Accelerating a local search algorithm for large instances of the - - PowerPoint PPT Presentation

▶
accelerating a local search algorithm for large instances
SMART_READER_LITE
LIVE PREVIEW

Accelerating a local search algorithm for large instances of the - - PowerPoint PPT Presentation

Jan 28, 2023 168 likes •370 views

Motivation Initial algorithm Adaptation Large instances Conclusion Accelerating a local search algorithm for large instances of the independent task scheduling problem with the GPU Frederic Pinel Johnatan Pecero Pascal Bouvry Computer

slide-1
SLIDE 1

Motivation Initial algorithm Adaptation Large instances Conclusion

Accelerating a local search algorithm for large instances of the independent task scheduling problem with the GPU

Frederic Pinel Johnatan Pecero Pascal Bouvry

Computer Science and Communications University of Luxembourg

GreenDays@Paris

slide-2
SLIDE 2

Motivation Initial algorithm Adaptation Large instances Conclusion

Outline

Motivation Initial algorithm Adaptation Large instances Conclusion

slide-3
SLIDE 3

Motivation Initial algorithm Adaptation Large instances Conclusion

Motivation

  • Independent tasks
  • Makespan, combines several perspectives:
  • User: flowtime
  • Provider: load balance, energy (low machine heterogeneity)
slide-4
SLIDE 4

Motivation Initial algorithm Adaptation Large instances Conclusion

Parallel CGA

Selection Recombination Mutation Replacement

Figure: Generating solution

slide-5
SLIDE 5

Motivation Initial algorithm Adaptation Large instances Conclusion

Parallel CGA

  • Parallel asynchronous cellular genetic algorithm
  • Initialized with heuristic (Min-Min)
  • Local search
slide-6
SLIDE 6

Motivation Initial algorithm Adaptation Large instances Conclusion

Feedback

Pop P_mutation Iter_mutation P_crossover P_search Iter_search Load_search Threads 0.0 0.2 0.4 0.6 0.8 1.0 main effect interactions

slide-7
SLIDE 7

Motivation Initial algorithm Adaptation Large instances Conclusion

Adaptation

  • Simplified algorithm
  • Min-Min, incremental formulation
  • Increased local search, complete-state formulation
slide-8
SLIDE 8

Motivation Initial algorithm Adaptation Large instances Conclusion

Results

700 800 900 1000 1100 1200 1300 Min-Min 2PH PA-CGA-1 PA-CGA-2 PACGA-3 PACGA-4 PACGA-5

Figure: Consistent, high-h. tasks, low-h. machines

slide-9
SLIDE 9

Motivation Initial algorithm Adaptation Large instances Conclusion

Results

650 700 750 800 850 900 950 1000 1050 1100 1150 Min-Min 2PH PA-CGA-1 PA-CGA-2 PA-CGA-3 PA-CGA-4 PA-CGA-5

Figure: Semi-consistent, high-h. tasks, low-h. machines

slide-10
SLIDE 10

Motivation Initial algorithm Adaptation Large instances Conclusion

Results

760 770 780 790 800 810 820 830 840 850 860 Min-Min 2PH PA-CGA-1 PA-CGA-2 PA-CGA-3 PA-CGA-4 PA-CGA-5

Figure: Consistent, low-h. tasks, low-h. machines

slide-11
SLIDE 11

Motivation Initial algorithm Adaptation Large instances Conclusion

Results

690 700 710 720 730 740 750 760 770 Min-Min 2PH PA-CGA-1 PA-CGA-2 PA-CGA-3 PA-CGA-4 PA-CGA-5

Figure: Semi-consistent, low-h. tasks, low-h. machines

slide-12
SLIDE 12

Motivation Initial algorithm Adaptation Large instances Conclusion

Min-Min on GPU

Figure: Parallel reduction in Min-Min

slide-13
SLIDE 13

Motivation Initial algorithm Adaptation Large instances Conclusion

Min-Min runtime

1 10 100 1000 10000 100000 1e+06 4096x128 8192x256 16384x512 32768x1024 65536x2048 Runtime (sec) Instance size CPU 1 thread CPU 4 threads CPU 8 threads GPU

Figure: Runtime Min-Min

slide-14
SLIDE 14

Motivation Initial algorithm Adaptation Large instances Conclusion

Performance

1650 1700 1750 1800 1850 1900 1950 2000 2050 512x16 4096x128 8192x256 16384x512 32768x1024 65536x2048 Makespan Instance size 32 320 3200 CGA

Figure: Makespan

slide-15
SLIDE 15

Motivation Initial algorithm Adaptation Large instances Conclusion

Performance

100 200 300 400 500 600 512x16 4096x128 8192x256 16384x512 32768x1024 65536x2048 Runtime (sec) Instance size 32 320 3200 CGA - Kgen

Figure: Runtime

slide-16
SLIDE 16

Motivation Initial algorithm Adaptation Large instances Conclusion

Conclusion

  • Failure
  • Solution → Feedback → loop
  • Learning process
slide-17
SLIDE 17

Motivation Initial algorithm Adaptation Large instances Conclusion

Machine learning opportunities

  • Learn on problem instance
  • Task profiling
  • Co-scheduling
  • Learn allocation rules
  • Adapt (parameters, heuristics)
  • Algorithm (oracle: solved instances)
slide-18
SLIDE 18

Motivation Initial algorithm Adaptation Large instances Conclusion

Questions

Thank you.