JIT Compilation heuristics and parallelism exploitation for Sulong

Transcript

1. JIT Compilation heuristics and parallelism exploitation for Sulong Swapnil Gaikwad, Andy Nisbet, Mikel Luján 29 September, 2017 VMM’17
2. Outline 2 • Sulong Project – Bitcode Parser • Ongoing work – Trace Based approach to help tuning compilation heuristics • Future Plan – Executing OpenMP programs on JVM using Sulong
3. Outline 2 • Sulong Project – Bitcode Parser • Ongoing work – Trace Based approach to help tuning compilation heuristics • Future Plan – Executing OpenMP programs on JVM using Sulong
4. Sulong Project Overview 3 LLVM IR Parser Truffle JVM Graal LLVM IR C C++ Fortran* Operating System *Fortran code is converted to LLVM IR using DragonEgg plugin of GCC
5. Sulong Execution Lifecycle 4 LLVM Tools LLVM IR LLVM Bitcode parser Truffle AST Generation Truffle Representation (AST form) Graal Compiled hot path code Sulong .c .cpp .f90 Interpretation
6. Contributions to BitCode parser 5 • Initiated work on the parser for binary bitcode files – Faster parsing compared to textual parser (for .ll files) • To support the latest LLVM IR the parser is improved to parse LLVM IR version 3.9 partially – Benefit from new static optimizations to LLVM IR – Sulong developer community led this work to completion
7. Outline • Sulong Project – Bitcode Parser • Ongoing work – Trace Based approach to help tuning compilation heuristics • Future Plan – Executing OpenMP programs on JVM using Sulong
8. Motivation 6 LLVM Tools LLVM IR LLVM Bitcode parser Truffle AST Generation Truffle Representation (AST form) Graal Compiled hot path code Sulong .c .cpp .f90 Interpretation
9. Motivation 7 • For automotive benchmarks from MiBench benchmark suite a speedups up to 41% (GeoMean 7.3%) is observed in execution time for the small input data set.
10. Trace based approach 7 a b c d e f g h j k l i … for( i=0; i<1000000; i++){ … switch(option){ case 1: …… case 2: …… if(isTrue){ func(); } break; default: …… } } …… LLVM IR
11. Trace based approach 7 a b c d e f g h j k l i … for( i=0; i<1000000; i++){ … switch(option){ case 1: …… case 2: …… if(isTrue){ func(); } break; default: …… } } …… LLVM IR
12. Trace based approach 7 a b c d e f g h j k l i a c f l i • Generate a trace when tracing threshold is crossed
13. Trace based approach 7 a b c d e f g h j k l i a c f l i • Use a trace for execution instead of original code
14. Tuning Compilation Heuristics 8 • Tune compilation threshold • Optimization pass selection • Order of optimization passes
15. Outline • Sulong Project – Bitcode Parser • Ongoing work – Trace Based approach to help tuning compilation heuristics • Future Plan – Executing OpenMP programs on JVM using Sulong
16. What is OpenMP? and why? 9 • Directive based parallel programming approach • Standardized specifications • Possible to get LLVM IR for OpenMP supported languages – C, C++ and Fortran • OpenMP supports accelerator directives
17. LLVM IR for OpenMP pragmas 10 void foo() { #pragma omp parallel for for(i = 0; i < n; i++){ … //Computation … } } LLVM IR define @foo { call @__kmpc_fork_call (@.omp_outlined.) } define @.omp_outlined.() { … //get loop bounds call @__kmpc_for() //LLVM IR for Computation … }
18. Implementation Challenges 11 • Mapping OpenMP memory model using Java threads • Thread scheduling in Java • Porting OpenMP runtime
19. Summary 12 • Described the bitcode parser work at University of Manchester • Trace based approach to help tuning the compilation heuristics • Utilize multicore CPU for running OpenMP parallel programs using JVM in Sulong