is 'open'. For example, the following are equivalent:
matlabpool open MyConfig 4 matlabpool('MyConfig', 4) matlabpool('open',..., 'FileDependencies', filecell) starts a
worker pool and allows you to specify file dependencies so that you can pass necessary files to the workers in the pool. The cell array filecell is appended to the FileDependencies specified in the configuration used for startup. The 'FileDependencies' property name is case sensitive, and must appear as shown. (Note: This form of the command does not allow you to directly specify any other job property-value pairs when opening a pool.)
matlabpool('addfiledependencies', filecell) allows you to add extra file dependencies to an already running pool. filecell is a cell
array of strings, identical in form to those you use when adding file dependencies to a job or when you open a MATLAB pool. Each string can specify either absolute or relative files, directories, or a file on the MATLAB path. The command transfers the files to each worker, placing the files in the file dependencies directory, exactly the same as if you set them at the time the pool was opened.
matlabpool updatefiledependencies checks all the file dependencies
of the current pool to see if they have changed, and replicates any changes to each of the labs in the pool. In this way, you can send code changes out to remote labs. This checks dependencies that you added with the matlabpool addfiledependencies command as well as those you specified when the pool was started (by a configuration or command-line argument).
Tips
When a pool of workers is open, the following commands entered in the clients Command Window also execute on all the workers:
14-167
matlabpool
cd addpath rmpath This enables you to set the working directory and the path on all the workers, so that a subsequent parfor-loop executes in the proper context. If any of these commands does not work on the client, it is not executed on the workers either. For example, if addpath specifies a directory that the client cannot see or access, the addpath command is not executed on the workers. However, if the working directory or path can be set on the client, but cannot be set as specified on any of the workers, you do not get an error message returned to the client Command Window. This slight difference in behavior is an issue especially in a mixed-platform environment where the client is not the same platform as the workers, where directories local to or mapped from the client are not available in the same way to the workers, or where directories are in a nonshared file system. For example, if you have a MATLAB client running on a Microsoft Windows operating system while the MATLAB workers are all running on Linux operating systems, the same argument to addpath cannot work on both. In this situation, you can use the function pctRunOnAll to assure that a command runs on all the workers. Another difference between client and workers is that any addpath arguments that are part of the matlabroot folder are not set on the workers. The assumption is that the MATLAB install base is already included in the workers paths. The rules for addpath regarding workers in the pool are: Subfolders of the matlabroot folder are not sent to the workers. Any folders that appear before the first occurrence of a matlabroot folder are added to the top of the path on the workers.
14-168
matlabpool
Any folders that appear after the first occurrence of a matlabroot folder are added after the matlabroot group of folders on the workers paths. For example, suppose that matlabroot on the client is C:\Applications\matlab\. With an open MATLAB pool, execute the following to set the path on the client and all workers:
addpath('P1', 'P2', 'C:\Applications\matlab\T3', 'C:\Applications\matlab\T4', 'P5', 'C:\Applications\matlab\T6', 'P7', 'P8');
Because T3, T4, and T6 are subfolders of matlabroot, they are not set on the workers paths. So on the workers, the pertinent part of the path resulting from this command is:
P1 P2 <worker original matlabroot folders...> P5 P7 P8
Examples
Start a pool using the default configuration to define the number of labs:
matlabpool
14-169
matlabpool
Run matlabpool as a function to check whether the worker pool is currently open:
isOpen = matlabpool('size') > 0
Start a pool with the default configuration, and pass two code files to the workers:
matlabpool('open', 'FileDependencies', {'mod1.m', 'mod2.m'})
Start a MATLAB pool with the scheduler and pool size determined by the default configuration:
schd = findResource('scheduler', ... 'configuration', defaultParallelConfig); matlabpool(schd)
See Also
14-170
methods
Purpose Syntax
Arguments
obj out
Description
methods(obj) returns the names of all methods for the class of which obj is an instance. out = methods(obj) returns the names of the methods as a cell array
of strings.
Examples
Create job manager, job, and task objects, and examine what methods are available for each.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); methods(jm) Methods for class distcomp.jobmanager: createJob demote pause resume createParallelJob findJob promote j1 = createJob(jm); methods(j1) Methods for class distcomp.job: cancel destroy getAllOutputArguments createTask findTask submit t1 = createTask(j1, @rand, 1, {3}); methods(t1) Methods for class distcomp.task: cancel destroy waitForState
waitForState
14-171
methods
See Also
help | get
14-172
mpiLibConf
primaryLib extras
MPI implementation library used by a parallel job. Cell array of other required library names.
Description
[primaryLib, extras] = mpiLibConf returns the MPI implementation library to be used by a parallel job. primaryLib is the name of the shared library file containing the MPI entry points. extras is a cell array of other library names required by the MPI library.
To supply an alternative MPI implementation, create a file named mpiLibConf.m, and place it on the MATLAB path. The recommended location is matlabroot/toolbox/distcomp/user. Your mpiLibConf.m file must be higher on the cluster workers path than matlabroot/toolbox/distcomp/mpi. (Sending mpiLibConf.m as a file dependency for this purpose does not work.)
Tips
Under all circumstances, the MPI library must support all MPI-1 functions. Additionally, the MPI library must support null arguments to MPI_Init as defined in section 4.2 of the MPI-2 standard. The library must also use an mpi.h header file that is fully compatible with MPICH2. When used with the MathWorks job manager or the local scheduler, the MPI library must support the following additional MPI-2 functions: MPI_Open_port MPI_Comm_accept MPI_Comm_connect
14-173
mpiLibConf
When used with any third-party scheduler, it is important to launch the workers using the version of mpiexec corresponding to the MPI library being used. Also, you might need to launch the corresponding process management daemons on the cluster before invoking mpiexec.
Examples
Use the mpiLibConf function to view the current MPI implementation library:
mpiLibConf mpich2.dll
14-174
mpiprofile
Purpose Syntax
Description
statistics on execution time and communication times. The statistics are collected in a manner similar to running the profile command on each MATLAB worker. By default, the parallel profiling extensions include array fields that collect information on communication with each of the other labs. This command in general should be executed in pmode or as part of a task in a parallel job.
mpiprofile on <options> starts the parallel profiler and clears
Option
-detail mmex -detail builtin
Description This option specifies the set of functions for which profiling statistics are gathered. -detail mmex (the default) records information about functions, subfunctions, and MEX-functions. -detail builtin additionally
14-175
mpiprofile
Option
fields, which reduces the profiling overhead. If you have a very large cluster, you might want to use this option; however, you will not get all the detailed inter-lab communication plots in the viewer. For information about the structure of returned data, see mpiprofile info below.
-history -nohistory -historysize <size> mpiprofile supports these options in the same way as the standard profile.
No other profile options are supported by mpiprofile. These three options have no effect on the data displayed by mpiprofile viewer.
mpiprofile off stops the parallel profiler. To reset the state of the profiler and disable collecting communication information, you should also call mpiprofile reset.
14-176
mpiprofile
recorded function statistics. This works only in pmode or in the same MATLAB worker session.
mpiprofile clear clears the profile information. mpiprofile status returns a valid status when it runs on the worker. mpiprofile reset turns off the parallel profiler and resets the data collection back to the standard profiler. If you do not call reset,
Field
BytesSent BytesReceived TimeWasted CommTime CommTimePerLab TimeWastedPerLab
Description Records the quantity of data sent Records the quantity of data received Records communication waiting time Records the communication time Vector of communication receive time for each lab Vector of communication waiting time for each lab
The three *PerLab fields are collected only on a per-function basis, and can be turned off by typing the following command in pmode:
mpiprofile on -messagedetail simplified mpiprofile viewer is used in pmode after running user code with mpiprofile on. Calling the viewer stops the profiler and opens the
14-177
mpiprofile
report displayed in the profiler window. The file listing at the bottom of the function profile page shows several columns to the left of each line of code. In the summary page: Column 1 indicates the number of calls to that line. Column 2 indicates total time spent on the line in seconds. Columns 36 contain the communication information specific to the parallel profiler
mpiprofile('viewer', <profinfoarray>) in function form can be used from the client. A structure <profinfoarray> needs be passed in as the second argument, which is an array of mpiprofile info structures. See pInfoVector in the Examples section below. mpiprofile does not accept -timer clock options, because the communication timer clock must be real.
For more information and examples on using the parallel profiler, see Profiling Parallel Code on page 6-35.
Examples
In pmode, turn on the parallel profiler, run your function in parallel, and call the viewer:
mpiprofile on; % call your function; mpiprofile viewer;
If you want to obtain the profiler information from a parallel job outside of pmode (i.e., in the MATLAB client), you need to return output arguments of mpiprofile info by using the functional form of the command. Define your function foo(), and make it the task function in a parallel job:
function [pInfo, yourResults] = foo mpiprofile on initData = (rand(100, codistributor()) ... * rand(100, codistributor()));
14-178
mpiprofile
After the job runs and foo() is evaluated on your cluster, get the data on the client:
A = getAllOutputArguments(yourJob);
See Also
14-179
mpiSettings
Purpose Syntax
Description
enable deadlock detection on all labs, this is the most useful option. The default value is 'off' for parallel jobs, and 'on' inside pmode sessions or spmd statements. Once the setting has been changed within a pmode session or an spmd statement, the setting stays in effect until either the pmode session ends or the MATLAB pool is closed.
mpiSettings('MessageLogging','on') turns on MPI message logging. The default is 'off'. The default destination is the MATLAB Command
Window.
mpiSettings('MessageLoggingDestination','CommandWindow') sends MPI logging information to the MATLAB Command Window. If the task within a parallel job is set to capture Command Window output, the MPI logging information will be present in the tasks CommandWindowOutput property. mpiSettings('MessageLoggingDestination','stdout') sends MPI
logging information to the standard output for the MATLAB process. If you are using a job manager, this is the mdce service log file; if you are using an mpiexec scheduler, this is the mpiexec debug log, which you can read with getDebugLog.
mpiSettings('MessageLoggingDestination','File','filename')
Tips
Setting the MessageLoggingDestination does not automatically enable message logging. A separate call is required to enable message logging.
14-180
mpiSettings
mpiSettings has to be called on the lab, not the client. That is, it should be called within the task function, within jobStartup.m, or within taskStartup.m.
Examples
Set deadlock detection for a parallel job inside the jobStartup.m file for that job:
% Inside jobStartup.m for the parallel job mpiSettings('DeadlockDetection', 'on'); myLogFname = sprintf('%s_%d.log', tempname, labindex); mpiSettings('MessageLoggingDestination', 'File', myLogFname); mpiSettings('MessageLogging', 'on');
Turn off deadlock detection for all subsequent spmd statements that use the same MATLAB pool:
spmd; mpiSettings('DeadlockDetection', 'off'); end
14-181
numlabs
the current job. This value is the maximum value that can be used with
labSend and labReceive.
See Also
14-182
parallel.gpu.CUDAKernel
Purpose Syntax
Description
KERN = parallel.gpu.CUDAKernel(PTXFILE, CPROTO) and KERN = parallel.gpu.CUDAKernel(PTXFILE, CPROTO, FUNC) create a kernel object that you can use to call a CUDA kernel on the GPU. PTXFILE is
the name of the file that contains the PTX code, or the contents of a PTX file as a string; and CPROTO is the C prototype for the kernel call that KERN represents. If specified, FUNC must be a string that unambiguously defines the appropriate kernel entry name in the PTX file. If FUNC is omitted, the PTX file must contain only a single entry point.
KERN = parallel.gpu.CUDAKernel(PTXFILE, CUFILE) and KERN = parallel.gpu.CUDAKernel(PTXFILE, CUFILE, FUNC) read the CUDA source file CUFILE, and look for a kernel definition starting with '__global__' to find the function prototype for the CUDA kernel that is defined in PTXFILE.
Examples
and simpleEx.ptx contains the PTX resulting from compiling simpleEx.cu into PTX, both of the following statements return a kernel object that you can use to call the addToVector CUDA kernel.
kern = parallel.gpu.CUDAKernel('simpleEx.ptx', ...
14-183
parallel.gpu.CUDAKernel
See Also
14-184
parfor
The following table lists some ranges that are not valid. Invalid parfor Range
parfor i = 1:2:25 parfor i = -7.5:7.5 A = [3 7 -2 6 4 -4 9 3 7]; parfor i = find(A>0) parfor i = [5;6;7;8] [5;6;7;8] is a column vector, not a
row vector. You can enter a parfor-loop on multiple lines, but if you put more than one segment of the loop statement on the same line, separate the segments with commas or semicolons:
parfor i = range; <loop body>; end parfor (loopvar = initval:endval, M), statements, end uses M to specify the maximum number of MATLAB workers that will evaluate statements in the body of the parfor-loop. M must be a
nonnegative integer. By default, MATLAB uses as many workers as it finds available. If you specify an upper limit, MATLAB employs no
14-185
parfor
more than that number, even if additional workers are available. If you request more resources than are available, MATLAB uses the maximum number available at the time of the call. If the parfor-loop cannot run on workers in a MATLAB pool (for example, if no workers are available or M is 0), MATLAB executes the loop on the client in a serial manner. In this situation, the parfor semantics are preserved in that the loop iterations can execute in any order. Note Because of independence of iteration order, execution of parfor does not guarantee deterministic results. The maximum amount of data that can be transferred in a single chunk between client and workers in the execution of a parfor-loop is determined by the JVM memory allocation limit. For details, see Object Data Size Limitations on page 6-48. For a detailed description of parfor-loops, see Chapter 2, Parallel for-Loops (parfor).
Examples
Suppose that f is a time-consuming function to compute, and that you want to compute its value on each element of array A and place the corresponding results in array B:
parfor i = 1:length(A) B(i) = f(A(i)); end
Because the loop iteration occurs in parallel, this evaluation can complete much faster than it would in an analogous for-loop. Next assume that A, B, and C are variables and that f, g, and h are functions:
parfor i = 1:n t = f(A(i));
14-186
parfor
If the time to compute f, g, and h is large, parfor will be significantly faster than the corresponding for statement, even if n is relatively small. Although the form of this statement is similar to a for statement, the behavior can be significantly different. Notably, the assignments to the variables i, t, and u do not affect variables with the same name in the context of the parfor statement. The rationale is that the body of the parfor is executed in parallel for all values of i, and there is no deterministic way to say what the final values of these variables are. Thus, parfor is defined to leave these variables unaffected in the context of the parfor statement. By contrast, the variable C has a different element set for each value of i, and these assignments do affect the variable C in the context of the parfor statement. Another important use of parfor has the following form:
s = 0; parfor i = 1:n if p(i) % assume p is a function s = s + 1; end end
The key point of this example is that the conditional adding of 1 to s can be done in any order. After the parfor statement has finished executing, the value of s depends only on the number of iterations for which p(i) is true. As long as p(i) depends only upon i, the value of s is deterministic. This technique generalizes to functions other than plus (+). Note that the variable s does refer to the variable in the context of the parfor statement. The general rule is that the only variables in the context of a parfor statement that can be affected by it are those like s (combined by a suitable function like +) or those like C in the previous example (set by indexed assignment).
14-187
parfor
See Also
14-188
pause
jm
pause(jm) pauses the job managers queue so that jobs waiting in the queued state will not run. Jobs that are already running also pause, after completion of tasks that are already running. No further jobs or tasks will run until the resume function is called for the job manager.
The pause function does nothing if the job manager is already paused.
See Also
resume | waitForState
14-189
pctconfig
Purpose Syntax
Arguments
p1 v1 config
Property to configure. Supported properties are 'portrange', 'hostname'. Value for corresponding property. Structure of configuration value.
Description
pctconfig('p1', v1, ...) sets the client configuration property p1 with the value v1.
Note that the property value pairs can be in any format supported by the set function, i.e., param-value string pairs, structures, and param-value cell array pairs. If a structure is used, the structure field names are the property names and the field values specify the property values. If the property is 'portrange', the specified value is used to set the range of ports to be used by the client session of Parallel Computing Toolbox software. This is useful in environments with a limited choice of ports. The value of 'portrange' should either be a 2-element vector [minport, maxport] specifying the range, or 0 to specify that the client session should use ephemeral ports. By default, the client session searches for available ports to communicate with the other sessions of MATLAB Distributed Computing Server software. If the property is 'hostname', the specified value is used to set the hostname for the client session of Parallel Computing Toolbox software. This is useful when the client computer is known by more than one hostname. The value you should use is the hostname by which the cluster nodes can contact the client computer. The toolbox supports both short hostnames and fully qualified domain names.
14-190
pctconfig
The field names of the structure reflect the property names, while the field values are set to the property values.
config = pctconfig(), without any input arguments, returns all the current values as a structure to config. If you have not set any values, these are the defaults.
Tips
The values set by this function do not persist between MATLAB sessions. To guarantee its effect, call pctconfig before calling any other Parallel Computing Toolbox functions. View the current settings for hostname and ports.
config = pctconfig() config = portrange: [27370 27470] hostname: 'machine32'
Examples
Set the current client session port range to 21000-22000 with hostname fdm4.
pctconfig('hostname', 'fdm4', 'portrange', [21000 22000]);
14-191
pctRunDeployedCleanup
client JVM can properly terminate when the deployed application exits. All deployed applications that use Parallel Computing Toolbox functionality need to call pctRunDeployedCleanup after the last call to Parallel Computing Toolbox functionality. After calling pctRunDeployedCleanup, you should not use any further Parallel Computing Toolbox functionality in the current MATLAB session.
14-192
pctRunOnAll
Note If you use pctRunOnAll to run a command such as addpath in a mixed-platform environment, it can generate a warning on the client while executing properly on the labs. For example, if your labs are all running on Linux operating systems and your client is running on a Microsoft Windows operating system, an addpath argument with Linux-based paths will warn on the Windows-based client.
Examples
See Also
matlabpool
14-193
pload
fileroot
pload(fileroot) loads the data from the files named [fileroot num2str(labindex)] into the labs running a parallel job. The files should have been created by the psave command. The number of
labs should be the same as the number of files. The files should be accessible to all the labs. Any codistributed arrays are reconstructed by this function. If fileroot contains an extension, the character representation of the labindex will be inserted before the extension. Thus, pload('abc') attempts to load the file abc1.mat on lab 1, abc2.mat on lab 2, and so on.
Examples
Create three variables one replicated, one variant, and one codistributed. Then save the data.
clear all; rep = speye(numlabs); var = magic(labindex); D = eye(numlabs,codistributor()); psave('threeThings');
This creates three files (threeThings1.mat, threeThings2.mat, threeThings3.mat) in the current working directory. Clear the workspace on all the labs and confirm there are no variables.
clear all whos
14-194
pload
Load the previously saved data into the labs. Confirm its presence.
pload('threeThings'); whos isreplicated(rep) iscodistributed(D)
See Also
14-195
pmode
Purpose Syntax
Description
pmode allows the interactive parallel execution of MATLAB commands. pmode achieves this by defining and submitting a parallel job, and
opening a Parallel Command Window connected to the labs running the job. The labs then receive commands entered in the Parallel Command Window, process them, and send the command output back to the Parallel Command Window. Variables can be transferred between the MATLAB client and the labs.
pmode start starts pmode, using the default configuration to
define the scheduler and number of labs. (The initial default configuration is local; you can change it by using the function defaultParallelConfig.) You can also specify the number of labs using pmode start numlabs, but note that the local scheduler allows for only up to twelve labs.
pmode start conf numlabs starts pmode using the Parallel Computing Toolbox configuration conf to locate the scheduler, submits a parallel job with the number of labs identified by numlabs, and connects the
Parallel Command Window with the labs. If the number of labs is specified, it overrides the minimum and maximum number of workers specified in the configuration.
pmode quit or pmode exit stops the parallel job, destroys it, and closes the Parallel Command Window. You can enter this command at the MATLAB prompt or the pmode prompt. pmode client2lab clientvar labs labvar copies the variable clientvar from the MATLAB client to the variable labvar on the labs
14-196
pmode
identified by labs. If labvar is omitted, the copy is named clientvar. labs can be either a single lab index or a vector of lab indices. You can enter this command at the MATLAB prompt or the pmode prompt.
pmode lab2client labvar lab clientvar copies the variable labvar from the lab identified by lab, to the variable clientvar on the MATLAB client. If clientvar is omitted, the copy is named labvar. You can enter this command at the MATLAB prompt or the pmode
prompt. Note: If you use this command in an attempt to transfer a codistributed array to the client, you get a warning, and only the local portion of the array on the specified lab is transferred. To transfer an entire codistributed array, first use the gather function to assemble the whole array into the labs workspaces.
pmode cleanup conf destroys all parallel jobs created by pmode for the current user running under the scheduler specified in the configuration conf, including jobs that are currently running. The configuration is optional; the default configuration is used if none is specified. You can enter this command at the MATLAB prompt or the pmode prompt.
You can invoke pmode as either a command or a function, so the following are equivalent.
pmode start conf 4 pmode('start', 'conf', 4)
Examples
In the following examples, the pmode prompt (P>>) indicates commands entered in the Parallel Command Window. Other commands are entered in the MATLAB Command Window. Start pmode using the default configuration to identify the scheduler and number of labs.
pmode start
Start pmode using the local configuration with four local labs.
pmode start local 4
14-197
pmode
Start pmode using the configuration myconfig and eight labs on the cluster.
pmode start myconfig 8
See Also
14-198
poolStartup
File for user-defined options to run on each worker when MATLAB pool starts
poolStartup poolStartup runs automatically on a worker each time the worker
forms part of a MATLAB pool. You do not call this function from the client session, nor explicitly as part of a task function. You add MATLAB code to the poolStartup.m file to define pool initialization on the worker. The worker looks for poolStartup.m in the following order, executing the one it finds first:
1 Included in the jobs FileDependencies property. 2 In a folder included in the jobs PathDependencies property. 3 In the workers MATLAB installation at the location
matlabroot/toolbox/distcomp/user/poolStartup.m
To create a version of poolStartup.m for FileDependencies or PathDependencies, copy the provided file and modify it as required. .
poolStartup is the ideal location for startup code required for parallel execution on the MATLAB pool. For example, you might want to include code for using mpiSettings. Because jobStartup and taskStartup execute before poolStartup, they are not suited to pool-specific code. In other words, you should use taskStartup for setup code on your worker regardless of whether the task is from a distributed job, parallel job, or using a MATLAB pool; while poolStartup is for setup code for pool usage only.
For further details on poolStartup and its implementation, see the text in the installed poolStartup.m file.
See Also
14-199
promote
jm job
The job manager object that contains the job. Job object promoted in the queue.
Description
promote(jm, job) promotes the job object job, that is queued in the job manager jm.
If job is not the first job in the queue, promote exchanges the position of job and the previous job.
Tips
After a call to promote or demote, there is no change in the order of job objects contained in the Jobs property of the job manager object. To see the scheduled order of execution for jobs in the queue, use the findJob function in the form [pending queued running finished] = findJob(jm). Create and submit multiple jobs to the scheduler identified by the default parallel configuration:
j1 = createJob('name','Job A'); j2 = createJob('name','Job B'); j3 = createJob('name','Job C'); submit(j1);submit(j2);submit(j3);
Examples
Assuming that the default parallel configuration uses a job manager, create an object for that job manager, and promote Job C by one position in its queue:
jm = findResource(); promote(jm, j3)
14-200
promote
[pjobs, qjobs, rjobs, fjobs] = findJob(jm); get(qjobs, 'Name') 'Job A' 'Job C' 'Job B'
See Also
14-201
psave
fileroot
psave(fileroot) saves the data from the labs workspace into the files named [fileroot num2str(labindex)]. The files can be loaded by using the pload command with the same fileroot, which should point to a directory accessible to all the labs. If fileroot contains an extension, the character representation of the labindex is inserted before the extension. Thus, psave('abc') creates the files 'abc1.mat', 'abc2.mat', etc., one for each lab.
Examples
Create three variables one replicated, one variant, and one codistributed. Then save the data.
clear all; rep = speye(numlabs); var = magic(labindex); D = eye(numlabs,codistributor()); psave('threeThings');
This creates three files (threeThings1.mat, threeThings2.mat, threeThings3.mat) in the current working directory. Clear the workspace on all the labs and confirm there are no variables.
clear all whos
14-202
psave
Load the previously saved data into the labs. Confirm its presence.
pload('threeThings'); whos isreplicated(rep) iscodistributed(D)
See Also
14-203
redistribute
Examples
See Also
14-204
resume
jm
resume(jm) resumes processing of the job managers queue so that jobs waiting in the queued state will be run. This call will do nothing if the job manager is not paused. pause | waitForState
See Also
14-205
set
Purpose Syntax
Arguments
An object or an array of objects. A property name for obj. A property value supported by PropertyName. A cell array of property names. A cell array of property values. A structure array whose field names are the property names for obj. A structure with property names and property values. Literal string to indicate usage of a configuration. Name of the configuration to use.
Description
set(obj) displays all configurable properties for obj. If a property has a finite list of possible string values, these values are also displayed. props = set(obj) returns all configurable properties for obj and their possible values to the structure props. The field names of props are the property names of obj, and the field values are cell arrays of possible
14-206
set
property values. If a property does not have a finite set of possible values, its cell array is empty.
set(obj,'PropertyName') displays the valid values for PropertyName
possible values.
set(obj,'PropertyName',PropertyValue,...) configures one or
Tips
You can use any combination of property name/property value pairs, structure arrays, and cell arrays in one call to set. Additionally, you can specify a property name without regard to case, and you can make use of property name completion. For example, if j1 is a job object, the following commands are all valid and have the same result:
set(j1,'Timeout',20) set(j1,'timeout',20) set(j1,'timeo',20)
14-207
set
Examples
This example illustrates some of the ways you can use set to configure property values for the job object j1.
set(j1,'Name','Job_PT109','Timeout',60); props1 = {'Name' 'Timeout'}; values1 = {'Job_PT109' 60}; set(j1, props1, values1); S.Name = 'Job_PT109'; S.Timeout = 60; set(j1,S);
See Also
get | inspect
14-208
setJobSchedulerData
Scheduler object identifying the generic third-party scheduler running the job. Job object identifying the job for which to store data. Information to store for this job.
Description
setJobSchedulerData(sched, job, userdata) stores data for the job job that is running under the generic scheduler sched. You can later retrieve the information with the function getJobSchedulerData. For
example, it might be useful to store the third-party schedulers external ID for this job, so that the function specified in GetJobStateFcn can later query the scheduler about the state of the job. Or the stored data might be an array with the schedulers ID for each task in the job. You should call the function setJobSchedulerData in the submit function (identified by the SubmitFcn property) and call getJobSchedulerData in any of the functions identified by the properties GetJobStateFcn, DestroyJobFcn, DestroyTaskFcn, CancelJobFcn, or CancelTaskFcn. For more information and examples on using these functions and properties, see Manage Jobs with Generic Scheduler on page 8-50.
See Also
getJobSchedulerData
14-209
setupForParallelExecution
Purpose Syntax
Arguments
Platform LSF, PBS Pro, or TORQUE scheduler object. Setting for parallel execution.
Description
to expect workers running on Microsoft Windows operating systems, and selects the wrapper script which expects to be able to call "mpiexec -delegate" on the workers. Note that you still need to supply SubmitArguments that ensure that the LSF or PBS Pro scheduler runs your job only on PC-based workers. For example, for LSF, including '-R type==NTX86' in your SubmitArguments causes the scheduler to select only workers on 32-bit Windows operating systems.
setupForParallelExecution(sched, 'pcNoDelegate') is similar to the 'pc' mode, except that the wrapper script does not attempt to call "mpiexec -delegate", and so assumes that you have installed some
14-210
setupForParallelExecution
Examples
From any client, set up the scheduler to run parallel jobs only on Windows-based (PC) workers.
lsf_sched = findResource('scheduler', 'Type', 'lsf'); setupForParallelExecution(lsf_sched, 'pc'); set(lsf_sched, 'SubmitArguments', '-R type==NTX86');
From any client, set up the scheduler to run parallel jobs only on UNIX-based workers.
lsf_sched = findResource('scheduler', 'Type', 'lsf'); setupForParallelExecution(lsf_sched, 'unix'); set(lsf_sched, 'SubmitArguments', '-R type==LINUX64');
See Also
createParallelJob | findResource
14-211
size
Purpose Syntax
Arguments
An object or an array of objects. The dimension of obj. The number of rows and columns in obj. The number of rows in obj, or the length of the dimension specified by dim. The number of columns in obj. The lengths of the first n dimensions of obj.
Description
d = size(obj) returns the two-element row vector d containing the number of rows and columns in obj. [m,n] = size(obj) returns the number of rows and columns in
See Also
length
14-212
sparse
Purpose Syntax
Description
sparse codistributed array in the manner specified above, but does not perform any global communication for error checking when constructing the array. This form of the syntax is most useful inside spmd, pmode, or a parallel job. Note To create a sparse codistributed array of underlying class logical, first create an array of underlying class double and then cast it using the logical function:
spmd SC = logical(sparse(m, n, codistributor1d())); end
Examples
14-213
sparse
creates a 1000-by-1000 codistributed sparse double array C. C is distributed by its second dimension (columns), and each lab contains a 1000-by-250 local piece of C.
spmd(4) codist = codistributor1d(2, 1:numlabs) C = sparse(10, 10, codist); end
creates a 10-by-10 codistributed sparse double array C, distributed by its columns. Each lab contains a 10-by-labindex local piece of C. Convert a distributed array into a sparse distributed array:
R = distributed.rand(1000); D = floor(2*R); % D also is distributed SD = sparse(D); % SD is sparse distributed
See Also
14-214
spmd
Purpose Syntax
Description
The general form of an spmd (single program, multiple data) statement is:
spmd statements end spmd, statements, end defines an spmd statement on a single line. MATLAB executes the spmd body denoted by statements on several MATLAB workers simultaneously. The spmd statement can be used
only if you have Parallel Computing Toolbox. To execute the statements in parallel, you must first open a pool of MATLAB workers using matlabpool. Inside the body of the spmd statement, each MATLAB worker has a unique value of labindex, while numlabs denotes the total number of workers executing the block in parallel. Within the body of the spmd statement, communication functions for parallel jobs (such as labSend and labReceive) can transfer data between the workers. Values returning from the body of an spmd statement are converted to Composite objects on the MATLAB client. A Composite object contains references to the values stored on the remote MATLAB workers, and those values can be retrieved using cell-array indexing. The actual data on the workers remains available on the workers for subsequent spmd execution, so long as the Composite exists on the client and the MATLAB pool remains open. By default, MATLAB uses as many workers as it finds available in the pool. When there are no MATLAB workers available, MATLAB executes the block body locally and creates Composite objects as necessary.
14-215
spmd
spmd(n), statements, end uses n to specify the exact number of MATLAB workers to evaluate statements, provided that n workers
are available from the MATLAB pool. If there are not enough workers available, an error is thrown. If n is zero, MATLAB executes the block body locally and creates Composite objects, the same as if there is no pool available.
spmd(m, n), statements, end uses a minimum of m and a maximum of n workers to evaluate statements. If there are not enough workers available, an error is thrown. m can be zero, which allows the block to run locally if no workers are available.
For more information about spmd and Composite objects, see Chapter 3, Single Program Multiple Data (spmd).
Tips Examples
For information about restrictions and limitations when using spmd, see Limitations on page 3-15. Perform a simple calculation in parallel, and plot the results:
matlabpool(3) spmd % build magic squares in parallel q = magic(labindex + 2); end for ii=1:length(q) % plot each magic square figure, imagesc(q{ii}); end matlabpool close
See Also
14-216
submit
obj
submit(obj) queues the job object, obj, in the scheduler queue. The scheduler used for this job was determined when the job was created.
When a job contained in a scheduler is submitted, the jobs State property is set to queued, and the job is added to the list of jobs waiting to be executed. The jobs in the waiting list are executed in a first in, first out manner; that is, the order in which they were submitted, except when the sequence is altered by promote, demote, cancel, or destroy.
Examples
Find the job manager named jobmanager1 using the lookup service on host JobMgrHost.
jm1 = findResource('scheduler','type','jobmanager', ... 'name','jobmanager1','LookupURL','JobMgrHost');
See Also
createJob | findJob
14-217
subsasgn
Purpose Syntax
Description
See Also
14-218
subsref
Purpose Syntax
Description
See Also
14-219
taskFinish
task
worker finishes evaluating a task for a particular job. You do not call this function from the client session, nor explicitly as part of a task function. You add MATLAB code to the taskFinish.m file to define anything you want executed on the worker when a task is finished. The worker looks for taskFinish.m in the following order, executing the one it finds first:
1 Included in the jobs FileDependencies property. 2 In a folder included in the jobs PathDependencies property. 3 In the workers MATLAB installation at the location
matlabroot/toolbox/distcomp/user/taskFinish.m
To create a version of taskFinish.m for FileDependencies or PathDependencies, copy the provided file and modify it as required. For further details on taskFinish and its implementation, see the text in the installed taskFinish.m file.
See Also
14-220
taskStartup
task
worker evaluates a task for a particular job. You do not call this function from the client session, nor explicitly as part of a task function. You add MATLAB code to the taskStartup.m file to define task initialization on the worker. The worker looks for taskStartup.m in the following order, executing the one it finds first:
1 Included in the jobs FileDependencies property. 2 In a folder included in the jobs PathDependencies property. 3 In the workers MATLAB installation at the location
matlabroot/toolbox/distcomp/user/taskStartup.m
To create a version of taskStartup.m for FileDependencies or PathDependencies, copy the provided file and modify it as required. For further details on taskStartup and its implementation, see the text in the installed taskStartup.m file.
See Also
14-221
wait
Purpose Syntax
Arguments
Job object whose change in state to wait for. Value of the job objects State property to wait for. Maximum time to wait, in seconds.
Description
wait(obj) blocks execution in the client session until the job identified by the object obj reaches the 'finished' state or fails. This occurs
when all the jobs tasks are finished processing on remote workers.
wait(obj, 'state') blocks execution in the client session until the specified job object changes state to the value of 'state'. The valid states to wait for are 'queued', 'running', and 'finished'.
If the object is currently or has already been in the specified state, a wait is not performed and execution returns immediately. For example, if you execute wait(job, 'queued') for a job already in the 'finished' state, the call returns immediately.
wait(obj, 'state', timeout) blocks execution until either the job reaches the specified 'state', or timeout seconds elapse, whichever happens first.
Note Simulink models cannot run while a MATLAB session is blocked by wait. If you must run Simulink from the MATLAB client while also running distributed or parallel jobs, you cannot use wait.
Examples
Submit a job to the queue, and wait for it to finish running before retrieving its results.
14-222
wait
Submit a batch job and wait for it to finish before retrieving its variables.
job = batch('myScript'); wait(job) load(job)
See Also
14-223
waitForState
Purpose Syntax
Arguments
Job or task object whose change in state to wait for. Value of the objects State property to wait for. Maximum time to wait, in seconds. Boolean true if wait succeeds, false if times out.
Description
waitForState(obj) blocks execution in the client session until the job or task identified by the object obj reaches the 'finished' state
or fails. For a job object, this occurs when all its tasks are finished processing on remote workers.
waitForState(obj, 'state') blocks execution in the client session until the specified object changes state to the value of 'state'. For a job object, the valid states to wait for are 'queued', 'running', and 'finished'. For a task object, the valid states are 'running' and 'finished'.
If the object is currently or has already been in the specified state, a wait is not performed and execution returns immediately. For example, if you execute waitForState(job, 'queued') for a job already in the 'finished' state, the call returns immediately.
waitForState(obj, 'state', timeout) blocks execution until either the object reaches the specified 'state', or timeout seconds elapse,
14-224
waitForState
Note Simulink models cannot run while a MATLAB session is blocked by waitForState. If you must run Simulink from the MATLAB client while also running distributed or parallel jobs, you cannot use waitForState.
Examples
Submit a job to the queue, and wait for it to finish running before retrieving its results.
submit(job) waitForState(job, 'finished') results = getAllOutputArguments(job)
See Also
14-225
waitForState
14-226
15
Property Reference
Job Manager (p. 15-2) Schedulers (p. 15-3) Jobs (p. 15-5) Tasks (p. 15-7) Workers (p. 15-8) Control job manager objects Control scheduler objects Control job objects Control task objects Control worker objects
15
Property Reference
Job Manager
BusyWorkers ClusterOsType ClusterSize Configuration HostAddress Hostname IdleWorkers IsUsingSecureCommunication Jobs
Workers currently running tasks Specify operating system of nodes on which scheduler will start workers Number of workers available to scheduler Specify configuration to apply to object or toolbox function IP address of host running job manager or worker session Name of host running job manager or worker session Idle workers available to run tasks True if job manager and workers use secure communication Jobs contained in job manager service or in schedulers data location Name of job manager, job, or worker object Number of workers currently running tasks Number of idle workers available to run tasks Specify if system should prompt for password when authenticating user Security level controlling access to job manager and its jobs Current state of task, job, job manager, or worker Type of scheduler object
15-2
Schedulers
UserData UserName
Specify data to associate with object User who created job or job manager object
Schedulers
CancelJobFcn CancelTaskFcn ClusterMatlabRoot ClusterName ClusterOsType ClusterSize ClusterVersion Configuration DataLocation DestroyJobFcn DestroyTaskFcn EnvironmentSetMethod
Specify function to run when canceling job on generic scheduler Specify function to run when canceling task on generic scheduler Specify MATLAB root for cluster Name of Platform LSF cluster Specify operating system of nodes on which scheduler will start workers Number of workers available to scheduler Version of HPC Server scheduler Specify configuration to apply to object or toolbox function Specify directory where job data is stored Specify function to run when destroying job on generic scheduler Specify function to run when destroying task on generic scheduler Specify means of setting environment variables for mpiexec scheduler Specify function to run when querying job state on generic scheduler
GetJobStateFcn
15-3
15
Property Reference
HasSharedFilesystem JobDescriptionFile
Specify whether nodes share data location Name of XML job description file for Microsoft Windows HPC Server scheduler Jobs contained in job manager service or in schedulers data location Name of job template for HPC Server 2008 scheduler Name of Platform LSF master node MATLAB command that generic scheduler runs to start lab Specify pathname of executable mpiexec command labs
Jobs
ParallelSubmissionWrapperScript Script that scheduler runs to start ParallelSubmitFcn RcpCommand ResourceTemplate RshCommand SchedulerHostname ServerName SubmitArguments
Specify function to run when parallel job submitted to generic scheduler Command to copy files from client Resource definition for PBS Pro or TORQUE scheduler Remote execution command used on worker nodes during parallel job Name of host running Microsoft Windows HPC Server scheduler Name of current PBS Pro or TORQUE server machine Specify additional arguments to use when submitting job to Platform LSF, PBS Pro, TORQUE, or mpiexec scheduler
15-4
Jobs
Specify function to run when job submitted to generic scheduler Type of scheduler object Specify data to associate with object Allow service-oriented architecture (SOA) submission on HPC Server 2008 cluster Specify operating system of nodes on which mpiexec scheduler will start labs
WorkerMachineOsType
Jobs
AuthorizedUsers Configuration CreateTime FileDependencies FinishedFcn FinishTime ID JobData MaximumNumberOfWorkers MinimumNumberOfWorkers
Specify users authorized to access job Specify configuration to apply to object or toolbox function When task or job was created Directories and files that worker can access Specify callback to execute after task or job runs When task or job finished Object identifier Data made available to all workers for jobs tasks Specify maximum number of workers to perform job tasks Specify minimum number of workers to perform job tasks
15-5
15
Property Reference
Name of job manager, job, or worker object Parent object of job or task Specify directories to add to MATLAB worker path Specify function file to execute when job is submitted to job manager queue Specify whether to restart MATLAB workers before evaluating job tasks Specify function file to execute when job or task starts running When job or task started Current state of task, job, job manager, or worker Specify additional arguments to use when submitting job to Platform LSF, PBS Pro, TORQUE, or mpiexec scheduler When job was submitted to queue Specify label to associate with job object First task contained in MATLAB pool job object Tasks contained in job object Specify time limit to complete task or job Specify data to associate with object User who created job or job manager object
15-6
Tasks
Tasks
AttemptedNumberOfRetries CaptureCommandWindowOutput CommandWindowOutput Configuration CreateTime Error ErrorIdentifier ErrorMessage FailedAttemptInformation FinishedFcn FinishTime Function ID InputArguments MaximumNumberOfRetries NumberOfOutputArguments OutputArguments Parent RunningFcn
Number of times failed task was rerun Specify whether to return Command Window output Text produced by execution of task objects function Specify configuration to apply to object or toolbox function When task or job was created Task error information Task error identifier Message from task error Information returned from failed task Specify callback to execute after task or job runs When task or job finished Function called when evaluating task Object identifier Input arguments to task object Specify maximum number of times to rerun failed task Number of arguments returned by task function Data returned from execution of task Parent object of job or task Specify function file to execute when job or task starts running
15-7
15
Property Reference
When job or task started Current state of task, job, job manager, or worker Specify time limit to complete task or job Specify data to associate with object Worker session that performed task
Workers
Computer CurrentJob CurrentTask HostAddress Hostname JobManager Name PreviousJob PreviousTask State
Information about computer on which worker is running Job whose task this worker session is currently evaluating Task that worker is currently running IP address of host running job manager or worker session Name of host running job manager or worker session Job manager that this worker is registered with Name of job manager, job, or worker object Job whose task this worker previously ran Task that this worker previously ran Current state of task, job, job manager, or worker
15-8
16
Properties Alphabetical List
AttemptedNumberOfRetries
Purpose Description
Number of times failed task was rerun If a task reruns because of certain system failures, the task property AttemptedNumberOfRetries stores a count of the number of attempted reruns. Note The AttemptedNumberOfRetries property is available only when using the MathWorks job manager as your scheduler.
Characteristics Usage
Read-only Data type
See Also
Properties
FailedAttemptInformation, MaximumNumberOfRetries
16-2
AuthorizedUsers
Purpose Description
Specify users authorized to access job The AuthorizedUsers property value is a cell array of strings which identify the users who are allowed to access the job. This controls who can set properties on the job, add tasks, destroy the job, etc. The person identified as the owner by the jobs UserName property does not have to be listed in the AuthorizedUsers property value. The following table explains the effect of AuthorizedUsers at different security levels. Security Level 0 1 Effect of AuthorizedUsers No effect. All users can access the job without hindrance. For users included in the property value, the system suppresses the dialog box that requires acknowledgment that the job belongs to another user. All other users must acknowledge job ownership every time they access the job. Only users who are authenticated in this session and are listed in AuthorizedUsers can access the job.
2 and 3
Note The AuthorizedUsers property is available only when using the MathWorks job manager as your scheduler.
Characteristics Usage
Read-only Data type
16-3
AuthorizedUsers
Values
You can populate AuthorizedUsers with the names of any users. At security levels 13, the users must be recognized by the job manager as authenticated in the session in which you are setting the property. This example creates a job named Job33, then adds the users sammy and bob to the jobs AuthorizedUsers.
jm = findResource('scheduler', 'Configuration', defaultparallelconfig); j = createJob(jm, 'Name', 'Job33'); set(j, 'AuthorizedUsers', {'sammy', 'bob'})
Examples
See Also
Properties
SecurityLevel, UserName
16-4
BlockSize
Purpose Description
For more information on 2dbc distribution and the block size of distributed arrays, see 2-Dimensional Distribution on page 5-17.
Characteristics Usage
Read-only Data type
See Also
Functions
codistributor2dbc
Properties
LabGrid, Orientation
16-5
BusyWorkers
Purpose Description
Workers currently running tasks The BusyWorkers property value indicates which workers are currently running tasks for the job manager.
Characteristics Usage
Read-only Data type
Values
As workers complete tasks and assume new ones, the lists of workers in BusyWorkers and IdleWorkers can change rapidly. If you examine these two properties at different times, you might see the same worker on both lists if that worker has changed its status between those times. If a worker stops unexpectedly, the job managers knowledge of that as a busy or idle worker does not get updated until the job manager runs the next job and tries to send a task to that worker.
Examples
Examine the workers currently running tasks for a particular job manager.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); workers_running_tasks = get(jm, 'BusyWorkers')
See Also
Properties
ClusterSize, IdleWorkers, MaximumNumberOfWorkers, MinimumNumberOfWorkers, NumberOfBusyWorkers, NumberOfIdleWorkers
16-6
CancelJobFcn
Purpose Description
For more information and examples on using these functions and properties, see Manage Jobs with Generic Scheduler on page 8-50.
Characteristics Usage
Read-only Data type
Values
You can set CancelJobFcn to any valid function handle or a cell array whose first element is a function handle.
See Also
Functions
cancel, getJobSchedulerData, setJobSchedulerData
Properties
CancelTaskFcn, DestroyJobFcn, DestroyTaskFcn
16-7
CancelTaskFcn
Purpose Description
for a task running on a generic scheduler. This function lets you communicate with the scheduler, to provide any instructions beyond the normal toolbox action of changing the state of the task. To identify the task for the scheduler, the function should include a call to getJobSchedulerData. For more information and examples on using these functions and properties, see Manage Jobs with Generic Scheduler on page 8-50.
Characteristics Usage
Read-only Data type
Values
You can set CancelTaskFcn to any valid function handle or a cell array whose first element is a function handle.
See Also
Functions
cancel, getJobSchedulerData, setJobSchedulerData
Properties
CancelJobFcn, DestroyJobFcn, DestroyTaskFcn
16-8
CaptureCommandWindowOutput
Purpose Description
If CaptureCommandWindowOutput is set true (or logical 1), the command window output will be stored in the CommandWindowOutput property of the task object. If the value is set false (or logical 0), the task does not retain command window output.
Characteristics Usage
Read-only Data type
Values
The value of CaptureCommandWindowOutput can be set to true (or logical 1) or false (or logical 0). When you perform get on the property, the value returned is logical 1 or logical 0. The default value is logical 0 to save network bandwidth in situations where the output is not needed; except for batch jobs, whose default is 1 (true). Set all tasks in a job to retain any command window output generated during task evaluation.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); createTask(j, @myfun, 1, {x}); createTask(j, @myfun, 1, {x}); . . . alltasks = get(j, 'Tasks'); set(alltasks, 'CaptureCommandWindowOutput', true)
Examples
16-9
CaptureCommandWindowOutput
See Also
Properties
Function, CommandWindowOutput
16-10
ClusterMatlabRoot
Purpose Description
to use for starting MATLAB worker processes. The path must be available from all nodes on which worker sessions will run. When using the generic scheduler interface, your scheduler script can construct a path to the executable by concatenating the values of ClusterMatlabRoot and MatlabCommandToRun into a single string.
Characteristics Usage
Read-only Data type
Values
the file system of the cluster nodes. The directory must be accessible as expressed in this string, from all cluster nodes on which MATLAB workers will run. If the value is empty, the MATLAB executable must be on the path of the worker.
See Also
Properties
DataLocation, MasterName, MatlabCommandToRun, PathDependencies
16-11
ClusterName
Purpose Description
Characteristics Usage
Read-only Data type
See Also
Properties
DataLocation, MasterName, PathDependencies
16-12
ClusterOsType
Purpose Description
scheduler will start workers, or whose workers are already registered with a job manager.
Characteristics Usage
Read-only Data type
Scheduler object For job manager or Microsoft Windows HPC Server (or CCS) scheduler object String
Values
The valid values for this property are 'pc', 'unix', and'mixed'. For Windows HPC Server, the setting is always 'pc'. A value of 'mixed' is valid only for distributed jobs with Platform LSF or generic schedulers; or for distributed or parallel jobs with a job manager. Otherwise, the nodes of the labs running a parallel job with LSF, Windows HPC Server, PBS Pro, TORQUE, mpiexec, or generic scheduler must all be the same platform. For parallel jobs with an LSF, PBS Pro, or TORQUE scheduler, this property value is set when you execute the function setupForParallelExecution, so you do not need to set the value directly.
See Also
Functions
createParallelJob, findResource, setupForParallelExecution
Properties
ClusterName, MasterName, SchedulerHostname
16-13
ClusterSize
Purpose Description
Characteristics Usage
Read-only Data type
Values
For job managers this property is read-only. The value for a job manager represents the number of workers registered with that job manager. For local or third-party schedulers, this property is settable, and its value specifies the maximum number of workers or labs that this scheduler can start for running a job. A parallel jobs MaximumNumberOfWorkers property value must not exceed the value of ClusterSize.
Remarks
If you change the value of ClusterVersion or SchedulerHostname, this resets the values of ClusterSize, JobTemplate, and UseSOAJobSubmission.
See Also
Properties
BusyWorkers, IdleWorkers, MaximumNumberOfWorkers, MinimumNumberOfWorkers, NumberOfBusyWorkers, NumberOfIdleWorkers
16-14
ClusterVersion
Purpose Description
Characteristics Usage
Read-only Data type
This property can have the value 'CCS' (for CCS) or 'HPCServer2008' (for HPC Server 2008). If you change the value of ClusterVersion, this resets the values of ClusterSize, JobTemplate, and UseSOAJobSubmission.
Properties
JobDescriptionFile, JobTemplate, UseSOAJobSubmission
16-15
codistributor2dbc.defaultBlockSize
Purpose Description
Characteristics Usage
Read-only Data type
See Also
Functions
codistributor2dbc, codistributor2dbc.defaultLabGrid
Properties
BlockSize, LabGrid
16-16
CommandWindowOutput
Purpose Description
For example, if the function specified in the Function property makes calls to the disp command, the output that would normally be printed to the Command Window on the worker is captured in the CommandWindowOutput property. Whether to store the CommandWindowOutput is specified using the CaptureCommandWindowOutput property. The CaptureCommandWindowOutput property by default is logical 0 to save network bandwidth in situations when the CommandWindowOutput is not needed.
Characteristics Usage
Read-only Data type
Values Examples
Before a task is evaluated, the default value of CommandWindowOutput is an empty string. Get the Command Window output from all tasks in a job.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); createTask(j, @myfun, 1, {x}); createTask(j, @myfun, 1, {x}); . . alltasks = get(j, 'Tasks') set(alltasks, 'CaptureCommandWindowOutput', true)
16-17
CommandWindowOutput
See Also
Properties
Function, CaptureCommandWindowOutput
16-18
Computer
Purpose Description
Information about computer on which worker is running The Computer property of a worker is set to the string that would be returned from running the computer function on that worker.
Characteristics Usage
Read-only Data type
Values
Some possible values for the Computer property are GLNX86, GLNXA64, MACI, PCWIN, and PCWIN64. For more information about specific values, see the computer function reference page.
See Also
Functions
computer MATLAB function reference page
Properties
HostAddress, Hostname, WorkerMachineOsType
16-19
Configuration
Purpose Description
Specify configuration to apply to object or toolbox function You use the Configuration property to apply a configuration to an object. For details about writing and applying configurations, see Parallel Configurations for Cluster Access on page 6-14. Setting the Configuration property causes all the applicable properties defined in the configuration to be set on the object.
Characteristics Usage
Read-only Data type
Values
The value of Configuration is a string that matches the name of a configuration. If a configuration was never applied to the object, or if any of the settable object properties have been changed since a configuration was applied, the Configuration property is set to an empty string. Use a configuration to find a scheduler.
jm = findResource('scheduler','configuration','myConfig')
Examples
16-20
Configuration
See Also
Functions
createJob, createParallelJob, createTask, dfeval, dfevalasync, findResource
16-21
CreateTime
Purpose Description
Characteristics Usage
Read-only Data type
Values Examples
See Also
Functions
createJob, createTask
Properties
FinishTime, StartTime, SubmitTime
16-22
CurrentJob
Purpose Description
Characteristics Usage
Read-only Data type
a task.
Properties
CurrentTask, PreviousJob, PreviousTask, Worker
16-23
CurrentTask
Purpose Description
present time.
Characteristics Usage
Read-only Data type
a task.
Properties
CurrentJob, PreviousJob, PreviousTask, Worker
16-24
DataLocation
Purpose Description
Characteristics Usage
Read-only Data type
Values
data. In a shared file system, the client, scheduler, and all worker nodes must have access to this location. In a nonshared file system, only the MATLAB client and scheduler access job data in this location. If DataLocation is not set, the default location for job data is the current working directory of the MATLAB client the first time you use findResource to create an object for this type of scheduler. All settable property values on a scheduler object are local to the MATLAB client, and are lost when you close the client session or when you remove the object from the client workspace with delete or clear all. Use a structure to specify the DataLocation in an environment of mixed platforms. The fields for the structure are named pc and unix. Each node then uses the field appropriate for its platform. See the examples below. When you examine a DataLocation property that was set by a structure in this way, the value returned is the string appropriate for the platform on which you are examining it. The job data stored in the folder identified by DataLocation might not be compatible between different versions of MATLAB Distributed Computing Server. Therefore, DataLocation should not be shared by parallel computing products running different versions, and each version on your cluster should have its own DataLocation.
Examples
16-25
DataLocation
set(sch, 'DataLocation','/depot/jobdata')
Use a structure to set the DataLocation property for a mixed platform cluster.
d = struct('pc', '\\ourdomain\depot\jobdata', ... 'unix', '/depot/jobdata') set(sch, 'DataLocation', d)
See Also
Properties
HasSharedFilesystem, PathDependencies
16-26
DestroyJobFcn
Purpose Description
for a job running on a generic scheduler. This function lets you communicate with the scheduler, to provide any instructions beyond the normal toolbox action of deleting the job data from disk. To identify the job for the scheduler, the function should include a call to getJobSchedulerData. For more information and examples on using these functions and properties, see Manage Jobs with Generic Scheduler on page 8-50.
Characteristics Usage
Read-only Data type
Values
You can set DestroyJobFcn to any valid function handle, string of a function name, or a cell array whose first element is one of these.
See Also
Functions
destroy, getJobSchedulerData, setJobSchedulerData
Properties
CancelJobFcn, CancelTaskFcn, DestroyTaskFcn
16-27
DestroyTaskFcn
Purpose Description
for a task running on a generic scheduler. This function lets you communicate with the scheduler, to provide any instructions beyond the normal toolbox action of deleting the task data from disk. To identify the task for the scheduler, the function should include a call to getJobSchedulerData. For more information and examples on using these functions and properties, see Manage Jobs with Generic Scheduler on page 8-50.
Characteristics Usage
Read-only Data type
Values
You can set DestroyTaskFcn to any valid function handle or a cell array whose first element is a function handle.
See Also
Functions
destroy, getJobSchedulerData, setJobSchedulerData
Properties
CancelJobFcn, CancelTaskFcn, DestroyJobFcn
16-28
Dimension
Purpose Description
Characteristics Usage
Read-only Data type
See Also
Functions
codistributor1d
Properties
Partition
16-29
EnvironmentSetMethod
Purpose Description
Specify means of setting environment variables for mpiexec scheduler The mpiexec scheduler needs to supply environment variables to the MATLAB processes (labs) that it launches. There are two means by which it can do this, determined by the EnvironmentSetMethod property.
Characteristics Usage
Read-only Data type
Values
A value of '-env' instructs the mpiexec scheduler to insert into the mpiexec command line additional directives of the form -env VARNAME value. A value of 'setenv' instructs the mpiexec scheduler to set the environment variables in the environment that launches mpiexec.
16-30
Error
Purpose Description
Task error information If an error occurs during the task evaluation, Error contains the MException object thrown. See the MException reference page for more about the returned information.
Characteristics Usage
Read-only Data type
Values
The Error property is empty before an attempt to run a task. When using a job manager, the Error property indicates if a worker has crashed during task execution, but when using a different scheduler Error does not indicate a crash. The Errors message and identifier contain the same information as the task object properties ErrorMessage and ErrorIdentifier, respectively.
Examples
Examine a tasks Error property to determine if there was an error during execution:
jm = findResource('scheduler','type','local'); j = createJob(jm); t = createTask(j, @myfun, 1, argcell); submit(j) waitForState(j) has_error = ~isempty(t.Error)
16-31
Error
See Also
Properties
ErrorIdentifier, ErrorMessage, Function
16-32
ErrorIdentifier
Purpose Description
Task error identifier If an error occurs during the task evaluation, ErrorIdentifier contains the identifier property of the MException thrown.
Characteristics Usage
Read-only Data type
Values
remains empty if the evaluation of a task function does not produce an error or if the error did not provide an identifier. ErrorIdentifier has the same value as the identifier property of the tasks Error. When using a job manager, ErrorIdentifier indicates if a worker has crashed during task execution, but when using a different scheduler ErrorIdentifier does not indicate a crash.
Examples
See Also
Properties
Error, ErrorMessage, Function
16-33
ErrorMessage
Purpose Description
Message from task error If an error occurs during the task evaluation, ErrorMessage contains the message property of the MException thrown.
Characteristics Usage
Read-only Data type
Values
empty if the evaluation of a task object function does not produce an error or if the error did not provide a message. ErrorMessage has the same value as the message property of the tasks Error. When using a job manager, ErrorMessage indicates if a worker has crashed during task execution, but when using a different scheduler ErrorMessage does not indicate a crash.
Examples
See Also
Properties
Error, ErrorIdentifier, Function
16-34
FailedAttemptInformation
Purpose Description
Information returned from failed task If a task reruns because of certain system failures, the task property FailedAttemptInformation stores information related to the failure and rerun attempts. Note The FailedAttemptInformation property is available only when using the MathWorks job manager as your scheduler.
Characteristics Usage
Read-only Data type
Values
The data type of FailedAttemptInformation is an array of objects, one object for each rerun of the task. The property values of each resulting object contain information about when the task was rerun and the error that caused it.
See Also
Properties
AttemptedNumberOfRetries, MaximumNumberOfRetries
16-35
FileDependencies
Purpose Description
worker will need to access for evaluating a jobs tasks. The value of the property is defined by the client session. You set the value for the property as a cell array of strings. Each string is an absolute or relative pathname to a directory or file. The toolbox makes a zip file of all the files and directories referenced in the property. (Note: If the files or directories change while they are being zipped, or if any of the directories are empty, a failure or error can result.) The first time a worker evaluates a task for a particular job, the scheduler passes to the worker the zip file of the files and directories in the FileDependencies property. On the worker machine, the file is unzipped, and a directory structure is created that is exactly the same as that accessed on the client machine where the property was set. Those entries listed in the property value are added to the top of the path in the MATLAB worker session. (The subdirectories of the entries are not added to the path, even though they are included in the directory structure.) To find out where the unzipping occurs on the worker machine, use the function getFileDependencyDir in code that runs on the worker. See Example 2, below. When the worker runs subsequent tasks for the same job, it uses the directory structure already set up by the jobs FileDependencies property for the first task it ran for that job. When you specify FileDependencies at the time of creating a job, the settings are combined with those specified in the applicable configuration, if any. (Setting FileDependencies on a job object after it is created does not combine the new setting with the configuration settings, but overwrites existing settings for that job.) The transfer and unzipping of FileDependencies occurs for each worker running a task for that particular job on a machine, regardless of how many workers run on that machine. Normally, the file dependencies are deleted from the worker machine when the job is completed, or when the next job begins.
16-36
FileDependencies
Characteristics Usage
Read-only Data type
Values Remarks
The value of FileDependencies is empty by default. If a pathname that does not exist is specified for the property value, an error is generated. There is a default limitation on the size of data transfers via the FileDependencies property. For more information on this limit, see Object Data Size Limitations on page 6-48. For alternative means of making data available to workers, see Share Code on page 8-29.
Examples
Example 1
Make available to a jobs workers the contents of the directories fd1 and fd2, and the file fdfile1.m.
set(job1,'FileDependencies',{'fd1' 'fd2' 'fdfile1.m'}) get(job1,'FileDependencies') ans = 'fd1' 'fd2' 'fdfile1.m'
Example 2
Suppose in your client MATLAB session you have the following folders on your MATLAB path:
dirA dirA\subdir1 dirA\subdir2 dirB
16-37
FileDependencies
Transfer the content of these folders to the worker machines, and add all these folders to the paths of the worker MATLAB sessions. On the client, execute the following code:
j = createJob(FileDependencies, {'dirA', 'dirB'}) % This includes the subfolders of dirA.
In the task function that executes on the workers, include the following code:
%First find where FileDependencies are unzipped: DepDir = getFileDependencyDir %The top folders are already on the path, so add subfolders: addpath(fullfile(DepDir,'dirA','subdir1'),... fullfile(DepDir,'dirA','subdir2'))
See Also
Functions
getFileDependencyDir, jobStartup, taskFinish, taskStartup
Properties
PathDependencies
16-38
FinishedFcn
Purpose Description
completes its execution. The callback executes in the local MATLAB session, that is, the session that sets the property, the MATLAB client. Notes The FinishedFcn property is available only when using the MathWorks job manager as your scheduler. The FinishedFcn property applies only in the client MATLAB session in which it is set. Later sessions that access the same job or task object do not inherit the setting from previous sessions.
Characteristics Usage
Read-only Data type
Values
The callback follows the same model as callbacks for Handle Graphics, passing to the callback function the object (job or task) that makes the call and an empty argument of event data.
Examples
Create a job and set its FinishedFcn property using a function handle to an anonymous function that sends information to the display.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm, 'Name', 'Job_52a'); set(j, 'FinishedFcn', ...
16-39
FinishedFcn
Run the job and note the output messages from the job and task FinishedFcn callbacks.
submit(j) Finished task: 1 Job_52a finished
See Also
Properties
QueuedFcn, RunningFcn
16-40
FinishTime
Purpose Description
If a task or job is stopped or is aborted due to an error condition, FinishTime will hold the time when the task or job was stopped or aborted.
Characteristics Usage
Read-only Data type
Values
FinishTime is assigned the job managers system time when the task or job has finished. If the job or task is in the failed state, its FinishTime property value is empty.
Examples
Create and submit a job, then get its StartTime and FinishTime.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); t1 = createTask(j, @rand, 1, {12,12}); t2 = createTask(j, @rand, 1, {12,12}); t3 = createTask(j, @rand, 1, {12,12}); t4 = createTask(j, @rand, 1, {12,12}); submit(j) waitForState(j,'finished') get(j,'StartTime') ans = Mon Jun 21 10:02:17 EDT 2004 get(j,'FinishTime') ans = Mon Jun 21 10:02:52 EDT 2004
16-41
FinishTime
See Also
Functions
cancel, submit
Properties
CreateTime, StartTime, SubmitTime
16-42
Function
Purpose Description
Characteristics Usage
Read-only Data type
See Also
Functions
createTask
Properties
InputArguments, NumberOfOutputArguments, OutputArguments
16-43
GetJobStateFcn
Purpose Description
running on a generic scheduler. This function lets you communicate with the scheduler, to provide any instructions beyond the normal toolbox action of retrieving the job state from disk. To identify the job for the scheduler, the function should include a call to getJobSchedulerData. The value returned from the function must be a valid State for a job, and replaces the value ordinarily returned from the original call to get, etc. This might be useful when the scheduler has more up-to-date information about the state of a job than what is stored by the toolbox. For example, the scheduler might be aware of a failure before the toolbox is aware. For more information and examples on using these functions and properties, see Manage Jobs with Generic Scheduler on page 8-50.
Characteristics Usage
Read-only Data type
Values
You can set GetJobStateFcn to any valid function handle or a cell array whose first element is a function handle.
See Also
Functions
get, getJobSchedulerData, setJobSchedulerData
Properties
State, SubmitFcn
16-44
HasSharedFilesystem
Purpose Description
the scheduler handles data transfers to and from the worker nodes. If HasSharedFilesystem is true (1), the workers access the job data directly.
Characteristics Usage
Read-only Data type
Values
The value of HasSharedFilesystem can be set to true (or logical 1) or false (or logical 0). When you perform get on the property, the value returned is logical 1 or logical 0.
See Also
Properties
DataLocation, FileDependencies, PathDependencies
16-45
HostAddress
Purpose Description
running the job manager or worker session to which the job manager object or worker object refers. You can match the HostAddress property to find a desired job manager or worker when creating an object with findResource.
Characteristics Usage
Read-only Data type
Examples
See Also
Functions
findResource
Properties
Computer, Hostname, WorkerMachineOsType
16-46
Hostname
Purpose Description
Name of host running job manager or worker session You can match the Hostname property to find a desired job manager or worker when creating the job manager or worker object with findResource.
Characteristics Usage
Read-only Data type
Examples
See Also
Functions
findResource
Properties
Computer, HostAddress, WorkerMachineOsType
16-47
ID
Purpose Description
Object identifier Each object has a unique identifier within its parent object. The ID value is assigned at the time of object creation. You can use the ID property value to distinguish one object from another, such as different tasks in the same job.
Characteristics Usage
Read-only Data type
Values
The first job created in a job manager has the ID value of 1, and jobs are assigned ID values in numerical sequence as they are created after that. The first task created in a job has the ID value of 1, and tasks are assigned ID values in numerical sequence as they are created after that.
Examples
The ID values are the only unique properties distinguishing these two tasks.
16-48
ID
See Also
Functions
createJob, createTask
Properties
Jobs, Tasks
16-49
IdleWorkers
Purpose Description
Idle workers available to run tasks The IdleWorkers property value indicates which workers are currently available to the job manager for the performance of job tasks.
Characteristics Usage
Read-only Data type
Values
As workers complete tasks and assume new ones, the lists of workers in BusyWorkers and IdleWorkers can change rapidly. If you examine these two properties at different times, you might see the same worker on both lists if that worker has changed its status between those times. If a worker stops unexpectedly, the job managers knowledge of that as a busy or idle worker does not get updated until the job manager runs the next job and tries to send a task to that worker.
Examples
Examine which workers are available to a job manager for immediate use to perform tasks.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); get(jm, 'NumberOfIdleWorkers')
See Also
Properties
BusyWorkers, ClusterSize, MaximumNumberOfWorkers, MinimumNumberOfWorkers, NumberOfBusyWorkers, NumberOfIdleWorkers
16-50
InputArguments
Purpose Description
Characteristics Usage
Read-only Data type
Values Examples
The forms and values of the input arguments are totally dependent on the task function. Create a task requiring two input arguments, then examine the tasks InputArguments property.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); t = createTask(j, @rand, 1, {2, 4}); get(t, 'InputArguments') ans = [2] [4]
See Also
Functions
createTask
Properties
Function, OutputArguments
16-51
IsUsingSecureCommunication
Purpose Description
True if job manager and workers use secure communication The IsUsingSecureCommunication property indicates whether secure communication is being used between the job manager and the workers. The mdce_def file sets the parameter that controls secure communication when the mdce process starts on the cluster nodes. Secure communication is required when running with SecurityLevel set to 3. It is optional at other security levels.
Characteristics Usage
Read-only Data type
See Also
Functions
changePassword, clearLocalPassword
Properties
PromptForPassword, SecurityLevel, UserName
16-52
JobData
Purpose Description
Data made available to all workers for jobs tasks The JobData property holds data that eventually gets stored in the local memory of the worker machines, so that it does not have to be passed to the worker for each task in a job that the worker evaluates. Passing the data only once per job to each worker is more efficient than passing data with each task. Note, that to access the data contained in a jobs JobData property, the worker session evaluating the task needs to have access to the job, which it gets from a call to the function getCurrentJob, as discussed in the example below.
Characteristics Usage
Read-only Data type
Values Examples
Create job1 and set its JobData property value to the contents of array1.
job1 = createJob(jm) set(job1, 'JobData', array1) createTask(job1, @myfunction, 1, {task_data})
Now the contents of array1 are available to all the tasks in the job. Because the job itself must be accessible to the tasks, myfunction must include a call to the function getCurrentJob. That is, the task function myfunction needs to call getCurrentJob to get the job object through which it can get the JobData property. So myfunction should contain lines like the following:
cj = getCurrentJob array1 = get(cj, 'JobData')
16-53
JobData
See Also
Functions
createJob, createTask
16-54
JobDescriptionFile
Purpose Description
Name of XML job description file for Microsoft Windows HPC Server scheduler The XML file you specify by the JobDescriptionFile property defines the base state from which the job is created. The file must exist on the MATLAB path or the property must specify the full path name to the file. Any job properties that are specified as part of MATLAB job objects (e.g., MinimumNumberOfWorkers, MaximumNumberOfWorkers, etc., for parallel or MATLAB pool jobs) override the values specified in the job description file. Scheduler properties (e.g., nonempty JobTemplate property) also override the values specified in the job description file. For SOA jobs the values in the job description file are ignored. For version 2 of Windows HPC Server 2008, the values for HPC Server job properties specified in the job description file must be compatible with the values in the job template that is applied to the job (either the default job template or the job template specified by the JobTemplate property). Incompatibilities between property values specified by the job description file and the job template might result in an error when you submit a job. For example, if the job template imposes property restrictions that you violate in your job description file, you get an error. For information about job description files, consult Microsoft online documentation at: http://technet.microsoft.com/en-us/library/cc972801(WS.10).aspx
Characteristics Usage
Read-only Data type
See Also
Properties
ClusterVersion, JobTemplate, UseSOAJobSubmission
16-55
JobManager
Purpose Description
Characteristics Usage
Read-only Data type
Properties
BusyWorkers, IdleWorkers
16-56
Jobs
Purpose Description
Jobs contained in job manager service or in schedulers data location The Jobs property contains an array of all the job objects in a scheduler. Job objects will be in the order indicated by their ID property, consistent with the sequence in which they were created, regardless of their State. (To see the jobs categorized by state or the scheduled execution sequence for jobs in the queue, use the findJob function.)
Characteristics Usage
Read-only Data type
Examples
Examine the Jobs property for a job manager, and use the resulting array of objects to set property values.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j1 = createJob(jm); j2 = createJob(jm); j3 = createJob(jm); j4 = createJob(jm); . . . all_jobs = get(jm, 'Jobs') set(all_jobs, 'MaximumNumberOfWorkers', 10);
The last line of code sets the MaximumNumberOfWorkers property value to 10 for each of the job objects in the array all_jobs.
16-57
Jobs
See Also
Functions
createJob, destroy, findJob, submit
Properties
Tasks
16-58
JobTemplate
Purpose Description
Name of job template for HPC Server 2008 scheduler JobTemplate identifies the name of a job template to use with your HPC Server scheduler. The property value is not case-sensitive. With HPC Server 2008, if you do not specify a value for the JobTemplate property, the scheduler uses the default job template to run the job. Ask your system administrator which job template you should use. For SOA jobs, the specified job template used for submitting SOA jobs must not impose any restrictions on the name of the job, otherwise these jobs fail.
Characteristics Usage
Read-only Data type
Values
JobTemplate is an empty string by default. Job templates apply only for HPC Server 2008 clusters, and your scheduler ClusterVersion property must be set to 'HPCServer2008'. If ClusterVersion is set to any other value, and you attempt to set JobTemplate to a nonempty
string, an error is generated and the property value remains as a nonempty string.
Remarks
If you change the value of ClusterVersion or SchedulerHostname, this resets the values of ClusterSize, JobTemplate, and UseSOAJobSubmission.
See Also
Properties
ClusterVersion, JobDescriptionFile, UseSOAJobSubmission
16-59
LabGrid
Purpose Description
For more information on 2dbc distribution and lab grids of distributed arrays, see 2-Dimensional Distribution on page 5-17.
Characteristics Usage
Read-only Data type
See Also
Functions
codistributor2dbc, numlabs
Properties
BlockSize, Orientation
16-60
MasterName
Purpose Description
Characteristics Usage
Read-only Data type
Properties
ClusterName
16-61
MatlabCommandToRun
Purpose Description
Characteristics Usage
Read-only Data type
is created.
Properties
ClusterMatlabRoot, SubmitFcn
16-62
MaximumNumberOfRetries
Purpose Description
Specify maximum number of times to rerun failed task If a task cannot complete because of certain system failures, the job manager can attempt to rerun the task. MaximumNumberOfRetries specifies how many times to try to run the task after such failures. The task reruns until it succeeds or until it reaches the specified maximum number of attempts. Note The MaximumNumberOfRetries property is available only when using the MathWorks job manager as your scheduler.
Characteristics Usage
Read-only Data type
Properties
AttemptedNumberOfRetries, FailedAttemptInformation
16-63
MaximumNumberOfWorkers
Purpose Description
Specify maximum number of workers to perform job tasks With MaximumNumberOfWorkers you specify the greatest number of workers to be used to perform the evaluation of the jobs tasks at any one time. Tasks may be distributed to different workers at different times during execution of the job, so that more than MaximumNumberOfWorkers might be used for the whole job, but this property limits the portion of the cluster used for the job at any one time.
Characteristics Usage
Read-only Data type
Values Examples
You can set the value to anything equal to or greater than the value of the MinimumNumberOfWorkers property. Set the maximum number of workers to perform a job.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); set(j, 'MaximumNumberOfWorkers', 12);
In this example, the job will use no more than 12 workers, regardless of how many tasks are in the job and how many workers are available on the cluster.
See Also
Properties
BusyWorkers, ClusterSize, IdleWorkers, MinimumNumberOfWorkers, NumberOfBusyWorkers, NumberOfIdleWorkers
16-64
MinimumNumberOfWorkers
Purpose Description
Specify minimum number of workers to perform job tasks With MinimumNumberOfWorkers you specify the minimum number of workers to perform the evaluation of the jobs tasks. When the job is queued, it will not run until at least this many workers are simultaneously available. If MinimumNumberOfWorkers workers are available to the job manager, but some of the task dispatches fail due to network or node failures, such that the number of tasks actually dispatched is less than MinimumNumberOfWorkers, the job will be canceled.
Characteristics Usage
Read-only Data type
Values Examples
The default value is 1. You can set the value anywhere from 1 up to or equal to the value of the MaximumNumberOfWorkers property. Set the minimum number of workers to perform a job.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); set(j, 'MinimumNumberOfWorkers', 6);
In this example, when the job is queued, it will not begin running tasks until at least six workers are available to perform task evaluations.
See Also
Properties
BusyWorkers, ClusterSize, IdleWorkers, MaximumNumberOfWorkers, NumberOfBusyWorkers, NumberOfIdleWorkers
16-65
MpiexecFileName
Purpose Description
Characteristics Usage
Read-only Data type
Remarks
See your network administrator to find out which mpiexec you should run. The default value of the property points the mpiexec included in your MATLAB installation.
See Also
Functions
mpiLibConf, mpiSettings
Properties
SubmitArguments
16-66
Name
Purpose Description
Name of job manager, job, or worker object The descriptive name of a job manager or worker is set when its service is started, as described in "Customizing Engine Services" in the MATLAB Distributed Computing Server System Administrators Guide. This is reflected in the Name property of the object that represents the service. You can use the name of the job manager or worker service to search for the particular service when creating an object with the findResource function. You can configure Name as a descriptive name for a job object at any time before the job is submitted to the queue.
Characteristics Usage
Read-only Data type
Job manager object, job object, or worker object Always for a job manager or worker object; after job object is submitted String
Values Examples
By default, a job object is constructed with a Name created by concatenating the Name of the job manager with _job. Construct a job manager object by searching for the name of the service you want to use.
jm = findResource('scheduler','type','jobmanager', ... 'Name','MyJobManager');
16-67
Name
Change the jobs Name property and verify the new setting.
set(j,'Name','MyJob') get(j,'Name') ans = MyJob
See Also
Functions
findResource, createJob
16-68
NumberOfBusyWorkers
Purpose Description
Number of workers currently running tasks The NumberOfBusyWorkers property value indicates how many workers are currently running tasks for the job manager.
Characteristics Usage
Read-only Data type
Values Examples
The value of NumberOfBusyWorkers can range from 0 up to the total number of workers registered with the job manager. Examine the number of workers currently running tasks for a job manager.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); get(jm, 'NumberOfBusyWorkers')
See Also
Properties
BusyWorkers, ClusterSize, IdleWorkers, MaximumNumberOfWorkers, MinimumNumberOfWorkers, NumberOfIdleWorkers
16-69
NumberOfIdleWorkers
Purpose Description
Number of idle workers available to run tasks The NumberOfIdleWorkers property value indicates how many workers are currently available to the job manager for the performance of job tasks. If the NumberOfIdleWorkers is equal to or greater than the MinimumNumberOfWorkers of the job at the top of the queue, that job can start running.
Characteristics Usage
Read-only Data type
Values Examples
The value of NumberOfIdleWorkers can range from 0 up to the total number of workers registered with the job manager. Examine the number of workers available to a job manager.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); get(jm, 'NumberOfIdleWorkers')
See Also
Properties
BusyWorkers, ClusterSize, IdleWorkers, MaximumNumberOfWorkers, MinimumNumberOfWorkers, NumberOfBusyWorkers
16-70
NumberOfOutputArguments
Purpose Description
Number of arguments returned by task function When you create a task with the createTask function, you define how many output arguments are expected from the task function.
Characteristics Usage
Read-only Data type
Values Examples
A matrix is considered one argument. Create a task and examine its NumberOfOutputArguments property.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); t = createTask(j, @rand, 1, {2, 4}); get(t,'NumberOfOutputArguments') ans = 1
This example returns a 2-by-4 matrix, which is a single argument. The NumberOfOutputArguments value is set by the createTask function, as the argument immediately after the task function definition; in this case, the 1 following the @rand argument.
See Also
Functions
createTask
Properties
OutputArguments
16-71
Orientation
Purpose Description
the labs are organized within the lab grid. Supported orientation values are 'row' and 'col'. You can read this property only by using dot-notation; not the get function. For more information on 2dbc distribution of arrays, see 2-Dimensional Distribution on page 5-17.
Characteristics Usage
Read-only Data type
See Also
Functions
codistributor2dbc
Properties
BlockSize, LabGrid
16-72
OutputArguments
Purpose Description
corresponds to each output argument requested from task evaluation. If the tasks NumberOfOutputArguments property value is 0, or if the evaluation of the task produced an error, the cell array is empty.
Characteristics Usage
Read-only Data type
Values Examples
The forms and values of the output arguments are totally dependent on the task function. Create a job with a task and examine its result after running the job.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); t = createTask(j, @rand, 1, {2, 4}); submit(j)
Because each task returns a cell array, allresults is a cell array of cell arrays.
16-73
OutputArguments
See Also
Functions
createTask, getAllOutputArguments
Properties
Function, InputArguments, NumberOfOutputArguments
16-74
ParallelSubmissionWrapperScript
Purpose Description
PBS Pro, or TORQUE scheduler to run when starting labs for a parallel job.
Characteristics Usage
Read-only Data type
Values
ParallelSubmissionWrapperScript is a string specifying the full path to the script. This property value is set when you execute the function setupForParallelExecution, so you do not need to set the value directly. The property value then points to the appropriate wrapper script in matlabroot/toolbox/distcomp/bin/util.
See Also
Functions
createParallelJob, setupForParallelExecution, submit
Properties
ClusterName, ClusterMatlabRoot, MasterName, SubmitArguments
16-75
ParallelSubmitFcn
Purpose Description
For more information, see MATLAB Client Submit Function on page 8-35.
Characteristics Usage
Read-only Data type
Values
that uses the user-defined parallel submit function. For more information about parallel submit functions and where to find example templates you can use, see Use the Generic Scheduler Interface on page 9-8.
See Also
Functions
createParallelJob, submit
Properties
MatlabCommandToRun, SubmitFcn
16-76
Parent
Purpose Description
Parent object of job or task A jobs Parent property indicates the job manager or scheduler object that contains the job. A tasks Parent property indicates the job object that contains the task.
Characteristics Usage
Read-only Data type
Job object or task object Always Job manager, scheduler, or job object
See Also
Properties
Jobs, Tasks
16-77
Partition
Purpose Description
an array among the labs. You can read this property only by using dot-notation; not the get function.
Characteristics Usage
Read-only Data type
Examples
returns [3 3 2 2] .
See Also
Functions
codistributor1d
Properties
Dimension
16-78
PathDependencies
Purpose Description
When you specify PathDependencies at the time of creating a job, the settings are combined with those specified in the applicable configuration, if any. (Setting PathDependencies on a job object after it is created does not combine the new setting with the configuration settings, but overwrites existing settings for that job.)
Characteristics Usage
Read-only Data type
Values
PathDependencies is empty by default. For a mixed-platform environment, the strings can specify both UNIX-based and Microsoft Windows-based paths; those not appropriate or not found for a particular node generate warnings and are ignored.
Remarks Examples
For alternative means of making data available to workers, see Share Code on page 8-29. Set the MATLAB worker path in a mixed-platform environment to use functions in both the central repository (/central/funcs) and the department archive (/dept1/funcs).
sch = findResource('scheduler','name','LSF') job1 = createJob(sch) p = {'/central/funcs','/dept1/funcs', ... '\\OurDomain\central\funcs','\\OurDomain\dept1\funcs'} set(job1, 'PathDependencies', p)
16-79
PathDependencies
See Also
Properties
ClusterMatlabRoot, FileDependencies
16-80
PreviousJob
Purpose Description
evaluated.
Characteristics Usage
Read-only Data type
Properties
CurrentJob, CurrentTask, PreviousTask, Worker
16-81
PreviousTask
Purpose Description
evaluated.
Characteristics Usage
Read-only Data type
Properties
CurrentJob, CurrentTask, PreviousJob, Worker
16-82
PromptForPassword
Purpose Description
Specify if system should prompt for password when authenticating user The PromptForPassword property is true by default, so that when you access a job manager object, if you do not already have a password stored, the system prompts you to enter it. Setting PromptForPassword to false causes the system to generate an error when a password is required. This can be useful when you have a noninteractive script or function that programmatically accesses the job manager, and you might prefer an error rather than a password prompt.
Characteristics Usage
Read-only Data type
See Also
Functions
changePassword, clearLocalPassword
Properties
IsUsingSecureCommunication, SecurityLevel, UserName
16-83
QueuedFcn
Purpose Description
Specify function file to execute when job is submitted to job manager queue
QueuedFcn specifies the function file to execute when a job is submitted
to a job manager queue. The callback executes in the local MATLAB session, that is, the session that sets the property. Notes The QueuedFcn property is available only when using the MathWorks job manager as your scheduler. The QueuedFcn property applies only in the client MATLAB session in which it is set. Later sessions that access the same job object do not inherit the setting from previous sessions.
Characteristics Usage
Read-only Data type
Values Examples
Create a job and set its QueuedFcn property, using a function handle to an anonymous function that sends information to the display.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm, 'Name', 'Job_52a'); set(j, 'QueuedFcn', ... @(job,eventdata) disp([job.Name ' now queued for execution.'])) . . .
16-84
QueuedFcn
See Also
Functions
submit
Properties
FinishedFcn, RunningFcn
16-85
RcpCommand
Purpose Description
Command to copy files from client When using a nonshared file system, the command specified by this propertys value is used on the cluster to copy files from the client machine. The syntax of the command must be compatible with standard rcp. On MicrosoftWindows operating systems, the cluster machines must have a suitable installation of rcp.
Characteristics Usage
Read-only Data type
16-86
ResourceTemplate
Purpose Description
Resource definition for PBS Pro or TORQUE scheduler The value of this property is used to build the resource selection portion of the qsub command, generally identified by the -l flag. The toolbox uses this to identify the number of tasks in a parallel job, and you might want to fill out other selection subclauses (such as the OS type of the workers). You should specify a value for this property that includes the literal string ^N^ , which the toolbox will replace with the number of workers in the parallel job prior to submission.
Characteristics Usage
Read-only Data type
Values
You might set the property value as follows, to accommodate your cluster size and to set the wall time limit of the job (i.e., how long it is allowed to run in real time) to one hour: '-l select=^N^,walltime=1:00:00' (for a PBS Pro scheduler) '-l nodes=^N^,walltime=1:00:00' (for a TORQUE scheduler)
16-87
RestartWorker
Purpose Description
Specify whether to restart MATLAB workers before evaluating job tasks In some cases, you might want to restart MATLAB on the workers before they evaluate any tasks in a job. This action resets defaults, clears the workspace, frees available memory, and so on.
Characteristics Usage
Read-only Data type
Values
Set RestartWorker to true (or logical 1) if you want the job to restart the MATLAB session on any workers before they evaluate their first task for that job. The workers are not reset between tasks of the same job. Set RestartWorker to false (or logical 0) if you do not want MATLAB restarted on any workers. When you perform get on the property, the value returned is logical 1 or logical 0. The default value is 0, which does not restart the workers. Create a job and set it so that MATLAB workers are restarted before evaluating tasks in a job.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); set(j, 'RestartWorker', true) . . . submit(j)
Examples
See Also
Functions
submit
16-88
RshCommand
Purpose Description
Remote execution command used on worker nodes during parallel job Used on only UNIX operating systems, the value of this property is the command used at the beginning of running parallel jobs, typically to start MPI daemon processes on the nodes allocated to run MATLAB workers. The remote execution must be able to proceed without user interaction, for example, without prompting for user credentials.
Characteristics Usage
Read-only Data type
16-89
RunningFcn
Purpose Description
The callback executes in the local MATLAB client session, that is, the session that sets the property. Notes The RunningFcn property is available only when using the MathWorks job manager as your scheduler. The RunningFcn property applies only in the client MATLAB session in which it is set. Later sessions that access the same job or task object do not inherit the setting from previous sessions.
Characteristics Usage
Read-only Data type
Values Examples
Create a job and set its QueuedFcn property, using a function handle to an anonymous function that sends information to the display.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm, 'Name', 'Job_52a'); set(j, 'RunningFcn', ... @(job,eventdata) disp([job.Name ' now running.'])) . . . submit(j)
16-90
RunningFcn
See Also
Functions
submit
Properties
FinishedFcn, QueuedFcn
16-91
SchedulerHostname
Purpose Description
Characteristics Usage
Read-only Data type
If you change the value of SchedulerHostname, this resets the values of ClusterSize, JobTemplate, and UseSOAJobSubmission.
Properties
ClusterOsType
16-92
SecurityLevel
Purpose Description
Security level controlling access to job manager and its jobs The SecurityLevel property indicates the degree of security applied to the job manager and its jobs. The mdce_def file sets the parameter that controls security level when the mdce process starts on the cluster nodes.
Characteristics Usage
Read-only Data type
Values
The property values indicating security level and their effects are shown in the following table. Security Level 0 1 Effect No security. All users can access all jobs; the
AuthorizedUsers property of the job is ignored.
You are warned when you try to access other users jobs and tasks, but can still perform all actions. You can suppress the warning by adding your user name to the AuthorizedUsers property of the job.
16-93
SecurityLevel
Security Level 2
Effect Authentication required. You must enter a password to access any jobs and tasks. You cannot access other users jobs unless your user name is included in the jobs AuthorizedUsers property. Same as level 2, but in addition, tasks run on the workers as the user to whom the job belongs. The user name and password for authentication in the client session need to be the same as the system password used to log on to a worker machine. NOTE: This level requires secure communication between job manager and workers. Secure communication is also set in the mdce_def file, and is indicated by a job managers IsUsingSecureCommunication property.
The job manager and the workers should run at the same security level. A worker running at too low a security level will fail to register with the job manager, because the job manager does not trust it.
See Also
Functions
changePassword, clearLocalPassword
Properties
AuthorizedUsers, IsUsingSecureCommunication, PromptForPassword, UserName
16-94
ServerName
Purpose Description
Characteristics Usage
Read-only Data type
See Also
Properties
ClusterOsType
16-95
StartTime
Purpose Description
Characteristics Usage
Read-only Data type
Values Examples
StartTime is assigned the job managers system time when the task or job has started running.
Create and submit a job, then get its StartTime and FinishTime.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); t1 = createTask(j, @rand, 1, {12,12}); t2 = createTask(j, @rand, 1, {12,12}); t3 = createTask(j, @rand, 1, {12,12}); t4 = createTask(j, @rand, 1, {12,12}); submit(j) waitForState(j, 'finished') get(j, 'StartTime') ans = Mon Jun 21 10:02:17 EDT 2004 get(j, 'FinishTime') ans = Mon Jun 21 10:02:52 EDT 2004
16-96
StartTime
See Also
Functions
submit
Properties
CreateTime, FinishTime, SubmitTime
16-97
State
Purpose Description
Current state of task, job, job manager, or worker The State property reflects the stage of an object in its life cycle, indicating primarily whether or not it has yet been executed. The possible State values for all Parallel Computing Toolbox objects are discussed below in the Values section. Note The State property of the task object is different than the State property of the job object. For example, a task that is finished may be part of a job that is running if other tasks in the job have not finished.
Characteristics Usage
Read-only Data type
Values
Task Object
For a task object, possible values for State are pending Tasks that have not yet started to evaluate the task objects Function property are in the pending state. running Task objects that are currently in the process of evaluating the Function property are in the running state. finished Task objects that have finished evaluating the task objects Function property are in the finished state. unavailable Communication cannot be established with the job manager.
16-98
State
Job Object
For a job object, possible values for State are pending Job objects that have not yet been submitted to a job queue are in the pending state. queued Job objects that have been submitted to a job queue but have not yet started to run are in the queued state. running Job objects that are currently in the process of running are in the running state. finished Job objects that have completed running all their tasks are in the finished state. failed Job objects when using a third-party scheduler and the job could not run because of unexpected or missing information. destroyed Job objects whose data has been permanently removed from the data location or job manager. unavailable Communication cannot be established with the job manager.
Job Manager
For a job manager, possible values for State are running A started job queue will execute jobs normally. paused The job queue is paused. unavailable Communication cannot be established with the job manager. When a job manager first starts up, the default value for State is
running.
16-99
State
Worker
For a worker, possible values for State are running A started job queue will execute jobs normally. unavailable Communication cannot be established with the worker.
Examples
Create a job manager object representing a job manager service, and create a job object; then examine each objects State property.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); get(jm, 'State') ans = running j = createJob(jm); get(j, 'State') ans = pending
See Also
Functions
createJob, createTask, findResource, pause, resume, submit
16-100
SubmitArguments
Purpose Description
Specify additional arguments to use when submitting job to Platform LSF, PBS Pro, TORQUE, or mpiexec scheduler
SubmitArguments is simply a string that is passed via the bsub or qsub
command to the LSF, PBS Pro, or TORQUE scheduler at submit time, or passed to the mpiexec command if using an mpiexec scheduler.
Characteristics Usage
Read-only Data type
Values
LSF Scheduler
Useful SubmitArguments values might be '-m "machine1 machine2"' to indicate that your scheduler should use only the named machines to run the job, or '-R "type==LINUX64"' to use only workers running on 64-bit machines with a Linux operating system. Note that by default the scheduler will attempt to run your job on only nodes with an architecture similar to the local machines unless you specify '-R "type==any"'.
mpiexec Scheduler
The following SubmitArguments values might be useful when using an mpiexec scheduler. They can be combined to form a single string when separated by spaces.
16-101
SubmitArguments
Value
-phrase MATLAB -noprompt -localonly -host <hostname> -machinefile <filename>
Description Use MATLAB as passphrase to connect with smpd. Suppress prompting for any user information. Run only on the local computer. Run only on the identified host. Run only on the nodes listed in the specified file (one hostname per line).
For a complete list, see the command-line help for the mpiexec command:
mpiexec -help mpiexec -help2
See Also
Functions
submit
Properties
MatlabCommandToRun, MpiexecFileName
16-102
SubmitFcn
Purpose Description
generic scheduler. The function runs in the MATLAB client. This user-defined submit function provides certain job and task data for the MATLAB worker, and identifies a corresponding decode function for the MATLAB worker to run. For further information, see MATLAB Client Submit Function on page 8-35.
Characteristics Usage
Read-only Data type
Values
SubmitFcn can be set to any valid MATLAB callback value that uses the user-defined submit function.
For a description of the user-defined submit function, how it is used, and its relationship to the worker decode function, see Use the Generic Scheduler Interface on page 8-34.
See Also
Functions
submit
Properties
MatlabCommandToRun
16-103
SubmitTime
Purpose Description
Characteristics Usage
Read-only Data type
Values Examples
SubmitTime is assigned the job managers system time when the job is
See Also
Functions
submit
Properties
CreateTime, FinishTime, StartTime
16-104
Tag
Purpose Description
Specify label to associate with job object You configure Tag to be a string value that uniquely identifies a job object.
Tag is particularly useful in programs that would otherwise need to define the job object as a global variable, or pass the object as an argument between callback routines.
You can return the job object with the findJob function by specifying the Tag property value.
Characteristics Usage
Read-only Data type
Values Examples
The default value is an empty string. Suppose you create a job object in the job manager jm.
job1 = createJob(jm);
You can identify and access job1 using the findJob function and the Tag property value.
job_one = findJob(jm,'Tag','MyFirstJob');
See Also
Functions
findJob
16-105
Task
Purpose Description
First task contained in MATLAB pool job object The Task property contains the task object for the MATLAB pool job, which has only this one task. This is the same as the first task contained in the Tasks property.
Characteristics Usage
Read-only Data type
See Also
Functions
createMatlabPoolJob, createTask
Properties
Tasks
16-106
Tasks
Purpose Description
Tasks contained in job object The Tasks property contains an array of all the task objects in a job, whether the tasks are pending, running, or finished. Tasks are always returned in the order in which they were created.
Characteristics Usage
Read-only Data type
Examples
Examine the Tasks property for a job object, and use the resulting array of objects to set property values.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); createTask(j, ...) . . . createTask(j, ...) alltasks = get(j, 'Tasks') alltasks = distcomp.task: 10-by-1 set(alltasks, 'Timeout', 20);
The last line of code sets the Timeout property value to 20 seconds for each task in the job.
16-107
Tasks
See Also
Functions
createTask, destroy, findTask
Properties
Jobs
16-108
Timeout
Purpose Description
Specify time limit to complete task or job Timeout holds a double value specifying the number of seconds to wait before giving up on a task or job. The time for timeout begins counting when the task State property value changes from the Pending to Running, or when the job object State property value changes from Queued to Running. When a task times out, the behavior of the task is the same as if the task were stopped with the cancel function, except a different message is placed in the task objects ErrorMessage property. When a job times out, the behavior of the job is the same as if the job were stopped using the cancel function, except all pending and running tasks are treated as having timed out.
Characteristics Usage
Read-only Data type
Values
The default value for Timeout is large enough so that in practice, tasks and jobs will never time out. You should set the value of Timeout to the number of seconds you want to allow for completion of tasks and jobs. Set a jobs Timeout value to 1 minute.
jm = findResource('scheduler','type','jobmanager', ... 'name','MyJobManager','LookupURL','JobMgrHost'); j = createJob(jm); set(j, 'Timeout', 60)
Examples
16-109
Timeout
See Also
Functions
submit
Properties
ErrorMessage, State
16-110
Type
Purpose Description
Characteristics Usage
Read-only Data type
Values
object.
16-111
UserData
Purpose Description
Specify data to associate with object You configure UserData to store data that you want to associate with an object. The object does not use this data directly, but you can access it using the get function or dot notation.
UserData is stored in the local MATLAB client session, not in the job manager, job data location, or worker. So, one MATLAB client session cannot access the data stored in this property by another MATLAB client session. Even on the same machine, if you close the client session where UserData is set for an object, and then access the same object from a later client session via the job manager or job data location, the original UserData is not recovered. Likewise, commands such as clear all clear functions
will clear an object in the local session, permanently removing the data in the UserData property.
Characteristics Usage
Read-only Data type
Values Examples
The default value is an empty vector. Suppose you create the job object job1.
job1 = createJob(jm);
16-112
UserData
16-113
UserName
Purpose Description
User who created job or job manager object On a job, the UserName property value is a string indicating the login name of the user who created the job. On a job manager object, the UserName property value indicates the user who created the object or who is using the job manager object to access jobs in its queue.
Characteristics Usage
Read-only
Job object or job manager object Always for job object Never for job manager object, but can be password protected
Data type
String
Examples
Change the user for a job manager object in your current MATLAB session. Certain security levels display a password prompt.
jm = findResource('scheduler','type','jobmanager','name','central-jm'); set(jm, 'UserName', 'MyNewName')
See Also
These references apply to using the UserName property for job manager objects.
Functions
changePassword, clearLocalPassword
Properties
IsUsingSecureCommunication, PromptForPassword, SecurityLevel
16-114
UseSOAJobSubmission
Purpose Description
Allow service-oriented architecture (SOA) submission on HPC Server 2008 cluster The value you assign to the UseSOAJobSubmission property specifies whether to allow SOA job submissions for the scheduler object representing a Microsoft Windows HPC Server 2008 cluster. If you enable SOA submission, MATLAB worker sessions can each evaluate multiple tasks in succession. If you disable SOA submission, a separate MATLAB worker starts for each task. Ensure that HPC Server 2008 is correctly configured to run SOA jobs on MATLAB Distributed Computing Server. For more details, see the online installation instructions at http://www.mathworks.com/distconfig. Note The MATLAB client from which you submit SOA jobs to the HPC Server 2008 scheduler must remain open for the duration of these jobs. Closing the MATLAB client session while SOA jobs are in the pending, queued, or running state causes the scheduler to cancel these jobs.
Characteristics Usage
Read-only Data type
Values
UseSOAJobSubmission is false by default. SOA job submission works only for HPC Server 2008 clusters, and your scheduler ClusterVersion property must be set to 'HPCServer2008'. If ClusterVersion is set to any other value, and you attempt to set UseSOAJobSubmission to true, an error is generated and the property value remains false.
16-115
UseSOAJobSubmission
Remarks
If you change the value of ClusterVersion or SchedulerHostname, this resets the values of ClusterSize, JobTemplate, and UseSOAJobSubmission. Set the scheduler to allow SOA job submissions.
s = findResource('scheduler', 'type', 'hpcserver'); s.UseSOAJobSubmission = true;
Examples
See Also
Properties
ClusterVersion, JobDescriptionFile, JobTemplate,
16-116
Worker
Purpose Description
Worker session that performed task The Worker property value is an object representing the worker session that evaluated the task.
Characteristics Usage
Read-only Data type
Values Examples
Before a task is evaluated, its Worker property value is an empty vector. Find out which worker evaluated a particular task.
submit(job1) waitForState(job1,'finished') t1 = findTask(job1,'ID',1) t1.Worker.Name ans = node55_worker1
See Also
Properties
Tasks
16-117
WorkerMachineOsType
Purpose Description
Specify operating system of nodes on which mpiexec scheduler will start labs
WorkerMachineOsType specifies the operating system of the nodes that an mpiexec scheduler will start labs on for the running of a parallel job.
Characteristics Usage
Read-only Data type
Values
The only value the property can have is 'pc' or 'unix'. The nodes of the labs running an mpiexec job must all be the same platform. The only heterogeneous mixing allowed in the cluster for the same mpiexec job is Intel Macintosh-based systems with 32-bit Linux-based systems.
See Also
Properties
Computer, HostAddress, Hostname
16-118
Glossary
Glossary
CHECKPOINTBASE The name of the parameter in the mdce_def file that defines the location of the job manager and worker checkpoint directories. checkpoint directory Location where job manager checkpoint information and worker checkpoint information is stored. client The MATLAB session that defines and submits the job. This is the MATLAB session in which the programmer usually develops and prototypes applications. Also known as the MATLAB client. client computer The computer running the MATLAB client. cluster A collection of computers that are connected via a network and intended for a common purpose. coarse-grained application An application for which run time is significantly greater than the communication time needed to start and stop the program. Coarse-grained distributed applications are also called embarrassingly parallel applications. codistributed array An array partitioned into segments, with each segment residing in the workspace of a different lab. Composite An object in a MATLAB client session that provides access to data values stored on the labs in a MATLAB pool, such as the values of variables that are assigned inside an spmd statement. computer A system with one or more processors.
Glossary-1
Glossary
distributed application The same application that runs independently on several nodes, possibly with different input parameters. There is no communication, shared data, or synchronization points between the nodes. Distributed applications can be either coarse-grained or fine-grained. distributed computing Computing with distributed applications, running the application on several nodes simultaneously. distributed computing demos Demonstration programs that use Parallel Computing Toolbox software, as opposed to sequential demos. DNS Domain Name System. A system that translates Internet domain names into IP addresses.
dynamic licensing The ability of a MATLAB worker or lab to employ all the functionality you are licensed for in the MATLAB client, while checking out only an engine license. When a job is created in the MATLAB client with Parallel Computing Toolbox software, the products for which the client is licensed will be available for all workers or labs that evaluate tasks for that job. This allows you to run any code on the cluster that you are licensed for on your MATLAB client, without requiring extra licenses for the worker beyond MATLAB Distributed Computing Server software. For a list of products that are not eligible for use with Parallel Computing Toolbox software, see http://www.mathworks.com/products/ineligible_programs/. fine-grained application An application for which run time is significantly less than the communication time needed to start and stop the program. Compare to coarse-grained applications. head node Usually, the node of the cluster designated for running the job manager and license manager. It is often useful to run all the nonworker related processes on a single machine.
Glossary-2
Glossary
heterogeneous cluster A cluster that is not homogeneous. homogeneous cluster A cluster of identical machines, in terms of both hardware and software. job The complete large-scale operation to perform in MATLAB, composed of a set of tasks.
job manager The MathWorks process that queues jobs and assigns tasks to workers. A third-party process that performs this function is called a scheduler. The general term "scheduler" can also refer to a job manager. job manager checkpoint information Snapshot of information necessary for the job manager to recover from a system crash or reboot. job manager database The database that the job manager uses to store the information about its jobs and tasks. job manager lookup process The process that allows clients, workers, and job managers to find each other. It starts automatically when the job manager starts. lab When workers start, they work independently by default. They can then connect to each other and work together as peers, and are then referred to as labs.
LOGDIR The name of the parameter in the mdce_def file that defines the directory where logs are stored. MathWorks job manager See job manager.
Glossary-3
Glossary
MATLAB client See client. MATLAB pool A collection of labs that are reserved by the client for execution of parfor-loops or spmd statements. See also lab. MATLAB worker See worker. mdce The service that has to run on all machines before they can run a job manager or worker. This is the engine foundation process, making sure that the job manager and worker processes that it controls are always running. Note that the program and service name is all lowercase letters. mdce_def file The file that defines all the defaults for the mdce processes by allowing you to set preferences or definitions in the form of parameter values. MPI Message Passing Interface, the means by which labs communicate with each other while running tasks in the same job. A computer that is part of a cluster.
node
parallel application The same application that runs on several labs simultaneously, with communication, shared data, or synchronization points between the labs. private array An array which resides in the workspaces of one or more, but perhaps not all labs. There might or might not be a relationship between the values of these arrays among the labs.
Glossary-4
Glossary
random port A random unprivileged TCP port, i.e., a random TCP port above 1024. register a worker The action that happens when both worker and job manager are started and the worker contacts job manager. replicated array An array which resides in the workspaces of all labs, and whose size and content are identical on all labs. scheduler The process, either third-party or the MathWorks job manager, that queues jobs and assigns tasks to workers. spmd (single program multiple data) A block of code that executes simultaneously on multiple labs in a MATLAB pool. Each lab can operate on a different data set or different portion of distributed data, and can communicate with other participating labs while performing the parallel computations. task One segment of a job to be evaluated by a worker.
variant array An array which resides in the workspaces of all labs, but whose content differs on these labs. worker The MATLAB session that performs the task computations. Also known as the MATLAB worker or worker process. worker checkpoint information Files required by the worker during the execution of tasks.
Glossary-5
Glossary
Glossary-6
Index
A
Index
arrays codistributed 5-4 local 5-11 private 5-4 replicated 5-2 types of 5-2 variant 5-3
AttemptedNumberOfRetries property 16-2 AuthorizedUsers property 16-3
B
batch function 14-5 BlockSize property 16-5 BusyWorkers property 16-6
C
cancel function 14-9 CancelJobFcn property 16-7 CancelTaskFcn property 16-8 CaptureCommandWindowOutput property 16-9 ccsscheduler object 12-2 changePassword function 14-11 clear function 14-12 clearLocalPassword function 14-13 ClusterMatlabRoot property 16-11 ClusterName property 16-12 ClusterOsType property 16-13 ClusterSize property 16-14 ClusterVersion property 16-15
codistributed object 12-4 codistributed.build function 14-16 codistributed.cell function 14-18 codistributed.colon function 14-20 codistributed.eye function 14-22 codistributed.false function 14-24 codistributed.Inf function 14-26 codistributed.NaN function 14-28 codistributed.ones function 14-30 codistributed.rand function 14-32 codistributed.randn function 14-34 codistributed.spalloc function 14-36 codistributed.speye function 14-38 codistributed.sprand function 14-40 codistributed.sprandn function 14-42 codistributed.true function 14-44 codistributed.zeros function 14-46 codistributor function 14-48 codistributor1d function 14-51 codistributor1d object 12-6 codistributor1d.defaultPartition
function 14-54
codistributor2dbc function 14-55 codistributor2dbc object 12-7 codistributor2dbc.defaultBlockSize
property 16-16
codistributor2dbc.defaultLabGrid
function 14-57
CommandWindowOutput property 16-17
codistributed arrays constructor functions 5-10 creating 5-7 defined 5-4 indexing 5-15 working with 5-5 codistributed function 14-14
Composite getting started 1-11 outside spmd 3-10 Composite function 14-58 Composite object 12-8 Computer property 16-19 Configuration property 16-20 configurations 6-14 importing and exporting 6-21 using in application 6-25 validating 6-22
Index-1
Index
E
EnvironmentSetMethod property 16-30 Error property 16-31 ErrorIdentifier property 16-33 ErrorMessage property 16-34 exist function 14-99
current working directory MATLAB worker 6-31 CurrentJob property 16-23 CurrentTask property 16-24
D
DataLocation property 16-25 defaultParallelConfig function 14-69 demote function 14-71 destroy function 14-73 DestroyJobFcn property 16-27 DestroyTaskFcn property 16-28 dfeval function 14-74 dfevalasync function 14-78 diary function 14-80 Dimension property 16-29 distributed function 14-81 distributed object 12-12 distributed.cell function 14-82 distributed.eye function 14-83 distributed.false function 14-84 distributed.Inf function 14-85 distributed.NaN function 14-86 distributed.ones function 14-87 distributed.rand function 14-88 distributed.randn function 14-89 distributed.spalloc function 14-90 distributed.speye function 14-91 distributed.sprand function 14-92 distributed.sprandn function 14-93 distributed.true function 14-94 distributed.zeros function 14-95 dload function 14-96
F
FailedAttemptInformation property 16-35 feval function 14-100 FileDependencies property 16-36
files sharing 8-14 using an LSF scheduler 8-29 findJob function 14-102 findResource function 14-104 findTask function 14-109 FinishedFcn property 16-39 FinishTime property 16-41 for loop distributed 14-111 Function property 16-43 functions arrayfun 14-2 batch 14-5 cancel 14-9 changePassword 14-11 clear 14-12 clearLocalPassword 14-13 codistributed 14-14 codistributed.build 14-16 codistributed.cell 14-18 codistributed.colon 14-20 codistributed.eye 14-22
Index-2
Index
codistributed.false 14-24 codistributed.Inf 14-26 codistributed.NaN 14-28 codistributed.ones 14-30 codistributed.rand 14-32 codistributed.randn 14-34 codistributed.spalloc 14-36 codistributed.speye 14-38 codistributed.sprand 14-40 codistributed.sprandn 14-42 codistributed.true 14-44 codistributed.zeros 14-46 codistributor 14-48 codistributor1d 14-51 codistributor1d.defaultPartition 14-54 codistributor2dbc 14-55 codistributor2dbc.defaultLabGrid 14-57 Composite 14-58 createJob 14-59 createMatlabPoolJob 14-61 createParallelJob 14-63 createTask 14-66 defaultParallelConfig 14-69 demote 14-71 destroy 14-73 dfeval 14-74 dfevalasync 14-78 diary 14-80 distributed 14-81 distributed.cell 14-82 distributed.eye 14-83 distributed.false 14-84 distributed.Inf 14-85 distributed.NaN 14-86 distributed.ones 14-87 distributed.rand 14-88 distributed.randn 14-89 distributed.spalloc 14-90 distributed.speye 14-91 distributed.sprand 14-92
distributed.sprandn 14-93 distributed.true 14-94 distributed.zeros 14-95 dload 14-96 dsave 14-98 exist 14-99 feval 14-100 findJob 14-102 findResource 14-104 findTask 14-109 for
distributed 14-111 drange 14-111 gather 14-113 gcat 14-116 get 14-117
getAllOutputArguments 14-119 getCodistributor 14-121 getCurrentJob 14-123 getCurrentJobmanager 14-124 getCurrentTask 14-125 getCurrentWorker 14-126 getDebugLog 14-127 getFileDependencyDir 14-129 getJobSchedulerData 14-130 getLocalPart 14-131 globalIndices 14-132 gop 14-134 gplus 14-136 gpuArray 14-137 gpuDevice 14-138 gpuDeviceCount 14-139 help 14-140 importParallelConfig 14-141 inspect 14-143 isaUnderlying 14-145 iscodistributed 14-146 isComplete 14-147 isdistributed 14-148 isreplicated 14-149
Index-3
Index
jobStartup 14-150 labBarrier 14-151 labBroadcast 14-152 labindex 14-154 labProbe 14-155 labReceive 14-156 labSend 14-158 labSendReceive 14-159 length 14-162 load 14-163 matlabpool 14-165 methods 14-171 mpiLibConf 14-173 mpiprofile 14-175 mpiSettings 14-180 numlabs 14-182 parallel.gpu.CUDAKernel 14-183 parfor 14-185 pause 14-189 pctconfig 14-190 pctRunDeployedCleanup 14-192 pctRunOnAll 14-193 pload 14-194 pmode 14-196 poolStartup 14-199 promote 14-200 psave 14-202 redistribute 14-204 resume 14-205 set 14-206 setJobSchedulerData 14-209 setupForParallelExecution 14-210 size 14-212 sparse 14-213 spmd 14-215 submit 14-217 subsasgn 14-218 subsref 14-219 taskFinish 14-220 taskStartup 14-221
G
gather function 14-113 gcat function 14-116
generic scheduler distributed jobs 8-34 parallel jobs 9-8 genericscheduler object 12-14 get function 14-117 getAllOutputArguments function 14-119 getCodistributor function 14-121 getCurrentJob function 14-123 getCurrentJobmanager function 14-124 getCurrentTask function 14-125 getCurrentWorker function 14-126 getDebugLogp function 14-127 getFileDependencyDir function 14-129 getJobSchedulerData function 14-130 GetJobStateFcn property 16-44 getLocalPart function 14-131 globalIndices function 14-132 gop function 14-134 gplus function 14-136 gpuArray function 14-137 gpuArray object 12-17 gpuDevice function 14-138 GPUDevice object 12-10 12-19 gpuDeviceCount function 14-139
H
HasSharedFilesystem property 16-45 help command-line 6-12 help function 14-140 HostAddress property 16-46 Hostname property 16-47
Index-4
Index
I
ID property 16-48 IdleWorkers property 16-50 importParallelConfig function 14-141 InputArguments property 16-51 inspect function 14-143 isaUnderlying function 14-145 iscodistributed function 14-146 isComplete function 14-147 isdistributed function 14-148 isreplicated function 14-149 IsUsingSecureCommunication property 16-52
L
labBarrier function 14-151 labBroadcast function 14-152 LabGrid property 16-60 labindex function 14-154 labProbe function 14-155 labReceive function 14-156 labSend function 14-158 labSendReceive function 14-159 length function 14-162 load function 14-163 localscheduler object 12-28
J
job creating example 8-10 creating on generic scheduler example 8-45 creating on LSF or HPC Server scheduler example 8-25 life cycle 6-8 local scheduler 8-3 submitting to generic scheduler queue 8-47 submitting to local scheduler 8-5 submitting to LSF or HPC Server scheduler queue 8-27 submitting to queue 8-13 job manager finding example 8-3 8-8 job object 12-22 JobData property 16-53 JobDescriptionFile property 16-55 jobmanager object 12-25 JobManager property 16-56 Jobs property 16-57 jobStartup function 14-150 JobTemplate property 16-59
M
MasterName property 16-61 MatlabCommandToRun property 16-62
matlabpool parfor 2-4 spmd 3-3 matlabpool function 14-165 matlabpooljob object 12-32 MaximumNumberOfRetries property 16-63 MaximumNumberOfWorkers property 16-64 methods function 14-171 MinimumNumberOfWorkers property 16-65 mpiexec object 12-35 MpiexecFileName property 16-66 mpiLibConf function 14-173 mpiprofile function 14-175 mpiSettings function 14-180
N
Name property 16-67 NumberOfBusyWorkers property 16-69 NumberOfIdleWorkers property 16-70 NumberOfOutputArguments property 16-71
Index-5
Index
O
objects 6-7
ccsscheduler 12-2 codistributed 12-4 codistributor1d 12-6 codistributor2dbc 12-7 Composite 12-8 distributed 12-12 genericscheduler 12-14 gpuArray 12-17 GPUDevice 12-10 12-19 job 12-22 jobmanager 12-25 localscheduler 12-28 lsfscheduler 12-30 matlabpooljob 12-32 mpiexec 12-35 paralleljob 12-37 pbsproscheduler 12-40 RemoteClusterAccess 12-42
property 16-75
ParallelSubmitFcn property 16-76 Parent property 16-77 parfor function 14-185 parfor-loops 2-1
P
parallel for-loops. See parfor-loops parallel jobs 9-2 supported schedulers 9-4 parallel.gpu.CUDAKernel function 14-183
break 2-14 broadcast variables 2-23 classification of variables 2-17 compared to for-loops 2-6 error handling 2-9 for-drange 2-16 global variables 2-15 improving performance 2-32 limitations 2-10 local vs. cluster workers 2-15 loop variable 2-18 MATLAB path 2-9 nested functions 2-12 nested loops 2-12 nesting with spmd 2-14 nondistributable functions 2-12 persistent variables 2-15 programming considerations 2-9 reduction assignments 2-24 reduction assignments, associativity 2-26 reduction assignments, commutativity 2-27 reduction assignments, overloading 2-28 reduction variables 2-23 release compatibility 2-16 return 2-14 sliced variables 2-19 temporary variables 2-30 transparency 2-10 Partition property 16-78 PathDependencies property 16-79 pause function 14-189 PBS Pro scheduler 8-21 pbsproscheduler object 12-40
Index-6
Index
platforms supported 6-7 pload function 14-194 pmode function 14-196 poolStartup function 14-199 PreviousJob property 16-81 PreviousTask property 16-82 programming basic session 8-8 guidelines 6-30 local scheduler 8-2 tips 6-30 promote function 14-200 PromptForPassword property 16-83 properties AttemptedNumberOfRetries 16-2 AuthorizedUsers 16-3 BlockSize 16-5 BusyWorkers 16-6 CancelJobFcn 16-7 CancelTaskFcn 16-8 CaptureCommandWindowOutput 16-9 ClusterMatlabRoot 16-11 ClusterName 16-12 ClusterOsType 16-13 ClusterSize 16-14 ClusterVersion 16-15
codistributor2dbc.defaultBlockSize 16-16 CommandWindowOutput 16-17 Computer 16-19 Configuration 16-20 CreateTime 16-22 CurrentJob 16-23 CurrentTask 16-24 DataLocation 16-25 DestroyJobFcn 16-27 DestroyTaskFcn 16-28
Dimension 16-29 EnvironmentSetMethod 16-30 Error 16-31 ErrorIdentifier 16-33 ErrorMessage 16-34 FailedAttemptInformation 16-35 FileDependencies 16-36 FinishedFcn 16-39 FinishTime 16-41 Function 16-43 GetJobStateFcn 16-44 HasSharedFilesystem 16-45 HostAddress 16-46 Hostname 16-47 ID 16-48 IdleWorkers 16-50 InputArguments 16-51 IsUsingSecureCommunication 16-52 JobData 16-53 JobDescriptionFile 16-55 JobManager 16-56 Jobs 16-57 JobTemplate 16-59 LabGrid 16-60 MasterName 16-61 MatlabCommandToRun 16-62 MaximumNumberOfRetries 16-63 MaximumNumberOfWorkers 16-64 MinimumNumberOfWorkers 16-65 MpiexecFileName 16-66 Name 16-67 NumberOfBusyWorkers 16-69 NumberOfIdleWorkers 16-70 NumberOfOutputArguments 16-71 Orientation 16-72 OutputArguments 16-73 ParallelSubmissionWrapperScript 16-75 ParallelSubmitFcn 16-76 Parent 16-77 Partition 16-78
Index-7
Index
PathDependencies 16-79 PreviousJob 16-81 PreviousTask 16-82 PromptForPassword 16-83 QueuedFcn 16-84 RcpCommand 16-86 ResourceTemplate 16-87 RestartWorker 16-88 RshCommand 16-89 RunningFcn 16-90 SchedulerHostname 16-92 SecurityLevel 16-93 ServerName 16-95 StartTime 16-96 State 16-98 SubmitArguments 16-101 SubmitFcn 16-103 SubmitTime 16-104 Tag 16-105 Task 16-106 Tasks 16-107 Timeout 16-109 Type 16-111 UserData 16-112 UserName 16-114 UseSOAJobSubmission 16-115 Worker 16-117 WorkerMachineOsType 16-118 psave function 14-202
RestartWorker property 16-88 results local scheduler 8-6 retrieving 8-14 retrieving from job on generic scheduler 8-47 retrieving from job on LSF scheduler 8-28 resume function 14-205 RshCommand property 16-89 RunningFcn property 16-90
S
saving objects 6-32 scheduler generic interface distributed jobs 8-34 parallel jobs 9-8 HPC Server 8-21 finding, example 8-23 LSF 8-21 finding, example 8-22 PBS Pro 8-21 TORQUE 8-21 SchedulerHostname property 16-92 SecurityLevel property 16-93 ServerName property 16-95 set function 14-206 setJobSchedulerData function 14-209 setupForParallelExecution function 14-210 simplejob object 12-47 simplematlabpooljob object 12-50 simpleparalleljob object 12-53 simpletask object 12-56 single program multiple data. See spmd size function 14-212 sparse function 14-213 spmd 3-1 break 3-17 error handling 3-15
Q
QueuedFcn property 16-84
R
RcpCommand property 16-86 redistribute function 14-204 RemoteClusterAccess object 12-42 ResourceTemplate property 16-87
Index-8
Index
getting started 1-11 global variables 3-17 limitations 3-15 MATLAB path 3-15 nested functions 3-16 nested spmd 3-17 nesting with parfor 3-17 persistent variables 3-17 programming considerations 3-15 return 3-17 transparency 3-15 spmd function 14-215 StartTime property 16-96 State property 16-98 submit function 14-217 SubmitArguments property 16-101 SubmitFcn property 16-103 SubmitTime property 16-104 subsasgn function 14-218 subsref function 14-219
example 8-26 local scheduler 8-5 task object 12-58 Task property 16-106 taskFinish function 14-220 Tasks property 16-107 taskStartup function 14-221 Timeout property 16-109 TORQUE scheduler 8-21 torquescheduler object 12-61 troubleshooting programs 6-48 Type property 16-111
U
user configurations 6-14 UserData property 16-112 UserName property 16-114 UseSOAJobSubmission property 16-115
T
Tag property 16-105
W
wait function 14-222 waitForState function 14-224
task creating example 8-12 creating on generic scheduler example 8-46 creating on LSF scheduler
Windows HPC Server scheduler 8-21 worker object 12-63 Worker property 16-117 WorkerMachineOsType property 16-118
Index-9