Real-Time SMP Scheduling

Ted Baker
Department of Computer Science
Florida State University
Tallahassee, FL 32312
[Link] baker
 Overview
1. taste of real-time scheduling theory
2. a research process
where an idea comes from
why doing research in ”backwaters” may lower stress
what to do when somebody else ”scoops” you
how to publish
what to expect from referees
3. a recent research result of mine
4. what I hope to do with the idea next
 Background: Periodic Task Model
set of tasks τ1 τn

each task has a period Ti

each task has a worst case compute time ci
each task has a relative deadline di.
ci
processor utilization of task τi Ti


c
total utilization = ∑in 1 Ti
i

 Gannt Chart of a Periodic Task’s Execution
release Tk release Tk release Tk release
ck ck




























































deadline deadline deadline
dk dk dk
 Period is Just a Lower Bound
release Tk release Tk
ck ck






























deadline dk deadline
dk
 Background: Liu & Layland EDF Utilization Bound
Theorem A set of n independent periodic tasks is schedulable by preemptive EDF
scheduling on one processor if
n c
∑ i
1
i 1 di

Q: How does this generalize for m processors?
 Bad Example for MP EDF Scheduling
τ1
τ2
.
..
τm τm+1
0 mx+1
τ1 τm
τ2
.
..
τm+1
0 mx+1
 Consequences
This example, which shows worst-case achievable processor utilization can be as
bad as 1 (compared to ideal value of m).
Everybody says EDF scheduling is no good for multiprocessors.
Everybody assumes tasks must be bound to processors in a static (or nearly static)
way, and single-processor scheduling applied to each processor.
Papers are written on how to partition tasks between processors, a bin packing
problem.
1990: My Observation about a ”Problem Window”
τk is released τk misses deadline
































t t + dk
m( d k − x) x
 1990: My Conjecture
If we have an upper bound on individual task utilizations we have a lower bound on
the worst-case achievable utilization.
Looking at the example, the worst-case achieveable utilization with EDF seems to
be close to m 1 λ where lambda maxin 1 Tci .


i



 Years Go By
I talk to Lui Sha (then CMU/SEI and now UIUC) about the idea. He doesn’t seem
to understand what I am talking about enough pick up on it.
I am still convinced it should not be too hard to prove something here.
I suggest to three different Ph.D. students that they work on the problem. They get
nowhere.
 2003: Revisiting the Problem
No longer department chair, with no current Ph.D. students, I decide to work out
the result myself.
 The Key Lemma
Lemma (upper bound on EDF load) For any busy window t t ∆ with respect to


τk the EDF load Wi ∆ due to τi is at most βi , where


ci Ti di ci
Ti 1

if λ


dk Ti
βi



ci Ti di ci λTi ci
if λ


Ti 1


dk dk Ti






 The Final Result
Theorem (EDF schedulability test) A set of periodic tasks τ1 τn is schedulable

on m processors using preemptive EDF scheduling if, for every task τk ,
n ck ck
∑ min 1 βi m1

dk dk


i 1


where β is as defined in the Lemma above.

 The Nice Corollary
Corollary A set of periodic tasks τ1 τn, all with deadline equal to period, is

guaranteed to be schedulable on m processors using preemptive EDF scheduling
if
n c
∑ i
m1 λ λ

i 1 Ti



 How I was “Scooped”
A periodic task set τ1 τ2 τn is light on m processors if:

1. ∑n
ci m2
i 1 Ti 2m 1



c m
2. Ti 2m 1 , for 1 i n.
i


Theorem (Srinivasan, Baruah[4]) Any periodic task system that is light on m pro-
cessors is scheduled to meet all deadlines on m processors by EDF.
 What I Thought I was able to Salvage
new proof technique
pre-period deadlines
2
m
more general utilization bound test: 2m 1 is just a special case of m 1 λ λ




proof that the utlization bound is tight
 The Second Scoop
Theorem (Goossens, Funk, Baruah[3]) A set of periodic tasks τ1 τn, all with

deadline equal to period, is guaranteed to be schedulable on m processors using
preemptive EDF scheduling if
n c
∑ i
m1 λ λ

i 1 Ti



where λ max ci Ti i 1 n .




They also provided a proof (like mine) that this result is “tight”.
 What I was able to Salvage
pre-period deadlines
new proof technique
decided to merge fixed-priority results into same paper
 What Referee 1 Said
“...Although the paper has some contributions to be presented .. the topic and
motivation is not that exciting. ...”
Consequence of being scooped.
 What Referee 1 Said
“Quantitative justification of the proposed analysis is required. ... more general in
the sense that it can handle preperiod deadlines. ... we can simply ... change the
original execution time C to C+(P-D) to assure P-D (D is the preperiod deadline)
earlier completion prior to the period P. Obviously, this simple modification of the
previous analysis may be much less accurate than the proposed analysis. How-
ever, how much accuracy improvement can be achieved by the proposed analysis
is questionable....”
There is an improvement, but to show it is a good idea for more research. One way
to do this is via simulation on a large randomly chosen collection of task sets.
 What Referee 1 Said
“... for some important theorems, only sketch of proof is given referring their two
technical reports. This makes readers hard to follow the theorems ...”
You can’t win on this, given the 20-page limit for papers. Putting in more proofs
means less results, and maybe an even less exciting paper.
 What Referee 2 Said
“...Given the originality of this work, I strongly recommend that this paper be ac-
cepted. ...”
 What Referee 3 Said
“... The paper is well written, and the results are of theoretical interest. ...”
 What Referee 3 Said
“... practical usage ... is limited ... unrealistic system model ... scalability and
processor cache considerations ... modern operating systems use a priority queue
per processor ... schedule the task on the processor where its previous instance
executed ... not ... the processor that is executing the lowest priority task ... ...
introduces a form of priority inversion when tasks are dynamically dispatched ...
challenging to dynamically schedule tasks in a multiprocessor in consistent priority
order ... many other factors make the assumption of perfect preemption invalid.”
A valid question. This is something we need to look into further. Clearly, there are
trade-offs involved.
 What Referee 3 Said
“...Are the bounds tight, in the sense that Liu and Layland bound is, while many
subsequent schedulability are not? Some statement on tightness of the bounds is
needed.”
We actually answered in the paper, but since it was just a few senteces and a
reference to the people who ”scooped” me, the reviewerd missed it.
 What Referee 3 Said
“...Lemma 9 is obvious. The proof obscures the result...”
You can’t please everybody on this kind of issue. Referee 1 wanted more details
on proofs.
 What I Hope to Do Next
Try to resolve Referee 1’s issue about how much is gained, and how often, by
the tighter preperiod deadline schedulability test
Try to resolve Referee 3’s issue about fixed vs. dyamic binding of tasks to
processors
1. simulate
2. implement and test, to determine real switching overheads
3. distribute implementation
Extend analysis to include blocking for mutexes
Revisit aperiodic server scheduling algorithms, in the MP context
 The Reasoning
Definition The demand of a time interval is the total amount of computation that
would need to be completed within the window for all the deadlines within the inter-
val to be met.
Definition The load of an interval t t

∆ is W ∆, where W is the demand of the




interval.
If we can find a lower bound on the load of a problem window that is necessary
for a job to miss its deadline, and we can show that a given set of tasks could not
possibly generate so much load in the problem window, that would be sufficient to
serve as a schedulability condition.
 Lower Bound on Load
τk is released τk misses deadline
































t t + dk
m( d k − x) x
Since the problem job misses its deadline, the sum of the lengths of all
the time intervals in which the problem job does not execute must exceed
its slack time, dk ck .

 Lower Bound on Load
Lemma(lower bound on load) If W dk is the load of the interval t t dk , where




t dk is a missed deadline of τk , then

W ck ck
m1

dk dk dk


 Analysis of Maximum Load
Ti Ti Ti
di di
ci ... ci ... ci
φ ε δ
t’ t head body tail t +∆
 Carried-in Load
Definition The carry-in of τi at time t is the residual compute time of the last job of
task τi released before t , if any, and is denoted by the symbol ε.
all m processors busy on other jobs of other tasks
m
ci
y x ε
φ
t’ t t +∆
 Upper Bound on EDF Demand
Lemma (EDF demand) For any busy window t t ∆ of task τk (i.e., the maxi-


mal λ-busy downward extension of a problem window) and any task τi, the EDF
demand Wi of τi in the busy window is no greater than
nci max 0 ci φλ

where φ nTi di ∆, n ∆ di Ti 1 if ∆ di, and n 0 otherwise.








Ti Ti Ti Ti
ci ci ci ci
φ Ti − φ (n−1) T i δ =d i
tail t + ∆
t’ t
head body
 Ti Ti Ti Ti
di
ci ci ... ci ci
φ Ti − φ (n−1) T i δ =c i
t’ t t +∆
head body tail
Tk Tk Tk Tk
dk
ck ck ... ck ck
φ Tk− φ (n−1) T k δ = dk
t’ t t +∆
head body tail
 References
1. C.L. Liu and J. W. Layland, “Scheduling Algorithms for Multiprogramming in a
Hard-Real-Time environment”, JACM 20.1 (January 1973) 46-61.
2. T.P. Baker, Multiprocessor EDF and Deadline Monotonic Schedulability Analy-
sis, [Link]
3. J. Goossens, S. Funk, S. Baruah, “Priority-driven scheduling of periodic task
systems on multiprocessors”, technical report UNC-CS TR01-024, University
of North Carolina Computer Science Department, Real Time Systems, Kluwer,
(to appear).
4. A. Srinivasan, S. Baruah, “Deadline-based scheduling of periodic task sys-
tems on multiprocessors”, Information Processing Letters 84 (2002) 93-98.x
5. C.A. Phillips, C. Stein, E. Torng, J Wein, “Optimal time-critical scheduling via
resource augmentation”, Proceedings of the Twenty-Ninth Annual ACM Sym-
posium on Theory of Computing (El Paso, Texas, 1997) 140-149.