Chapter 6 – Cloud Resource

Management and Scheduling

Contents
n  Resource management and scheduling.
n  Policies and mechanisms.
n  Applications of control theory to cloud resource allocation.
n  Stability of a two-level resource allocation architecture.
n  Proportional thresholding.
n  Coordinating power and performance management.
n  A utility-based model for cloud-based Web services.
n  Resource bundling and combinatorial auctions.
n  Scheduling algorithms.
n  Fair queuing.
n  Start-up fair queuing.
n  Borrowed virtual time.
n  Cloud scheduling subject to deadlines.

Dan C. Marinescu Cloud Computing: Theory and Practice. Chapter 6 2

Resource management and scheduling
n  Critical function of any man-made system.
n  It affects the three basic criteria for the evaluation of a system:
¨  Functionality.
¨  Performance.
¨  Cost.
n  Scheduling in a computing system à deciding how to allocate
resources of a system, such as CPU cycles, memory, secondary
storage space, I/O and network bandwidth, between users and
tasks.
n  Policies and mechanisms for resource allocation.
¨  Policy à principles guiding decisions.
¨  Mechanisms à the means to implement policies.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 3
Motivation
n  Cloud resource management .
¨  Requires complex policies and decisions for multi-objective
optimization.
¨  It is challenging - the complexity of the system makes it impossible to
have accurate global state information.
¨  Affected by unpredictable interactions with the environment, e.g.,
system failures, attacks.
¨  Cloud service providers are faced with large fluctuating loads which
challenge the claim of cloud elasticity.

n  The strategies for resource management for IaaS, PaaS, and SaaS
are different.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 4
Cloud resource management (CRM) policies
1.  Admission control à prevent the system from accepting workload
in violation of high-level system policies.

2.  Capacity allocation à allocate resources for individual activations

of a service.

3.  Load balancing à distribute the workload evenly among the

servers.

4.  Energy optimization à minimization of energy consumption.

5.  Quality of service (QoS) guarantees à ability to satisfy timing or

other conditions specified by a Service Level Agreement.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 5
Mechanisms for the implementation of resource
management policies

n  Control theory à uses the feedback to guarantee system stability

and predict transient behavior.

n  Machine learning à does not need a performance model of the

system.

n  Utility-based à require a performance model and a mechanism to

correlate user-level performance with cost.

n  Market-oriented/economic à do not require a model of the system,

e.g., combinatorial auctions for bundles of resources.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 6
Tradeoffs
n  To reduce cost and save energy we may need to concentrate the
load on fewer servers rather than balance the load among them.
n  We may also need to operate at a lower clock rate; the
performance decreases at a lower rate than does the energy.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 7
Control theory application to cloud resource
management (CRM)

n  The main components of a control system:

¨  The inputs à the offered workload and the policies for admission
control, the capacity allocation, the load balancing, the energy
optimization, and the QoS guarantees in the cloud.

¨  The control system components à sensors used to estimate relevant

measures of performance and controllers which implement various
policies.

¨  The outputs à the resource allocations to the individual applications.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 8
Feedback and Stability
n  Control granularity à the level of detail of the information used to
control the system.
¨  Fine control à very detailed information about the parameters
controlling the system state is used.
¨  Coarse control à the accuracy of these parameters is traded for the
efficiency of implementation.
n  The controllers use the feedback provided by sensors to stabilize
the system. Stability is related to the change of the output.
n  Sources of instability in any control system:
¨  The delay in getting the system reaction after a control action.
¨  The granularity of the control, the fact that a small change enacted by
the controllers leads to very large changes of the output.
¨  Oscillations, when the changes of the input are too large and the
control is too weak, such that the changes of the input propagate
directly to the output.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 9
The structure of a cloud controller

disturbance

r s λ (k )
u* (k)
Predictive Optimal Queuing
filter controller dynamics
external forecast ω (k )
traffic

state feedback q(k)

The controller uses the feedback regarding the current state and the estimation
of the future disturbance due to environment to compute the optimal inputs
over a finite horizon. r and s are the weighting factors of the performance
index.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 10
Two-level cloud controller

Application 1 Application n
Application 1 SLA 1 …. SLA n Application n
VM VM
…. VM VM

Application …. Application
controller controller
Monitor Monitor

Decision …. Decision

Cloud Controller Actuator Actuator

Cloud Platform

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 11
Lessons from the two-level experiment

n  The actions of the control system should be carried out in a rhythm
that does not lead to instability.
n  Adjustments should only be carried out after the performance of the
system has stabilized.
n  If upper and a lower thresholds are set, then instability occurs when
they are too close to one another if the variations of the workload
are large enough and the time required to adapt does not allow the
system to stabilize.
n  The actions consist of allocation/deallocation of one or more virtual
machines. Sometimes allocation/dealocation of a single VM
required by one of the threshold may cause crossing of the other,
another source of instability.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 12
Control theory application to CRM
n  Regulate the key operating parameters of the system based on
measurement of the system output.
n  The feedback control assumes a linear time-invariant system
model, and a closed-loop controller.
n  The system transfer function satisfies stability and sensitivity
constraints.
n  A threshold à the value of a parameter related to the state of a
system that triggers a change in the system behavior.
n  Thresholds à used to keep critical parameters of a system in a
predefined range.
n  Two types of policies:
1.  threshold-based à upper and lower bounds on performance trigger
adaptation through resource reallocation; such policies are simple
and intuitive but require setting per-application thresholds.
2.  sequential decision à based on Markovian decision models.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 13
Design decisions

n  Is it beneficial to have two types of controllers:

¨  application controllers à determine if additional resources are needed.
¨  cloud controllers à arbitrate requests for resources and allocates the
physical resources.

n  Choose fine versus coarse control.

n  Dynamic thresholds based on time averages better versus static

ones.

n  Use a high and a low threshold versus a high threshold only.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 14
Proportional thresholding

n  Algorithm
¨  Compute the integral value of the high and the low threshold as
averages of the maximum and, respectively, the minimum of the
processor utilization over the process history.
¨  Request additional VMs when the average value of the CPU utilization
over the current time slice exceeds the high threshold.
¨  Release a VM when the average value of the CPU utilization over the
current time slice falls below the low threshold.

n  Conclusions
¨  Dynamic thresholds perform better than the static ones.
¨  Two thresholds are better than one.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 15
Coordinating power and performance management

n  Use separate controllers/managers for the two objectives.

n  Identify a minimal set of parameters to be exchanged between the
two managers.
n  Use a joint utility function for power and performance.
n  Set up a power cap for individual systems based on the utility
-optimized power management policy.
n  Use a standard performance manager modified only to accept
input from the power manager regarding the frequency
determined according to the power management policy.
n  Use standard software systems.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 16
 Communication between
autonomous managers

Performance manager Power manager

Control policy Power Control policy

Performance
data data

Blade

Blade

Workload Workload Blade Power

Power
generator distribution assignment
Blade

Blade

Autonomous performance and power managers cooperate to

ensure prescribed performance and energy optimization; they
are fed with performance and power data and implement the
performance and power management policies
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 17
A utility-based model for cloud-based web services

n  A service level agreement (SLA) à specifies the rewards as well

as penalties associated with specific performance metrics.
n  The SLA for cloud-based web services uses the average response
time to reflect the Quality of Service.
n  We assume a cloud providing K different classes of service, each
class k involving Nk applications.
n  The system is modeled as a network of queues with multi-queues
for each server.
n  A delay center models the think time of the user after the
completion of service at one server and the start of processing at
the next server.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 18
 Utility function U(R)

P
e
n
a
t
l
t
y

0 R0 R1 R2
R R - response time
e
w
a
r
d

The utility function U(R) is a series of step functions with jumps

corresponding to the response time, R=R0 | R1 | R2, when the
reward and the penalty levels change according to the SLA. The
dotted line shows a quadratic approximation of the utility function.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 19
 Si2

Si3

Si1

Si6
Si4
vk
vkmax

Si5

rk (b)
rkmax
(a)

(a) The utility function: vk the revenue (or the penalty)

function of the response time rk for a request of class k.
(b) A network of multiqueues.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 20
The model requires a large number of parameters

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 21
Resource bundling
u  Resources in a cloud are allocated in bundles.

u  Users get maximum benefit from a specific combination of resources:

CPU cycles, main memory, disk space, network bandwidth, and so
on.

u  Resource bundling complicates traditional resource allocation models

and has generated an interest in economic models and, in particular,
in auction algorithms.

u  The bidding process aims to optimize an objective function f(x,p).

u  In the context of cloud computing, an auction is the allocation of

resources to the highest bidder.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 22
Combinatorial auctions for cloud resources
u  Simultaneous clock auction, clock proxy auction, and ascending
clock auction (ASCA) are 3 known combinatorial auctions.

u  Users provide bids for desirable bundles and the price they are
willing to pay.

u  Prices and allocation are set as a result of an auction.

u  Ascending Clock Auction, (ASCA) à the current price for each
resource is represented by a “clock” seen by all participants at the
auction.

u  The algorithm involves user bidding in multiple rounds; to address

this problem the user proxies automatically adjust their demands
on behalf of the actual bidders.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 23
 u1 Proxy x1(t)

u2 Proxy x2(t)

x3(t)
u3 Proxy Auctioneer

∑x
u
u (t ) > 0

uU Proxy xU(t)

p(t+1)

The schematics of the ASCA algorithm; to allow for a single round auction
users are represented by proxies which place the bids xu(t). The auctioneer
determines if there is an excess demand and, in that case, it raises the price of
resources for which the demand exceeds the supply and requests new bids.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 24
 Pricing and allocation algorithms
n  A pricing and allocation algorithm partitions the set of users in two
disjoint sets, winners and losers.
n  Desirable properties of a pricing algorithm:
¨  Be computationally tractable; traditional combinatorial auction algorithms
e.g., Vickey-Clarke-Groves (VLG) are not computationally tractable.
¨  Scale well - given the scale of the system and the number of requests
for service, scalability is a necessary condition.
¨  Be objective - partitioning in winners and losers should only be based on
the price of a user's bid; if the price exceeds the threshold then the user
is a winner, otherwise the user is a loser.
¨  Be fair - make sure that the prices are uniform, all winners within a
given resource pool pay the same price.
¨  Indicate clearly at the end of the auction the unit prices for each
resource pool.
¨  Indicate clearly to all participants the relationship between the supply
and the demand in the system.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 25
 Cloud scheduling algorithms (1/2)
n  Scheduling à responsible for resource sharing at several levels:
¨  A server can be shared among several virtual machines.
¨  A virtual machine could support several applications.
¨  An application may consist of multiple threads.
n  A scheduling algorithm should be efficient, fair, and starvation-free.
n  The objectives of a scheduler:
¨  Batch system à maximize throughput and minimize turnaround time.
¨  Real-time system à meet the deadlines and be predictable.
n  Best-effort: batch applications and analytics.
n  Common algorithms for best effort applications:
¨  Round-robin.
¨  First-Come-First-Serve (FCFS).
¨  Shortest-Job-First (SJF).
¨  Priority algorithms.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 26
 Cloud scheduling algorithms (2/2)
n  Multimedia applications (e.g., audio and video streaming)
¨  Have soft real-time constraints.
¨  Require statistically guaranteed maximum delay and throughput.
n  Real-time applications have hard real-time constraints.
n  Scheduling algorithms for real-time applications:
¨  Earliest Deadline First (EDF).
¨  Rate Monotonic Algorithms (RMA).
n  Algorithms for integrated scheduling of several classes of
applications (best-effort, multimedia, real-time):
¨  Resource Allocation/Dispatching (RAD) .
¨  Rate-Based Earliest Deadline (RBED).

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 27
 Quantity

Hard real-time
strict
Hard-requirements

Soft-requirements

Best-effort

Timing
loose

loose strict

Best-effort policies à do not impose requirements regarding either the amount of

resources allocated to an application, or the timing when an application is scheduled.
Soft-requirements policies à require statistically guaranteed amounts and
timing constraints
Hard-requirements policies à demand strict timing and precise amounts of resources.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 28
Fair queuing - schedule multiple flows through a switch
R(t) R(t)

Fai(tai)=Sai(tai)+Pai Fai(tai)=Sai(tai)+Pai

Sai(tai)=Rai(tai) Sai(tai)=Fai-1(tai-1)

Fai-1(tai-1) Rai(tai)

tai-1 tai tai-1 tai

(a) (b)

The transmission of packet i of a flow can only start after the packet is
available and the transmission of the previous packet has finished.
(a) The new packet arrives after the previous has finished.
(b) The new packet arrives before the previous one was finished.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 29
Start-time fair queuing

n  Organize the consumers of the CPU bandwidth in a tree structure.

n  The root node is the processor and the leaves of this tree are the
threads of each application.
¨  When a virtual machine is not active, its bandwidth is reallocated to the
other VMs active at the time.
¨  When one of the applications of a virtual machine is not active, its
allocation is transferred to the other applications running on the same VM.
¨  If one of the threads of an application is not runnable then its allocation is
transferred to the other threads of the applications.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 30
 Server

VM1 VM2
(1) (3)

A1 A2 A3
(3) (1) (1)

t1,1 t1,2 t1,3 t2 vs1 vs2 vs3

(1) (1) (1) (1) (1) (1) (1)

The SFQ tree for scheduling when two virtual machines VM1 and VM2 run
on a powerful server

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 31
 Thread  a

0 12 24 36
t
12 24 36 48

thread  b  is  suspended thread  b  is  reactivated

Thread  b

0 3 6 9 12

3 6 9 12 15

Virtual  time

36

24

12
6 Real  time

0 3 15 18 21 24 36 48 60

Top à the virtual startup time and the virtual finish time and function of the real time t
for each activation of threads a and b.
Bottom à the virtual time of the scheduler v(t) function of the real time
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 32
Borrowed virtual time (BVT)
n  Objective - support low-latency dispatching of real-time applications,
and weighted sharing of CPU among several classes of applications.
n  A thread i has
¨  an effective virtual time, Ei.
¨  an actual virtual time, Ai.
¨  a virtual time warp, Wi.
n  The scheduler thread maintains its own scheduler virtual time (SVT)
defined as the minimum actual virtual time of any thread.
n  The threads are dispatched in the order of their effective virtual time,
policy called the Earliest Virtual Time (EVT).
n  Context switches are triggered by events such as:
¨  the running thread is blocked waiting for an event to occur.
¨  the time quantum expires.
¨  an interrupt occurs.
¨  when a thread becomes runnable after sleeping.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 33
 Effective  virtual  time
450

390

360

300

270

210

180

120

90

30
Real  time  (mcu)

2 5 11 14 20 23 29 32 38 41
9 18 27 36 45
The effective virtual time and the real time of the threads a
(solid line) and b (dotted line) with weights wa = 2 wb when
the actual virtual time is incremented in steps of 90 mcu.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 34
 Effective  virtual  time
450

390

360

300

270

210

180

120

90

30
Real  time  (mcu)

2 5 12 14 21 23 30 32 39 41
9 18 27 36 45

-­‐60

The effective virtual time and the real time of the threads a (solid line), b (dotted line),
and the c with real-time constraints (thick solid line). Thread c wakes up periodically
at times t=9, 18, 27, 36,…, is active for 3 units of time and has a time warp of 60 mcu.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 35
Cloud scheduling subject to deadlines

n  Hard deadlines à if the task is not completed by the deadline, other
tasks which depend on it may be affected and there are penalties; a
hard deadline is strict and expressed precisely as milliseconds, or
possibly seconds.

n  Soft deadlines à more of a guideline and, in general, there are no

penalties; soft deadlines can be missed by fractions of the units
used to express them, e.g., minutes if the deadline is expressed in
hours, or hours if the deadlines is expressed in days. (cloud
schedules are usually in this category)

n  We consider only aperiodic tasks with arbitrarily divisible workloads.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 36
 System Model

n  Aperiodic tasks with arbitrarily divisible workloads only.

n  The application runs on a partition of a cloud, a virtual cloud with a
head node and n worker nodes.
n  The system is homogeneous, all workers are identical, and the
communication time from the head node to any worker node is the
same.
n  The problems to be resolved are:
1.  The order of execution of the tasks.
2.  The workload partitioning and task mapping to the worker nodes.
u  The most common scheduling policies to determine the order of
execution of the tasks are:
1.  FIFO
2.  Earliest deadline first (EDF)
3.  Maximum workload derivative first (MWF)
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 37
 Scheduling Policies

n  First in, First out (FIFO) à The tasks are scheduled for execution
in the order of their arrival.

n  Earliest deadline first (EDF) à The task with the earliest deadline
is scheduled first.

n  Maximum workload derivative first (MWF) à The tasks are

scheduled in the order of their derivatives, the one with the highest
derivative first. The number n of nodes assigned to the application
is kept to a minimum.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 38
 Workload Partitioning Rules

n  Optimal Partitioning Rule (OPR) à the workload is partitioned to

ensure the earliest possible completion time and all tasks are
required to complete at the same time.
Ø  The head node distributes sequentially the data to individual worker nodes.
Ø  Worker nodes start processing the data as soon as the transfer is complete.

n  Equal Partitioning Rule (EPR) à assigns an equal workload to

individual worker nodes.
Ø  The head node distributes sequentially the data to individual worker nodes.
Ø  Worker nodes start processing the data as soon as the transfer is complete.
Ø  The workload is partitioned in equal segments.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 39
 S0 Γ1 Γ2 Γ3 Γn

S1 Δ1

S2 Δ2

S3 Δ3

Sn Δn

The timing diagram for the Optimal Partitioning Rule; the algorithm requires
worker nodes to complete execution at the same time. The head node, S0,
distributes sequentially the data to individual worker nodes.
Cloud Computing: Theory and Practice.
Dan C. Marinescu Chapter 6 40
 S0 Γ1 Γ2 Γ3 Γn

S1 Δ1

S2 Δ2

S3 Δ3

Sn Δn

The timing diagram for the Equal Partitioning Rule; the algorithm assigns an equal
workload to individual worker nodes.

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 41
 Scheduling MapReduce Applications
subject to deadlines

Four commonly referenced scheduling are:

n  The default FIFO schedule
n  The Fair Scheduler
n  The Capacity Scheduler
n  The Dynamic Proportional Scheduler

They are all based upon the assumptions:

1.  The system is homogeneous
2.  Load equipartition

Cloud Computing: Theory and Practice.

Dan C. Marinescu Chapter 6 42