Workload-oriented Programmingtbasten/mocc2008/presentations/MoCC2008... · Workload-oriented Programming Silviu S. Craciunas Department of Computer Sciences University of Salzburg,

Workload-oriented Programming

Silviu S. Craciunas

Department of Computer SciencesUniversity of Salzburg, Austria

-joint work with

C. Kirsch, H. Röck, and A. Sokolova

July 2008

Silviu Craciunas (U Salzburg, Austria) MoCC Artist Workshop 2008 Eindhoven, Netherlands 1 / 26

Tiptoe RTOS

TiptoePrototype real-time operating system.http://tiptoe.cs.uni-salzburg.at

Processes are sequences of actionsActions may have workload parameters that determine theamount of work involvedWe are interested in the temporal performance of actions withrespect to the workload involved


http://tiptoe.cs.uni-salzburg.at

Example

Consider a process that:allocates memory for a video streamreads the video stream from a network connectioncompresses the streamstores it on diskdeallocates the used memory


Example

Pseudo-codeloop {

int number_of_frames=determine_rate();

allocate_memory(number_of_frames);read_from_network(number_of_frames);

compress_data(number_of_frames);

write_to_disk(number_of_frames);deallocate_memory(number_of_frames);

} until (done); <– requirements


Example

Pseudo-codeloop {


allocate_memory(number_of_frames);<– requirements 1read_from_network(number_of_frames);<– requirements 2

compress_data(number_of_frames);<– requirements 3

write_to_disk(number_of_frames);<– requirements 4deallocate_memory(number_of_frames);<– requirements 5

} until (done);


Example

Pseudo-codeloop {


> allocate_memory(number_of_frames);read_from_network(number_of_frames);

compress_data(number_of_frames);

write_to_disk(number_of_frames);deallocate_memory(number_of_frames);

} until (done);


Response-time function

80

60

40

20

0

Memory al location requests in number of frames

Time (ms)

Bad

Good

0 5 10 15 20 25

f R

100


Improving latency over throughput

80

60

40

20

0


Time (ms)

Bad

Good

0 5 10 15 20 25

f R

100

8

1

allocation rate of 125 fps


Improving throughput over latency

80

60

40

20

0


Time (ms)

Bad

Good

0 5 10 15 20 25

f R

1 0 0

2 4

allocation rate of 240 fps



80

60

40

20

0


Time (ms)

Bad

Good

0 5 10 15 20 25

f R

100



InformallyThe response-time (RT) function fR characterizes the action’sresponse time bound for a given workload, independently of anyprevious or concurrent actions.

FormallyA response-time (RT) function is a discrete function

fR : N→ Q+


Execution-time function

80

60

40

20

0


Time (ms)

Bad

0 5 10 15 20 25

f R

100

fE

10



80

60

40

20

0


Time (ms)

Bad

0 5 10 15 20 25

f R

100

fE

10

f is a parametric version of WCETE



InformallyThe execution-time (ET) function fE characterizes the action’sexecution time bound for a given workload, in the absence of anyconcurrent actions.

FormallyAn execution-time (ET) function is a discrete function

fE : ED → Q+.

where ED ⊆ N is called the execution domain.


Scheduling and Admission

Process schedulingHow do we efficiently schedule processes on the level of individualprocess actions?

Process admissionHow do we efficiently test schedulability of newly arriving processes?


Virtual periodic resources


Period π Limit λ Utilizationλ

π

We use our modified version of the CBS1 which allows different π andλ for a process.

1Luca Abeni and Giorgio Buttazzo - Resource Reservation in Dynamic Real-TimeSystems, Journal of Real-Time Systems, Volume 27 (2), 2004Silviu Craciunas (U Salzburg, Austria) MoCC Artist Workshop 2008 Eindhoven, Netherlands 14 / 26


Each process declares a finite set of virtual periodic resourcesEach process action of a process uses exactly one virtual periodicresource declared by the process


Scheduled response time

The scheduled response time fS(w) is the time from the arrival of anaction until it completes.


Schedulabilty result

A set of processes is schedulable under EDF ifFor every action of every process, λ and π “sample” fR such that

∀w ∈ ED, fS(w) ≤ fR(w)

The schedulability test holds for the set of processes∑P

maxR

λPR

πPR≤ 1


Response time sampling


Tiptoe compositionality

∀fS, fS′,∀w ∈ ED

0 ≤ |fS(w)− fS′(w)| ≤ πa − 1

if the schedulability test holds


Tiptoe compositionality

Response-time jitterAny individual action a of a process maintains its response time withinπa − 1 even when adding/removing concurrent processes.


Implementation and performance

list array matrixtime O(n2) O(log(t) + n · log(t)) Θ(t)space Θ(n) Θ(t + n) Θ(t2 + n)

5 0 150 250 350 450 550 650

2 04 06 0

8 0100120

140160180

200220240260

280300 bitmap_array_max

list_maxmatrix_max

5 0 150 250 350 450 550 650

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

5.5

6

6.5

7bitmap_array_stddevlist_stddevmatrix_stddev




0 3 3 6 7 100 150 200 250 300 349

5

2 0

100

500

2000

10000

50000

200000




0 3 3 6 5 9 8 131 180 229 278 327

5

2 0

100

500

2000

10000

50000

200000

1000000




0 3 3 6 5 9 8 130 179 228 276 325

5

2 0

100

500

2000

10000

50000

200000

1000000




20

25

210

215

220

matrix

matrix-tree

array

list2

5 28 2

11 214

5KB

100KB

5MB

100MB

1GB

memory usage

KB

time instants (t)

memory usage


Conclusion

The model enablestrading off throughput and latencysequential and concurrent process compositioncontrolling the response-time variance (jitter) of individual processactions


Conclusion



Conclusion



Current work

Fly the JAviator with Tiptoe

Homepagehttp://javiator.cs.uni-salzburg.at


http://javiator.cs.uni-salzburg.at

Thank you!



∀w ∈ UD, fS(w) ≤ fR(w) + π ifλ

π= cU

π divides fR(w) evenly∑P maxR

λPRπPR≤ 1



∀w ∈ UD, fS(w) ≤ fR(w) ifλ

π= cU

0 < π ≤ dR − dEcU

π divides dR and fR(w)− dR evenly∑P maxR

λPRπPR≤ 1


Documents

Workload-oriented Programmingtbasten/mocc2008/presentations/MoCC2008... · Workload-oriented Programming Silviu S. Craciunas Department of Computer Sciences University of Salzburg,