heuristics for profile driven method driven method
play

Heuristics for Profile- -driven Method driven Method- - - PowerPoint PPT Presentation

Heuristics for Profile- -driven Method driven Method- - Heuristics for Profile level Speculative Parallelization level Speculative Parallelization John Whaley and Christos Kozyrakis Stanford University June 15, 2005 Speculative


  1. Heuristics for Profile- -driven Method driven Method- - Heuristics for Profile level Speculative Parallelization level Speculative Parallelization John Whaley and Christos Kozyrakis Stanford University June 15, 2005

  2. Speculative Multithreading • Speculatively parallelize an application – Uses speculation to overcome ambiguous dependencies – Uses hardware support to recover from misspeculation – Promising technique for automatically extracting parallelism from programs • Problem: Where to put the threads? June 15, 2005 Heuristics for Profile-driven Method- 1 level Speculative Parallelization

  3. Method-Level Speculation • Idea: Use method boundaries as speculative threads – Computation is naturally partitioned into methods – Execution often independent – Well-defined interface • Extract parallelism from irregular, non-numerical applications June 15, 2005 Heuristics for Profile-driven Method- 2 level Speculative Parallelization

  4. Method-Level Speculation Example main() foo() { { work_A; work_B; // writes *p } foo(); work_C; // reads *q } June 15, 2005 Heuristics for Profile-driven Method- 3 level Speculative Parallelization

  5. Method-Level Speculation Example main() { work_A; foo() { work_B; // writes *p } work_C; // reads *q } June 15, 2005 Heuristics for Profile-driven Method- 4 level Speculative Parallelization

  6. Method-Level Speculation Example work_A main() { foo() work_A; work_B foo() { work_B; // writes *p } work_C; // reads *q work_C } Sequential execution June 15, 2005 Heuristics for Profile-driven Method- 5 level Speculative Parallelization

  7. Method-Level Speculation Example work_A main() fork { overhead foo() work_A; work_B foo() { work_C work_B; // writes *p p!=q } No violation work_C; // reads *q } TLS execution – no violation June 15, 2005 Heuristics for Profile-driven Method- 6 level Speculative Parallelization

  8. Method-Level Speculation Example work_A main() fork { overhead foo() work_A; work_C work_B foo() { (aborted) work_B; // writes *p overhead p=q } Violation! work_C; // reads *q } work_C TLS execution – violation June 15, 2005 Heuristics for Profile-driven Method- 7 level Speculative Parallelization

  9. Method-Level Speculation Example Sequential TLS – no violation TLS – violation work_A work_A work_A fork fork overhead overhead foo() foo() foo() work_C work_B work_B work_B (aborted) work_C p!=q p=q overhead No violation Violation! work_C work_C June 15, 2005 Heuristics for Profile-driven Method- 8 level Speculative Parallelization

  10. Nested Speculation fork main() overhead { work_B foo() { foo() fork work_A; work_A overhead } bar() work_B; work_C work_D bar() { work_C; } work_D; Sequences of method calls can } cause nested speculation. June 15, 2005 Heuristics for Profile-driven Method- 9 level Speculative Parallelization

  11. This Talk: Choosing Speculation Points • Which methods to speculate? – Low chance of violation – Not too short, not too long – Not too many stores • Idea: Use profile data to choose good speculation points – Used for profile-driven and dynamic compiler – Should be low-cost but accurate • We evaluated 7 different heuristics – ~80% effective compared to perfect oracle June 15, 2005 Heuristics for Profile-driven Method- 10 level Speculative Parallelization

  12. Difficulties in Method-Level Speculation • Method invocations can have varying execution times – Too short: Doesn’t overcome speculation overhead – Too long: More likely to violate or overflow, prevents other threads from retiring • Return values – Mispredicted return value causes violation June 15, 2005 Heuristics for Profile-driven Method- 11 level Speculative Parallelization

  13. Classes of Heuristics • Simple Heuristics – Use only simple information, such as method runtime • Single-Pass Heuristics – More advanced information, such as sequence of store addresses – Single pass through profile data • Multi-Pass Heuristics – Multiple passes through profile data June 15, 2005 Heuristics for Profile-driven Method- 12 level Speculative Parallelization

  14. Classes of Heuristics • Simple Heuristics – Use only simple information, such as method runtime • Single-Pass Heuristics – More advanced information, such as sequence of store addresses – Single pass through profile data • Multi-Pass Heuristics – Multiple passes through profile data June 15, 2005 Heuristics for Profile-driven Method- 13 level Speculative Parallelization

  15. Runtime Heuristic (SI-RT) • Speculate on all methods with: – MIN < runtime < MAX • Idea: Should be long enough to amortize overhead, but not long enough to violate • Data required: – Average runtime of each method June 15, 2005 Heuristics for Profile-driven Method- 14 level Speculative Parallelization

  16. Store Heuristic (SI-SC) • Speculate on all methods with: – dynamic # of stores < MAX • Idea: Stores cause violations, so speculate on methods with few stores • Data required: – Average dynamic store count of each method June 15, 2005 Heuristics for Profile-driven Method- 15 level Speculative Parallelization

  17. Classes of Heuristics • Simple Heuristics – Use only simple information, such as method runtime • Single-Pass Heuristics – More advanced information, such as sequence of store addresses – Single pass through profile data • Multi-Pass Heuristics – Multiple passes through profile data June 15, 2005 Heuristics for Profile-driven Method- 16 level Speculative Parallelization

  18. Stalled Threads fork foo() overhead work_B { bar() { bar() work_A; work_A idle } work_B; } Speculative threads may stall while waiting to become main thread. June 15, 2005 Heuristics for Profile-driven Method- 17 level Speculative Parallelization

  19. Fork at intermediate points foo() { bar() { bar() work_A; work_A } fork work_B; overhead work_B } Fork at an intermediate point within a method to avoid violations and stalling June 15, 2005 Heuristics for Profile-driven Method- 18 level Speculative Parallelization

  20. Best Speedup Heuristic (SP-SU) • Speculate on methods with: – predicted speedup > THRES • Calculate predicted speedup by: expected sequential run time expected parallel run time • Scan store stream backwards to find fork point – Choose fork point to avoid violations and stalling June 15, 2005 Heuristics for Profile-driven Method- 19 level Speculative Parallelization

  21. Most Cycles Saved Heuristic (SP-CS) • Speculate on methods with: – predicted cycle savings > THRES • Calculate predicted cycle savings by: sequential cycle count – parallel cycle count • Place fork point such that: – predicted probability of violation < RATIO • Uses same information as SP-SU June 15, 2005 Heuristics for Profile-driven Method- 20 level Speculative Parallelization

  22. Classes of Heuristics • Simple Heuristics – Use only simple information, such as method runtime • Single-Pass Heuristics – More advanced information, such as sequence of store addresses – Single pass through profile data • Multi-Pass Heuristics – Multiple passes through profile data June 15, 2005 Heuristics for Profile-driven Method- 21 level Speculative Parallelization

  23. Nested Speculation fork main() overhead foo() { work_A foo() { work_D fork work_A; bar() { overhead bar() work_B; work_B foo() } idle work_C work_C; } work_D; } Effectiveness of speculation choice depends on choices for caller methods! June 15, 2005 Heuristics for Profile-driven Method- 22 level Speculative Parallelization

  24. Best Speedup Heuristic with Parent Info (MP-SU) • Iterative algorithm: – Choose speculation with best speedup – Readjust all callee methods to account for speculation in caller – Repeat until best speedup < THRES • Max # of iterations: depth of call graph June 15, 2005 Heuristics for Profile-driven Method- 23 level Speculative Parallelization

  25. Most Cycles Saved Heuristic with Parent Info (MP-CS) • Iterative algorithm: 1.Choose speculation with most cycles saved and predicted violations < RATIO 2.Readjust all callee methods to account for speculation in caller 3.Repeat until most cycles saved < THRES • Multi-pass version of SP-CS June 15, 2005 Heuristics for Profile-driven Method- 24 level Speculative Parallelization

  26. Most Cycles Saved Heuristic with No Nesting (MP-CSNN) • Iterative algorithm: – Choose speculation with most cycles saved and predicted violations < RATIO. – Eliminate all callee methods from consideration. – Repeat until most cycles saved < THRES. • Disallows nested speculation to avoid double-counting the benefits • Faster to compute than MP-CS June 15, 2005 Heuristics for Profile-driven Method- 25 level Speculative Parallelization

  27. Experimental Results Experimental Results June 15, 2005 Heuristics for Profile-driven Method- 26 level Speculative Parallelization

Recommend


More recommend