A Power-Aware Online Scheduling Algorithm for Streaming ...patmos2010.univ-grenoble-alpes.fr/slides/patmos2010_s1p1.pdf · A Power-Aware Online Scheduling Algorithm for Streaming

A Power-Aware Online Scheduling Algorithm for Streaming

Applications in Embedded MPSoC

PATMOS 2010: 7-10 September 2010,Grenoble,France

T. Sassolas, N. Ventroux, G. Blanc

CEA LIST,Embedded Computing Laboratory

contact: [email protected]

Table of content

• Context

• Previous works

• Proposed solution

• Implementation

• Results

• Conclusion

2

Table of content

• Context

• Previous works

• Proposed solution

• Implementation

• Results

• Conclusion

3

Context of the Study• Embedded systems must support

Various application domainsmore computation intensive applicationsApplication that become more and more dynamic

• Move to multiprocessor Architectures

1 GOPS

0.1

10

100

1 TOPS

HD Audio

Multimedia

OpenGL1.1

OpenGL 2.0

H264

Digital TV

Mobile multimedia

MPEG2

3D Graphics

UMTS

EDGE

GPRS

GSM

WIMAX

3GPP-LTESDR

Telecom

DVB-S2

4

T3D2

T3 D3

T3 D1

Multiprocessor issues• Need to maximize resource usage

Increase task parallelismStreaming applicationsSet of tasks with data dependenciesScheduling of dependent tasksExecution speed determined by slowest task

5

T3D2

T3 D3

T3 D1

Worst case executionSome energy savings can be performed

T1 T2 T3in out

• Need to reduce power consumption Real case executionDynamism implies loss of energyNeed of a dynamic control

T1 D3

T2D2

T2D3

T2D4

T1 D4

T1 D5

P2

P1

P0

time

T1 D2

T2D1

T1 D2

Slack Slack SlackSlack

DVFS vs DPM• Dynamic Voltage and Frequency Scaling (DVFS)

Low mode switching penaltyReduces mainly dynamic power consumption

T1 T2

time

Power

T1T2

time

Power

6

• Dynamic power management (DPM)High energy and time switching penaltyReduces both static and dynamic power consumption

• Optimal functioning points are highly dependent on the technological process

Table of content

• Context

• Previous works

• Proposed solution

• Implementation

• Results

• Conclusion

7

Streaming schedulings: offline solutions• Scheduling on a multiprocessor is an NP complete problem [1]• Adding power optimization adds complexity

• Monoprocessor solutions [2] [3] …Find minimum power consumption given data production rate and communication buffer sizesWith DPM or DVFS functionalitiesVariable production rate following probability rule

• Multiprocessor solutionsMinimize energy consumption by finding optimum number of resources and their speed to meet QoS requirements [4]Various models : communication costs, consumption model, optimization techniques… [5]

• But regular workload was assumed: Application dynamism imposes online solutions

8

Streaming scheduling: Online solutions

• MonoprocessorSlack time reclamation: GSR [6]Offline and online partitioning

9

T1 D2

T2D2

T3 D2

T1 D2

T2D2 T3 D2

T1 D2

T2D2

T4 D2

T3D2

Resulting execution : No slack time! Buffer added

• MultiprocessorMany solutions for independent tasks -> do not applyPartitioning -> apply monoprocessor solution to every processor [7] [8]

T0

T1

T3

T2

out

in

T1 D2

T2D2 T4 D2

T3D2

time time

P1

P0 P0

P1

P0

P0

Table of content

• Context

• Previous works

• Proposed solution

• Implementation

• Results

• Conclusion

10

Power-aware streaming application scheduling

• PropertiesThroughput constrained by slowest task

» Other tasks can be slowed down to reach the same throughput -> DVFS

Task deeper in the pipeline can be blocked waiting for availabledata

» Preemption mechanisms are required for a higher resource usage rate» Unused resources can be shut down -> DPM

• Objective: keep the throughput while making substantial energy savings

11

Static Priorities

• If PE number < task number : need to specify static priorityDescribes the position in the pipelineAllows to execute oldest data first.Prevents to buffer instead of executing critical tasks

12

T0

T1

T3 T4

T2

in out

Prio = 0 Prio = 1 Prio = 2 Prio = 3

buffers monitors

Buffer full threshold : Preempt Writer

Buffer empty threshold: Preempt Reader(s)

Buffer filling threshold : reduce DVFS couple of Writer

Buffer emptying threshold: increase DVFS couple of Writer

Change QoS

Change QoS

13

• Priority impact Task is blockedTask executes at fastest speedApplication priorityTask priority

Table of content

• Context

• Previous works

• Proposed solution

• Implementation

• Results

• Conclusion

14

Consumption model• SESAM[9] simulation environment

SystemCAT-TLMIP: Noc, caches, memories…Processors ArchC ISS [10]Statistics

TurboA=1,B=1

Consumption 923 mW

Half-TurboA=1,B=2

Consumption 390 mW

Deep IdleA=0,B=1

Consumption 64 mW

1 µs1 µs

2 µs

3 µs

2 µs3 µs

15

• Modified ArchC modelsMIPS32 ISS annotatedwith PXA270 [11] PSMmode power consumptionExecution speed variationMode switching penalties

» Energy» t ime

Implemented platform

16

Centralmemory

CPUController

Processing elements

Shared memory banks

SchedulingAlgorithm

Task 1 Task 2

T1 T2in

T1T2

D1

D1

Threshold reached

Task 1

The scheduling loop

17

Update Dynamictask priorities

Buffer Statuses

Task Statuses

Order task alongwith priority

Keep already allocatedtasks on the same PE

Allocate remainingtasks on remaining PE

Update PEconsumption mode

along with buffer status

Execution / preemption demands

Mode switching demands

Table of content

• Context

• Previous works

• Proposed solution

• Implementation

• Results

• Conclusion

18

The WCDMA test case• Wideband code division multiple access application [12]

13 tasksVariable workload : pilot frame once every 10 framesIrregular pipeline task lengths

19

Results – Energy saving• 3 scheduling solutions: Standard, DPM only and DVFS + DPM• Substantial energy saving

0

10

20

30

40

50

60

70

80

90

100

1 2 4 8 13 16

Ener

gy sa

ving

(%)

PE ef

fect

ive

occu

panc

y (%

)

Number of PE

Standard dpm only dpm + dvfsenergy saving dpm only energy saving dpm+dvfs

20

Results – Execution time• No deviation in execution time

21

Results – pipeline balancing• Blocked states are reduced by the use of DVFS• More could be achieved with other DVFS couples

22

Table of content

• Context

• Previous works

• Proposed solution

• Implementation

• Results

• Conclusion

23

Conclusion

• Power reduction for variable pipeline Substantial powers saving when PE load drops : 45% on 13 processorsNo performance lossLight execution to reduce control overhead

• This work was partly funded by project SCALOPES (ARTEMIS)

• Upcoming worksImplementation on hardware multiprocessor platformEvaluation with other applications from various domainsEvaluation of optimal buffer sizes

24

References

[1] M. L. Dertouzos and A. K. Mok. Multiprocessor Online Scheduling of Hard-Real-Time Tasks. IEEE Transactions on Software Engineering, 15(12):1497-1506, 1989.[2] Y.-H. Lu, L. Benini, and G. De Micheli. Dynamic Frequency Scaling with Buffer Insertion for Mixed Workloads. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 21(5):1284-1305, 2002.[3] N. Pettis, L. Cai, and Y.-H. Lu. Statistically Optimal Dynamic Power Management for Streaming Data. IEEE Transactions on Computers, 55(7):800-814, 2006.[4] Xu, R., Melhem, R., and Mosse, D. 2007. Energy-Aware Scheduling for Streaming Applications on Chip Multiprocessors. In Proceedings of the 28th IEEE international Real-Time Systems Symposium (RTSS),pages 25-38, 2007.[5] L. Benini, D. Bertozzi, A. Guerri, and M. Milano. Allocation, Scheduling and Voltage Scaling on Energy Aware MPSoCs. In Conference on Integration of AI and OR Techniques in Constraint Programming for Combinatorial Optimization Problems(CPAIOR), pages 44-58, 2006.[6] D. Mosse, H. Aydin, B. Childers and R. Melhem. Compiler-Assisted DynamicPower-Aware Scheduling for Real-Time Applications, In Workshop on Compilers and Operating Systems for Low Power, 2000.

25

References

[7] P. Choudhury, P. P. Chakrabarti, and R. Kumar. Online Dynamic Voltage Scalingusing Task Graph Mapping Analysis for Multiprocessors. In International Conferenceon VLSI Design (VLSID), pages 89-94, 2007.[8] S. Hua, G. Qu, and S. S. Bhattacharyya. Energy-Ecient Embedded Software Implementation on Multiprocessor System-on-Chip with Multiple Voltages. ACM Transactions on Embedded Computing Systems (TECS), 5(2):321-341, 2006.[9] N. Ventroux, A. Guerre, T. Sassolas, L. Moutaoukil, C. Bechara, and R. David. SESAM: an MPSoC Simulation Environment for Dynamic Application Processing. In IEEE International Conference on Embedded Software and Systems (ICESS),2010.[10] M. Bartholomeu G. Araujo C. Araujo R. Azevedo, S. Rigo and E. Barros. The ArchC Architecture Description Language and Tools. Parallel Programming, 33(5):453–484, 2005.[11] Intel PXA27x Processor Family, Electrical, Mechanical, and Thermal Specication,2005.[12]A. Richardson. WCDMA Design Handbook. 2006.

26

Thank you for your attention

We value your opinion and questions