COORDINATION'16

Transcript

1. PART: AN ASYNCHRONOUS PARALLEL ABSTRACTION FOR SPECULATIVE PIPELINE COMPUTATIONS Kiko

   Fernandez-Reyes, Dave Clarke and Daniel S.McCain Uppsala University, Sweden

2. PART: AN ASYNCHRONOUS PARALLEL ABSTRACTION FOR SPECULATIVE PIPELINE COMPUTATIONS Asynchronous:

   Don’t stop the world

3. PART: AN ASYNCHRONOUS PARALLEL ABSTRACTION FOR SPECULATIVE PIPELINE COMPUTATIONS Asynchronous:

   Don’t stop the world Parallel: manycore era

4. PART: AN ASYNCHRONOUS PARALLEL ABSTRACTION FOR SPECULATIVE PIPELINE COMPUTATIONS Asynchronous:

   Don’t stop the world Parallel Abstraction: manycore era + parallel collection

5. PART: AN ASYNCHRONOUS PARALLEL ABSTRACTION FOR SPECULATIVE PIPELINE COMPUTATIONS Asynchronous:

   Don’t stop the world Parallel Abstraction: manycore era + parallel collection Speculative: kill-off computations Better use for CPU

6. PART: AN ASYNCHRONOUS PARALLEL ABSTRACTION FOR SPECULATIVE PIPELINE COMPUTATIONS Asynchronous:

   Don’t stop the world Parallel Abstraction: manycore era + parallel collection Speculative: kill-off computations Better use of CPU Pipeline: highly parallelizable

7. Motivation Encore Programming Language: • Asynchronous • Parallel • Futures

   as synchronisation point 1

8. Motivation Encore Programming Language: • Asynchronous • Parallel • Futures

   as synchronisation point Actors & Tasks 1

9. Tasks calculateF ƒ1 get ƒ1 (Block until future is fulfilled)

   return Maybe(bool) 1. 2. 3. 4. Steps: 2 let fut : Fut(Maybe bool) = async { calculateF(f, a); } in return get(fut) (get :: Fut t → t)

10. Tasks 3 let fut : Maybe bool = async {

    calculateF(f, a); } in fut ⤳ (\(m: Maybe bool) -> match m with Just true => a …) calculateF ƒ1 1. 2. Steps: ƒ1 ⤳ (\(m: Maybe bool) -> match m with … ƒ2 = ⤳ :: Fut t → (t → t’) → Fut t’

11. Motivation Encore Programming Language: • Asynchronous • Parallel • Futures

    as synchronisation point Actors & Tasks 4

12. Motivation Encore Programming Language: • Asynchronous • Parallel • Futures

    as synchronisation point Actors & Tasks Actors and tasks lack of an abstraction for the asynchronous and parallel nature of their computations 4

13. Motivation 5 Parallel Portfolio-based SAT solver Goal: is a Formula

    satisfiable or not? Input: Array of Strategies, Formula Output: Maybe Assignment

14. Motivation 6 Parallel Portfolio-based SAT solver foreach s: strategies async

    solve(formula, s, new Assignment())

15. Motivation 6 Parallel Portfolio-based SAT solver foreach s: strategies async

    solve(formula, s, new Assignment()) Futures

16. Motivation 6 Parallel Portfolio-based SAT solver foreach s: strategies async

    solve(formula, s, new Assignment()) Futures Asynchronously continue running

17. Motivation 6 Parallel Portfolio-based SAT solver foreach s: strategies async

    solve(formula, s, new Assignment()) Futures Asynchronously continue running How can I select the first one that finishes?

18. Motivation 7 Parallel Portfolio-based SAT solver solve solve solve

19. Motivation 7 Parallel Portfolio-based SAT solver solve solve solve Strategy

    1. Take Variable 2. Split variable 3. Evaluate formula

20. Motivation 7 Parallel Portfolio-based SAT solver solve solve solve Strategy

    1. Take Variable 2. Split variable 3. Evaluate formula async

21. Motivation 7 Parallel Portfolio-based SAT solver solve solve solve Strategy

    1. Take Variable 2. Split variable 3. Evaluate formula async async

22. Motivation 7 Parallel Portfolio-based SAT solver solve solve solve Strategy

    1. Take Variable 2. Split variable 3. Evaluate formula EvalFormula async async

23. Motivation 7 Parallel Portfolio-based SAT solver solve EvalFormula solve solve

    Strategy 1. Take Variable 2. Split variable 3. Evaluate formula EvalFormula async async

24. Motivation 7 Parallel Portfolio-based SAT solver solve EvalFormula solve solve

    Strategy 1. Take Variable 2. Split variable 3. Evaluate formula EvalFormula ⤳ satisfiable? async async

25. Motivation 8 EvalFormula solve solve solve Strategy 1. Take Variable

    2. Split variable 3. Evaluate formula EvalFormula ⤳ satisfiable? async async Parallel Portfolio-based SAT solver ⤳ satisfiable?

26. Motivation 8 EvalFormula solve solve solve Strategy 1. Take Variable

    2. Split variable 3. Evaluate formula EvalFormula ⤳ satisfiable? Don’t know solve async async Parallel Portfolio-based SAT solver ⤳ satisfiable?

27. Motivation 8 EvalFormula solve solve solve Strategy 1. Take Variable

    2. Split variable 3. Evaluate formula EvalFormula False ⤳ satisfiable? Don’t know solve async async Parallel Portfolio-based SAT solver ⤳ satisfiable?

28. Motivation 9

29. Motivation 9

30. Motivation 9

31. Motivation 9

32. Motivation 9 value ƒ2 ƒ1 ƒ3 ƒ2

33. Motivation 10 value ƒ2 ƒ1 ƒ3 ƒ2

34. Motivation 10 value ƒ2 ƒ1 ƒ3 ƒ2

35. Motivation 10 value ƒ2 ƒ1 ƒ3 ƒ2

36. Motivation 10 value ƒ2 ƒ1 ƒ3 ƒ2 ƒ1

37. Motivation 10 value ƒ2 ƒ1 ƒ3 ƒ2 ƒ1 true

38. Motivation 10 value ƒ2 ƒ1 ƒ3 ƒ2 ƒ1 true

39. Motivation 11 ƒ1 ƒ3 ƒ1 true value ƒ2 ƒ2

40. Motivation 11 ƒ1 ƒ3 ƒ1 true

41. Motivation 11 ƒ1 ƒ1 true

42. Contributions • Parallel abstraction for handling complex coordination patterns •

    Asynchronous Combinators • Kills-off speculative computations when no longer needed • Prototypes in Encore and Clojure 12

43. Index • ParT & Combinators • Example: Parallel Portfolio-based SAT

    solver • Implementation: under the hood • Future work & Limitations • Conclusion 19

44. ParT & Combinators 13 Functional data structure Encapsulates parallel &

    asynchronous computations Controlled via parallel combinators derived from Orc… typed & asynchronous async computation async computation async computation ParT

45. ParT & Combinators 13 Functional data structure Encapsulates parallel &

    asynchronous computations Controlled via parallel combinators derived from Orc… typed & asynchronous async computation async computation async computation ParT

46. ParT & Combinators 13 Functional data structure Encapsulates parallel &

    asynchronous computations Controlled via parallel combinators derived from Orc… typed & asynchronous async computation async computation async computation ParT Handle to ongoing parallel computations

47. Example • SAT solver • Multiple strategies • Pipeline and

    Speculative parallelism 14

48. Example: SAT solver Formula [Strategies] 15

49. Example: SAT solver Formula [Strategies] s1 s2 sn 15 each

50. Example: SAT solver Formula [Strategies] s1 s2 sn solve 15

    each

51. Example: SAT solver Formula [Strategies] s1 s2 sn solve Strategy

    1. Take variable 2. Split variable 15 each

52. Example: SAT solver olve Strategy 1. Take variable 2. Split

    variable 16

53. Example: SAT solver olve Strategy 1. Take variable 2. Split

    variable evalFormula evalFormula 16 ||

54. Example: SAT solver olve Strategy 1. Take variable 2. Split

    variable evalFormula evalFormula 16 || doesn’t spawn new asynchronous computations! It rather declares there is possibility for parallelism ||

55. Example: SAT solver olve Strategy 1. Take variable 2. Split

    variable evalFormula evalFormula 16 || doesn’t spawn new asynchronous computations! It rather declares there is possibility for parallelism || >>= >>= satisfiable? satisfiable?

56. Example: SAT solver Formula [Strategies] s1 s2 sn 17

57. Example: SAT solver Formula [Strategies] s1 s2 sn formula not

    solved 17

58. Example: SAT solver Formula [Strategies] s1 s2 sn formula not

    solved false x 17

59. Example: SAT solver Formula [Strategies] s1 s2 sn formula not

    solved false true x 17

60. Example: SAT solver Formula [Strategies] s1 s2 sn true 18

61. Example: SAT solver Formula [Strategies] s1 s2 sn true 18

62. Example: SAT solver Formula [Strategies] s1 s2 sn true 18

63. Example: SAT solver Formula [Strategies] s1 s2 sn true 18

64. Example: SAT solver Formula [Strategies] s1 s2 sn true 18

65. Example: SAT solver Formula [Strategies] s1 s2 sn true 19

66. Example: SAT solver Formula [Strategies] s1 s2 sn true 19

67. Example: SAT solver Formula [Strategies] 20

68. Example: SAT solver Formula [Strategies] 20 ParT

69. Example: SAT solver Formula [Strategies] 20 ParT

70. Example: SAT solver Formula [Strategies] << 20 ParT

71. Example: SAT solver Formula [Strategies] << Computation 20 ParT

72. Example: SAT solver Formula [Strategies] << Computation Parallel and Asynchronous

    computations captured in the ParT abstraction This computation can start immediately 21

73. ParT Example: SAT solver Formula [Strategies] << Computation Parallel and

    Asynchronous computations captured in the ParT abstraction This computation can start immediately 21

74. Example: SAT solver << fut2Par << (each(sts) >>= \(s: Strategy)

    -> (async sat(s, fml, new Assignment())✝)) 22 ParT Computation

75. Example: SAT solver << fut2Par << (each(sts) >>= \(s: Strategy)

    -> (async sat(s, fml, new Assignment())✝)) 22 ParT Computation

76. Example: SAT solver << fut2Par << (each(sts) >>= \(s: Strategy)

    -> (async sat(s, fml, new Assignment())✝)) 22 ParT Computation

77. Implementation • Futures are essential to ParT • Futures are

    handles to single parallel computations • ParT are handle to collection of parallel computations 23

78. Implementation ( )✝ :: Fut (Par t) → Par t

    ( ) :: Fut t → Par t ⚬ { v } :: t → Par t ƒParT LIFT 24 ƒ1 VALUE

79. Par Implementation ( )✝ :: Fut (Par t) → Par

    t ( ) :: Fut t → Par t ⚬ { v } :: t → Par t ƒParT LIFT 24 Par Par ƒ1 VALUE

80. Implementation >> :: Par t → (t → t’) →

    Par t’ ParT ƒ1 >> function ParT ƒ2 = SEQUENCE 25

81. Implementation >> :: Par t → (t → t’) →

    Par t’ ParT ƒ1 >> function ParT ƒ2 ƒ1 ⤳ = SEQUENCE 25 =

82. Implementation >> :: Par t → (t → t’) →

    Par t’ ParT ƒ1 >> function ParT ƒ2 ƒ1 ⤳ = ParT >> function ParT = VALUE VALUE’ SEQUENCE 25 =

83. Combinators << :: (Fut(Maybe t) → Par t’) → Par

    t→ Par t’ ƒ1 → → ƒ2 → function Par t’ ParT << VALUE PRUNE 33

84. Combinators << :: (Fut(Maybe t) → Par t’) → Par

    t→ Par t’ ƒ1 → → ƒ2 → function Par t’ ParT << VALUE PRUNE 33

85. Combinators << :: (Fut(Maybe t) → Par t’) → Par

    t→ Par t’ ƒ1 → → ƒ2 → function Par t’ ParT << VALUE PRUNE 33

86. Combinators << :: (Fut(Maybe t) → Par t’) → Par

    t→ Par t’ ƒ1 → function Par t’ ParT << VALUE PRUNE 33

87. Combinators << :: (Fut(Maybe t) → Par t’) → Par

    t→ Par t’ ƒ1 → function Par t’ ParT << VALUE PRUNE 33 ParT

88. Combinators << :: (Fut(Maybe t) → Par t’) → Par

    t→ Par t’ ƒ1 → function Par t’ ParT << VALUE PRUNE 33 ParT

89. Future work & Limitations • Limitation: Computations must be side-effect

    free • Integration with actors possible via software transactional memory • Adding more combinators • Improving the current implementation 26

90. Conclusions • ParT: abstraction to handle parallel and async computations

    • Non-blocking combinators • Tracks down and kills-off speculative work 27

91. Questions 28