Understanding Energy Behaviors of Thread Management Constructs

Transcript

1. Understanding Energy Behaviors
   of Thread Management Constructs
   Gustavo Pinto1 Fernando Castor1 David Liu2
   {ghlp, castor}@cin.ufpe.br1
   davidL@binghamton.edu2
   

   View Slide

2. 2
   • First, the proliferation of multicore CPUs
   • Second, the prevalence of multi-threaded programs
   Motivation
   

   View Slide

3. 3
   • First, more cores more power consumed
   • Second, little is known about energy behaviors of
   multi-threaded program on the application and
   programming language level
   The Problem
   

   View Slide

4. 4
   1. programming abstractions of thread management
   on energy efficiency
   2. programmer choices of thread management on
   energy efficiency
   This Talk
   for Java multi-threaded programs
   

   View Slide

5. 5
   • Explicit threading (the Thread-style): Using the
   java.lang.Thread class
   • Thread pooling (the Executor-style): Using the
   java.util.concurrent.Executor framework
   • Working Stealing (the ForkJoin-style): Using the
   java.util.concurrent.ForkJoin framework
   Thread management
   constructs
   

   View Slide

6. 6
   • Embarrassingly parallel: spectralnorm, sunflow,
   n-queens
   • Leaning parallel: xalan, knucleotide, tomcat
   • Leaning serial: mandelbrot, largestImage
   • Embarrassingly serial: h2
   Benchmarks
   

   View Slide

7. 7
   • Embarrassingly parallel: spectralnorm, sunflow,
   n-queens
   • Leaning parallel: xalan, knucleotide, tomcat
   • Leaning serial: mandelbrot, largestImage
   • Embarrassingly serial: h2
   Benchmarks
   Micro-benchmarks
   DaCapo benchmarks
   

   View Slide

8. 8
   Experimental Environment
   A 2×16-core AMD CPUs, running Debian
   Linux, 64GB of memory, JDK version 1.7.0
   11, build 21, “ondemand” governor
   

   View Slide

9. 9
   Experimental Environment
   A 2×16-core AMD CPUs, running Debian
   Linux, 64GB of memory, JDK version 1.7.0
   11, build 21, “ondemand” governor
   

   View Slide

10. 10
    Experimental Environment
    A 2×16-core AMD CPUs, running Debian
    Linux, 64GB of memory, JDK version 1.7.0
    11, build 21.
    

    View Slide

11. 11
    Experimental Environment
    A 2×16-core AMD CPUs, running Debian
    Linux, 64GB of DDR3 1600 memory, and
    JDK version 1.7.0 11, build 21.
    

    View Slide

12. Energy Consumption When Varying the
    Number of Threads
    12
    

    View Slide

13. 13
    The Λ Curve
    

    View Slide

14. 14
    The Λ Curve
    

    View Slide

15. 15
    The Λ Curve
    

    View Slide

16. 16
    More cores idle
    CPU frequency
    at a lower level
    The Λ Curve
    

    View Slide

17. 17
    More cores idle
    CPU frequency
    at a lower level
    More threads used,
    program completes sooner
    The greater the ratio
    between speedup and
    power, the steeper the \
    The Λ Curve
    

    View Slide

18. Which programming style should I use?
    18
    ?
    ?
    

    View Slide

19. Overpopulating Cores with Threads
    19
    

    View Slide

20. Overpopulating Cores with Threads
    20
    

    View Slide

21. Faster ≠ Greener
    21
    

    View Slide

22. Copying vs Sharing
    22
    

    View Slide

23. Copying vs Sharing
    23
    Copying
    

    View Slide

24. Copying vs Sharing
    24
    Copying
    Sharing
    

    View Slide

25. Copying vs Sharing
    25
    Copying
    Sharing
    ±15% of energy savings!
    

    View Slide

26. 26
    Copy-Fork
    

    View Slide

27. 27
    After
    Before
    Copy-Fork
    

    View Slide

28. 28
    After
    Before
    Copy-Fork
    Copy/Fork/Copy/Fork/…
    

    View Slide

29. 29
    After
    Before
    Copy-Fork
    Copy/Fork/Copy/Fork/…
    

    View Slide

30. 30
    After
    Before
    Copy-Fork
    Copy/Fork/Copy/Fork/…
    Copy/../Copy/Fork/…/Fork
    

    View Slide

31. 31
    After
    Before
    Copy-Fork
    ±10% of energy
    savings!
    Copy/Fork/Copy/Fork/…
    Copy/../Copy/Fork/…/Fork
    

    View Slide

32. Data Size
    32
    

    View Slide

33. Data Size
    33
    For most of the cases, energy
    consumption is linear to data
    size!
    

    View Slide

34. View Slide

35. View Slide

36. View Slide

37. View Slide

38. Understanding Energy Behaviors
    of Thread Management Constructs
    Gustavo Pinto1 Fernando Castor1 David Liu2
    {ghlp, castor}@cin.ufpe.br1
    davidL@binghamton.edu2
    

    View Slide