cross layer workload characterization of meta tracing jit
play

Cross-Layer Workload Characterization of Meta-Tracing JIT VMs Berkin - PowerPoint PPT Presentation

Dec 20, 2023 •385 likes •1.48k views

Cross-Layer Workload Characterization of Meta-Tracing JIT VMs Berkin Ilbeyi 1 , Carl Friedrich Bolz-Tereick 2 , and Christopher Batten 1 1 Cornell University, 2 Heinrich-Heine-Universitt Dsseldorf Dynamic languages are popular S. Cass. The

  1. Python-based interpreter ... b += a ... Application: FooLang ... compile 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) Application: Bytecode 27 INPLACE_ADD interpret 28 STORE_FAST 1 (b) ... Interpreter: Python while True: compile bc = bcs[bci] bci += bc.length Interpreter: Bytecode if bc.type == INPLACE_ADD: v1 = stack.pop() v2 = stack.pop() if (type(v1) == int and type(v2) == int): stack.push(v1 + v2) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 7

  2. Python-based interpreter ... b += a ... Application: FooLang ... compile 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) Application: Bytecode 27 INPLACE_ADD interpret 28 STORE_FAST 1 (b) ... Interpreter: Python while True: compile bc = bcs[bci] bci += bc.length Interpreter: Bytecode if bc.type == INPLACE_ADD: interpret v1 = stack.pop() v2 = stack.pop() Interpreter Interpreter if (type(v1) == int and type(v2) == int): stack.push(v1 + v2) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 7

  3. RPython Framework Application: Python compile Application: Bytecode Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 8

  4. RPython Framework Application: Python Interpreter: RPython compile Application: Bytecode Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 8

  5. RPython Framework Application: Python Interpreter: RPython Framework: RPython compile Application: Bytecode Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 8

  6. RPython Framework Application: Python Interpreter: RPython Framework: RPython translate compile Interpreter + Framework: C Application: Bytecode Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 8

  7. RPython Framework Application: Python Interpreter: RPython Framework: RPython translate compile Interpreter + Framework: C Application: Bytecode compile interpret PyPy: Binary Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 8

  8. RPython Framework Application: Python Interpreter: RPython Framework: RPython translate compile Interpreter + Framework: C Application: Bytecode compile interpret trace and optimize PyPy: Binary Meta-trace: JIT IR assemble JIT-ed code: Binary Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 8

  9. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) ... Interpreter while True: bc = bcs[bci] bci += bc.length if bc.type == INPLACE_ADD: v1 = stack.pop() v2 = stack.pop() if (type(v1) == int and type(v2) == int): stack.push(v1 + v2) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  10. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) Meta-interpreter ... Interpreter while True: Meta-trace bc = bcs[bci] bci += bc.length ... if bc.type == INPLACE_ADD: v1 = stack.pop() v2 = stack.pop() if (type(v1) == int and type(v2) == int): stack.push(v1 + v2) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  11. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) Meta-interpreter ... Interpreter while True: Meta-trace bc = bcs[bci] bci += bc.length ... if bc.type == INPLACE_ADD: v1 = stack.pop() v2 = stack.pop() if (type(v1) == int and type(v2) == int): stack.push(v1 + v2) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  12. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) Meta-interpreter ... Interpreter while True: Meta-trace bc = bcs[bci] bci += bc.length ... if bc.type == INPLACE_ADD: v1 = stack.pop() v2 = stack.pop() if (type(v1) == int and type(v2) == int): stack.push(v1 + v2) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  13. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) Meta-interpreter ... Interpreter while True: Meta-trace bc = bcs[bci] bci += bc.length ... if bc.type == INPLACE_ADD: p1 = getarrayitem(p0, 1) v1 = stack.pop() v2 = stack.pop() if (type(v1) == int and type(v2) == int): stack.push(v1 + v2) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  14. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) Meta-interpreter ... Interpreter while True: Meta-trace bc = bcs[bci] bci += bc.length ... if bc.type == INPLACE_ADD: p1 = getarrayitem(p0, 1) v1 = stack.pop() v2 = stack.pop() if (type(v1) == int and type(v2) == int): stack.push(v1 + v2) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  15. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) Meta-interpreter ... Interpreter while True: Meta-trace bc = bcs[bci] bci += bc.length ... if bc.type == INPLACE_ADD: p1 = getarrayitem(p0, 1) v1 = stack.pop() p2 = getarrayitem(p0, 0) v2 = stack.pop() if (type(v1) == int and type(v2) == int): stack.push(v1 + v2) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  16. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) Meta-interpreter ... Interpreter while True: Meta-trace bc = bcs[bci] bci += bc.length ... if bc.type == INPLACE_ADD: p1 = getarrayitem(p0, 1) v1 = stack.pop() p2 = getarrayitem(p0, 0) v2 = stack.pop() if (type(v1) == int and type(v2) == int): stack.push(v1 + v2) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  17. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) Meta-interpreter ... Interpreter while True: Meta-trace bc = bcs[bci] bci += bc.length ... if bc.type == INPLACE_ADD: p1 = getarrayitem(p0, 1) v1 = stack.pop() p2 = getarrayitem(p0, 0) v2 = stack.pop() if (type(v1) == int and type(v2) == int): stack.push(v1 + v2) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  18. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) Meta-interpreter ... Interpreter while True: Meta-trace bc = bcs[bci] bci += bc.length ... if bc.type == INPLACE_ADD: p1 = getarrayitem(p0, 1) v1 = stack.pop() p2 = getarrayitem(p0, 0) v2 = stack.pop() guard_class(p1, int) if (type(v1) == int and guard_class(p2, int) type(v2) == int): stack.push(v1 + v2) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  19. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) Meta-interpreter ... Interpreter while True: Meta-trace bc = bcs[bci] bci += bc.length ... if bc.type == INPLACE_ADD: p1 = getarrayitem(p0, 1) v1 = stack.pop() p2 = getarrayitem(p0, 0) v2 = stack.pop() guard_class(p1, int) if (type(v1) == int and guard_class(p2, int) type(v2) == int): i3 = getfield(p1, intval) stack.push(v1 + v2) i4 = getfield(p2, intval) elif ... elif bc.type == LOAD_FAST: stack.push(local[bc.varnum]) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  20. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) Meta-interpreter ... Interpreter while True: Meta-trace bc = bcs[bci] bci += bc.length ... if bc.type == INPLACE_ADD: p1 = getarrayitem(p0, 1) v1 = stack.pop() p2 = getarrayitem(p0, 0) v2 = stack.pop() guard_class(p1, int) if (type(v1) == int and guard_class(p2, int) type(v2) == int): i3 = getfield(p1, intval) stack.push(v1 + v2) i4 = getfield(p2, intval) elif ... i5 = int_add_ovf(i3, i4) elif bc.type == LOAD_FAST: guard_no_overflow() stack.push(local[bc.varnum]) ... ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  21. Meta-trace Application bytecode ... 21 LOAD_FAST 1 (b) 24 LOAD_FAST 0 (a) 27 INPLACE_ADD 28 STORE_FAST 1 (b) Meta-interpreter ... Interpreter while True: Meta-trace bc = bcs[bci] bci += bc.length ... if bc.type == INPLACE_ADD: p1 = getarrayitem(p0, 1) v1 = stack.pop() p2 = getarrayitem(p0, 0) v2 = stack.pop() guard_class(p1, int) if (type(v1) == int and guard_class(p2, int) Deoptimization back to interpreter on guard failure type(v2) == int): i3 = getfield(p1, intval) stack.push(v1 + v2) i4 = getfield(p2, intval) elif ... i5 = int_add_ovf(i3, i4) elif bc.type == LOAD_FAST: guard_no_overflow() stack.push(local[bc.varnum]) ... ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 9

  22. Cross-layer annotations Application: Python Interpreter: RPython Framework: RPython translate compile Interpreter + Application: C Application: Bytecode compile interpret trace and optimize PyPy: Binary Meta-trace: JIT IR assemble JIT-ed code: Binary Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 10

  23. Cross-layer annotations application annotations Application: Python Interpreter: RPython Framework: RPython translate compile Interpreter + Application: C Application: Bytecode compile interpret trace and optimize PyPy: Binary Meta-trace: JIT IR assemble JIT-ed code: Binary Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 10

  24. Cross-layer annotations application annotations interpreter annotations Application: Python Interpreter: RPython Framework: RPython translate compile Interpreter + Application: C Application: Bytecode compile interpret trace and optimize PyPy: Binary Meta-trace: JIT IR assemble JIT-ed code: Binary Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 10

  25. Cross-layer annotations application annotations interpreter annotations framework annotations Application: Python Interpreter: RPython Framework: RPython translate compile Interpreter + Application: C Application: Bytecode compile interpret trace and optimize PyPy: Binary Meta-trace: JIT IR assemble JIT-ed code: Binary Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 10

  26. Cross-layer annotations application annotations interpreter annotations framework annotations Application: Python Interpreter: RPython Framework: RPython translate compile Interpreter + Application: C Application: Bytecode compile interpret trace and optimize PyPy: Binary Meta-trace: JIT IR IR node of interest assemble JIT-ed code: Binary asm of interest Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 10

  27. Cross-layer annotations application annotations interpreter annotations framework annotations Application: Python Interpreter: RPython Framework: RPython translate compile Interpreter + Application: C Application: Bytecode compile interpret trace and optimize PyPy: Binary Meta-trace: JIT IR IR node of interest assemble perf counters JIT-ed code: Binary asm of interest using PAPI Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 10

  28. Cross-layer annotations application annotations interpreter annotations framework annotations Application: Python Interpreter: RPython Framework: RPython translate compile Interpreter + Application: C Application: Bytecode compile interpret trace and optimize PyPy: Binary Meta-trace: JIT IR IR node of interest assemble perf counters JIT-ed code: Binary asm of interest using PAPI Dynamic Binary Instrumentation Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 10

  29. Cross-layer annotations application annotations interpreter annotations framework annotations Application: Python Interpreter: RPython Framework: RPython translate compile Interpreter + Application: C Application: Bytecode compile interpret trace and optimize PyPy: Binary Meta-trace: JIT IR IR node of interest assemble perf counters JIT-ed code: Binary asm of interest using PAPI phase counters, Dynamic Binary Instrumentation IR node counters Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 10

  30. Cross-layer workload characterization of meta-tracing JIT VMs PyPy >> CPython ▪ How can meta-tracing JITs significantly improve the performance of multiple dynamic languages? PyPy << C ▪ Why are meta-tracing JITs for dynamic programming still slower than C? Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 11

  31. Meta-tracing JIT improves the performance significantly PyPy with meta-tracing JIT speedup over CPython: 10 15 20 25 30 0 5 51.2 richards crypto_pyaes 30.2 chaos telco spectral-norm django twisted_iteration spitfire_cstringio raytrace-simple hexiom2 float ai nbody_modified twisted_pb fannkuch genshi_text pyflate-fast bm_mako twisted_names json_bench Motivation • Meta-tracing • PyPy >> CPython • PyPy << C genshi_xml bm_chameleon pypy_interp twisted_tcp html5lib meteor-contest sympy_sum spitfire spambayes rietveld deltablue eparse sympy_expand slowspitfire sympy_integrate pidigits bm_mdp sympy_str 12

  32. PyPy speedup over CPython and Pycket speedup over Racket: Meta-tracing JIT improves performance significantly across multiple languages PyPy speedup 12 11 10 9 8 7 6 5 4 3 2 1 0 binarytrees chameneosredux fannkuchredux fasta knucleotide mandelbrot meteor nbody pidigits regexdna revcomp spectralnorm threadring Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 13

  33. PyPy speedup over CPython and Pycket speedup over Racket: Meta-tracing JIT improves performance significantly across multiple languages Pycket speedup PyPy speedup 2 12 11 10 1.5 9 8 7 1 6 5 4 0.5 3 2 1 0 0 binarytrees fannkuchredux fasta mandelbrot meteor nbody pidigits revcomp spectralnorm binarytrees chameneosredux fannkuchredux fasta knucleotide mandelbrot meteor nbody pidigits regexdna revcomp spectralnorm threadring Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 13

  34. Meta-tracing JIT VM phases richards calls to AOT funs JIT GC deoptimization tracing & opt interpreter 0 2B 4B 6B 8B 10B instructions Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 14

  35. Meta-tracing JIT VM phases richards calls to AOT funs JIT GC deoptimization tracing & opt interpreter 0 2B 4B 6B 8B 10B instructions sympy_str calls to AOT funs JIT GC deoptimization tracing & opt interpreter 0 2B 4B 6B 8B 10B instructions Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 14

  36. Meta-tracing JIT VM phases Fastest on PyPy Slowest on PyPy JIT calls JIT GC deopt tracing interp Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 15

  37. The JIT phase: The fastest benchmarks tend to execute JIT-compiled code the most JIT + JIT call to AOT 1 0.75 0.5 0.25 0 Benchmarks Fastest on PyPy Benchmarks Slowest on PyPy Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 16

  38. Meta-tracing inlines all loops and can hurt performance Interpreter while True: ... memcpy(d, s, n) ... def memcpy(dest, src, n): i = 0 while i < n: dest[i] = src[i] i += 1 Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 17

  39. Meta-tracing inlines all loops and can hurt performance Interpreter while True: ... memcpy(d, s, n) Meta-interpreter ... def memcpy(dest, src, n): i = 0 while i < n: dest[i] = src[i] Meta-trace i += 1 ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 17

  40. Meta-tracing inlines all loops and can hurt performance Interpreter while True: ... memcpy(d, s, n) Meta-interpreter ... def memcpy(dest, src, n): i = 0 while i < n: dest[i] = src[i] Meta-trace i += 1 ... guard_gt(i0, 0) i3 = getarrayitem(p1, 0) setarrayitem(p2, 0, i3) Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 17

  41. Meta-tracing inlines all loops and can hurt performance Interpreter while True: ... memcpy(d, s, n) Meta-interpreter ... def memcpy(dest, src, n): i = 0 while i < n: dest[i] = src[i] Meta-trace i += 1 ... guard_gt(i0, 0) i3 = getarrayitem(p1, 0) setarrayitem(p2, 0, i3) guard_gt(i0, 1) i4 = getarrayitem(p1, 1) setarrayitem(p2, 1, i4) Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 17

  42. Meta-tracing inlines all loops and can hurt performance Interpreter while True: ... memcpy(d, s, n) Meta-interpreter ... def memcpy(dest, src, n): i = 0 while i < n: dest[i] = src[i] Meta-trace i += 1 ... guard_gt(i0, 0) i3 = getarrayitem(p1, 0) setarrayitem(p2, 0, i3) guard_gt(i0, 1) i4 = getarrayitem(p1, 1) setarrayitem(p2, 1, i4) guard_gt(i0, 2) i5 = getarrayitem(p1, 2) setarrayitem(p2, 2, i5) ... Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 17

  43. Examples of significant AOT-compiled functions Benchmark % Source Function ai 19.4 interpreter setobject.get_storage_from_list bm_chameleon 17.9 RPython types rordereddict.ll_call_lookup_function bm_mako 26.1 RPython lib runicode.unicode_encode_ucs1_helper json_bench 18.5 PyPy module _pypyjson.raw_encode_basestring_ascii nbody_modified 44.6 external lib pow Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 18

  44. JIT calls to AOT-compiled functions: AOT-compiled functions can improve performance by avoiding long traces JIT JIT call to AOT functions 1 0.75 0.5 0.25 0 Benchmarks Fastest on PyPy Benchmarks Slowest on PyPy Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 19

  45. PyPy bytecode execution rate compared to CPython: Benchmarks that perform the best also warm up the fastest Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 20

  46. PyPy bytecode execution rate compared to CPython: Benchmarks that perform the best also warm up the fastest richards 50 30 10 0 2B 4B 6B 8B 10B instructions Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 20

  47. PyPy bytecode execution rate compared to CPython: Benchmarks that perform the best also warm up the fastest Breakeven point: the performance of the two VMs at this point is equal richards 50 30 10 0 2B 4B 6B 8B 10B instructions PyPy w/o JIT breakeven point CPython breakeven point Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 20

  48. PyPy bytecode execution rate compared to CPython: Benchmarks that perform the best also warm up the fastest Breakeven point: the performance of the two VMs at this point is equal richards html5lib 3 50 2 30 1 10 0 2B 4B 6B 8B 10B 0 2B 4B 6B 8B 10B instructions instructions Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 20

  49. PyPy bytecode execution rate compared to CPython: Benchmarks that perform the best also warm up the fastest Breakeven point: the performance of the two VMs at this point is equal richards html5lib 3 50 2 30 1 10 0 2B 4B 6B 8B 10B 0 2B 4B 6B 8B 10B instructions instructions PyPy w/o JIT breakeven point CPython breakeven point Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 20

  50. PyPy bytecode execution rate compared to CPython: Benchmarks that perform the best also warm up the fastest Breakeven point: the performance of the two VMs at this point is equal richards html5lib sympy_str 3 50 2 2 30 1 1 10 0 2B 4B 6B 8B 10B 0 2B 4B 6B 8B 10B 0 2B 4B 6B 8B 10B instructions instructions instructions PyPy w/o JIT breakeven point Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 20

  51. Cross-layer workload characterization of meta-tracing JIT VMs PyPy >> CPython ▪ How can meta-tracing JITs significantly improve the performance of multiple dynamic languages? PyPy << C ▪ Why are meta-tracing JITs for dynamic programming still slower than C? Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 21

  52. Cross-layer workload characterization of meta-tracing JIT VMs PyPy >> CPython ▪ How can meta-tracing JITs significantly improve the performance of multiple dynamic languages? ▪ Meta-tracing JIT compilation significantly improves the performance PyPy << C ▪ Why are meta-tracing JITs for dynamic programming still slower than C? Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 21

  53. Cross-layer workload characterization of meta-tracing JIT VMs PyPy >> CPython ▪ How can meta-tracing JITs significantly improve the performance of multiple dynamic languages? ▪ Meta-tracing JIT compilation significantly improves the performance ▪ AOT-compiled functions are good to break pathological traces PyPy << C ▪ Why are meta-tracing JITs for dynamic programming still slower than C? Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 21

  54. Cross-layer workload characterization of meta-tracing JIT VMs PyPy >> CPython ▪ How can meta-tracing JITs significantly improve the performance of multiple dynamic languages? ▪ Meta-tracing JIT compilation significantly improves the performance ▪ AOT-compiled functions are good to break pathological traces ▪ Easier-to-JIT programs perform the best and warm up the fastest PyPy << C ▪ Why are meta-tracing JITs for dynamic programming still slower than C? Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 21

  55. PyPy and Pycket slowdown over C/C++: Meta-tracing JIT has a big performance gap between static languages 1374 PyPy slowdown 30 31 25 20 15 10 5 0 binarytrees chameneosredux fannkuchredux fasta knucleotide mandelbrot meteor nbody pidigits regexdna revcomp spectralnorm threadring Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 22

  56. PyPy and Pycket slowdown over C/C++: Meta-tracing JIT has a big performance gap between static languages 1374 Pycket slowdown PyPy slowdown 12 30 31 10 25 8 20 6 15 4 10 2 5 0 0 binarytrees fannkuchredux fasta mandelbrot meteor nbody pidigits revcomp spectralnorm binarytrees chameneosredux fannkuchredux fasta knucleotide mandelbrot meteor nbody pidigits regexdna revcomp spectralnorm threadring Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 22

  57. Meta-tracing JIT phases JIT JIT call to AOT functions 1 0.75 0.5 0.25 0 Benchmarks Fastest on PyPy Benchmarks Slowest on PyPy Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 23

  58. Meta-tracing JIT IR node breakdown: Likely a big part of JIT compiled code is overhead Fastest on PyPy Slowest on PyPy Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 24

  59. Meta-tracing JIT IR node breakdown: Likely a big part of JIT compiled code is overhead Fastest on PyPy Slowest on PyPy Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 24

  60. Meta-tracing JIT IR node breakdown: Likely a big part of JIT compiled code is overhead Fastest on PyPy Slowest on PyPy Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 24

  61. Meta-tracing JIT phases JIT JIT call to AOT functions 1 0.75 0.5 0.25 0 Benchmarks Fastest on PyPy Benchmarks Slowest on PyPy Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 25

  62. Interpreter phase Interpreter 1 0.75 0.5 0.25 0 Benchmarks Fastest on PyPy Benchmarks Slowest on PyPy Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 26

  63. RPython-to-C translation has overheads PyPy without meta-tracing JIT speedup over CPython: 0.2 0.4 0.6 0.8 1.2 0 1 richards crypto_pyaes chaos telco spectral-norm django twisted_iteration spitfire_cstringio raytrace-simple hexiom2 float ai nbody_modified twisted_pb fannkuch genshi_text pyflate-fast bm_mako twisted_names json_bench Motivation • Meta-tracing • PyPy >> CPython • PyPy << C genshi_xml bm_chameleon pypy_interp twisted_tcp html5lib meteor-contest sympy_sum spitfire spambayes rietveld deltablue eparse sympy_expand slowspitfire sympy_integrate pidigits bm_mdp sympy_str 27

  64. Tracing and optimization phase Tracing & optimization 1 0.75 0.5 0.25 0 Benchmarks Fastest on PyPy Benchmarks Slowest on PyPy Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 28

  65. Deoptimization phase Deoptimization 1 0.75 0.5 0.25 0 Benchmarks Fastest on PyPy Benchmarks Slowest on PyPy Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 29

  66. Garbage collection phase Garbage collection 1 0.75 0.5 0.25 0 Benchmarks Fastest on PyPy Benchmarks Slowest on PyPy Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 30

  67. Meta-tracing JIT VM overheads: Overheads are diverse and can add up to significant portion of execution Interpreter Tracing & optimization Deoptimization Garbage collection 1 0.75 0.5 0.25 0 Benchmarks Fastest on PyPy Benchmarks Slowest on PyPy Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 31

  68. Iron law of processor performance: Does meta-tracing VM code execute poorly in addition to more instructions? Time Instructions Cycle Time × = × Program Program Instructions Cycle Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 32

  69. Comparing meta-tracing JIT IPC to C/C++: Meta-tracing has a similar IPC for most benchmarks C/C++ IPC PyPy IPC Pycket IPC 2.25 1.875 1.5 1.125 0.75 0.375 0 s x x a e t r y s a p m g o o e u u t d d t n m n s r e i r e g d d i o d o i b a t t o r r i e e o e b x n d t f l d e c r r y e m n e l a i s h d v a r p g l e a o c c n e r e t r n e u r u a c h r i n k n m e t b e n k p m n s a a f h c Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 33

  70. Comparing meta-tracing JIT IPC to C/C++: Meta-tracing has a similar IPC for most benchmarks C/C++ IPC PyPy IPC Pycket IPC 2.25 1.875 1.5 1.125 0.75 0.375 0 s x x a e t r y s a p m g o o e u u t d d t n m n s r e i r e g d d i o d o i b a t t o r r i e e o e b x n d t f l d e c r r y e m n e l a i s h d v a r p g l e a o c c n e r e t r n e u r u a c h r i n k n m e t b e n k p m n s a a f h c Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 33

  71. IPC measurements can be accurately matched against VM phases JIT GC deopt trace interp Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 34

  72. Microarchitectural characterization by the VM phase: Meta-tracing-JIT-compiled code has a similar IPC, fewer branches and mispredictions IPC 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 Interp Trace Deopt GC JIT C/C++ Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 35

  73. Microarchitectural characterization by the VM phase: Meta-tracing-JIT-compiled code has a similar IPC, fewer branches and mispredictions IPC Branch per instruction 1.8 0.2 1.6 1.4 0.15 1.2 1 0.1 0.8 0.6 0.05 0.4 0.2 0 0 Interp Trace Deopt GC JIT C/C++ Interp Trace Deopt GC JIT C/C++ Motivation • Meta-tracing • PyPy >> CPython • PyPy << C 35

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend

DAY 2 Agenda for Today Introduce the workload characterization problem. Discuss a

DAY 2 Agenda for Today Introduce the workload characterization problem. Discuss a

DAY 2 Agenda for Today Introduce the workload characterization problem. Discuss a simple example of characterizing the workload for an intranet. What is workload characterization? What workload we want to characterize?

589 views • 43 slides
ISA-Independent W ISA-Independent Workload Characterization and orkload Characterization and

ISA-Independent W ISA-Independent Workload Characterization and orkload Characterization and

ISA-Independent W ISA-Independent Workload Characterization and orkload Characterization and Implications for Specialized Ar Implications for Specialized Architectur chitectures es Yakun Sophia Shao and David Brooks Harvard University

691 views • 35 slides
Automated Workload Characterization for I/O Performance Analysis in

Automated Workload Characterization for I/O Performance Analysis in

Automated Workload Characterization for I/O Performance Analysis in Virtualized Environments Axel Busch, Qais Noorshams, Samuel Kounev, February 4 th , 2015 Anne Koziolek, Ralf Reussner, Erich Amrehn ICPE 2015 , Austin,

561 views • 23 slides
Design a Cross-Layer Cognitive Engine using Cross-Layer Optimization with Case- Based Reasoning

Design a Cross-Layer Cognitive Engine using Cross-Layer Optimization with Case- Based Reasoning

Design a Cross-Layer Cognitive Engine using Cross-Layer Optimization with Case- Based Reasoning and Reinforcement Learning 4 th Workshop of COST Action IC0902 Rome, Italy 09 11, October, 2013 Ali Haider Mahdi, Andreas Mitschele-Thiel

672 views • 40 slides
1 ISO/OSI reference model: end ISO/OSI reference model: end- -to to- -end dialogue 2/2 end

1 ISO/OSI reference model: end ISO/OSI reference model: end- -to to- -end dialogue 2/2 end

Topics in Computer Networks 2010 Introduction, What is CrossLayer Design ? Mesh Network Cross Layer Design : Seminar 3 Taxonomy of cross-layer methods Detailed overview on cross-layer interactions Jonas Karlsson Karlstad Universitet

307 views • 10 slides
Workload, Fatigue, and Sleep Disruption 1 Workload 1.What is workload? 2.What is the

Workload, Fatigue, and Sleep Disruption 1 Workload 1.What is workload? 2.What is the

Workload, Fatigue, and Sleep Disruption 1 Workload 1.What is workload? 2.What is the relationship between workload and performance? 3.What is the difference between stress and workload? 4.What are some factors that mediate workload? 5.How

216 views • 4 slides
Workload Characterization of Selected JEE-based Web 2.0 Applications Priya Nagpurkar , William

Workload Characterization of Selected JEE-based Web 2.0 Applications Priya Nagpurkar , William

Workload Characterization of Selected JEE-based Web 2.0 Applications Priya Nagpurkar , William Horn, U Gopalakrishnan, Niteesh Dubey, Joefon Jann, Pratap Pattnaik IBM TJ Watson Research Center IISWC 2008 Introduction Motivation : Web 2.0

577 views • 16 slides
Day 3 Agenda for Today Formulate simple problem statement Revisit the workload

Day 3 Agenda for Today Formulate simple problem statement Revisit the workload

Day 3 Agenda for Today Formulate simple problem statement Revisit the workload characterization problem. Present detailed (step by step) derivation of the workload. Workload Modeling Workload modeling is used to generate

554 views • 39 slides
Workload Characterization of a Leadership Class Storage Cluster Technology Integration Group

Workload Characterization of a Leadership Class Storage Cluster Technology Integration Group

Workload Characterization of a Leadership Class Storage Cluster Technology Integration Group National Center for Computational Sciences Presented by Youngjae Kim Youngjae Kim, Raghul Gunasekaran, Galen M. Shipman, David A. Dillow, Zhe Zhang,

458 views • 21 slides
Analysis of peer-to-peer systems: workload characterization and ef- fects on traffic cacheability

Analysis of peer-to-peer systems: workload characterization and ef- fects on traffic cacheability

Analysis of peer-to-peer systems: workload characterization and ef- fects on traffic cacheability Mauro Andreolini University of Rome Tor Vergata Riccardo Lancellotti University of Modena and Reggio Emilia Philip S. Yu IBM T.J. Watson

358 views • 20 slides
CROSS-LAYER CROSS-LAYER LATENCY-AWARE AND -PREDICTABLE LATENCY-AWARE AND -PREDICTABLE DATA

CROSS-LAYER CROSS-LAYER LATENCY-AWARE AND -PREDICTABLE LATENCY-AWARE AND -PREDICTABLE DATA

CROSS-LAYER CROSS-LAYER LATENCY-AWARE AND -PREDICTABLE LATENCY-AWARE AND -PREDICTABLE DATA COMMUNICATION DATA COMMUNICATION Andreas Schmidt Apr 07th, 2020 Anywhere on Earth Doctoral Colloquium 1

702 views • 59 slides
Workload Characterization of 3D Games Jordi Roca, Victor Moya, Carlos Gonzlez, Chema Solis,

Workload Characterization of 3D Games Jordi Roca, Victor Moya, Carlos Gonzlez, Chema Solis,

Workload Characterization of 3D Games Jordi Roca, Victor Moya, Carlos Gonzlez, Chema Solis, Agustn Fernandez and Roger Espasa (Intel DEG Barcelona) Computer Architecture Department 1 Outline Introduction Game selection &amp;

460 views • 27 slides
GaaS Workload Characterization under NUMA Architecture for Virtualized GPU Huixiang Chen, Meng

GaaS Workload Characterization under NUMA Architecture for Virtualized GPU Huixiang Chen, Meng

GaaS Workload Characterization under NUMA Architecture for Virtualized GPU Huixiang Chen, Meng Wang, Yang Hu, Mingcong Song, Tao Li Presented by Huixiang Chen ISPASS 2017 April 24, 2017, Santa Rosa, California IDEAL (Intelligent Design of

276 views • 27 slides
A Hierarchical Characterization of a Live Streaming Media Workload Eveline Veloso Computer

A Hierarchical Characterization of a Live Streaming Media Workload Eveline Veloso Computer

A Hierarchical Characterization of a Live Streaming Media Workload Eveline Veloso Computer Science Department Virglio Almeida Federal University of Minas Gerais Wagner Meira Brazil Computer Science Department Azer Bestavros Boston

495 views • 20 slides
Recent Workload Characterization Activities at NERSC Harvey Wasserman NERSC Science Driven System

Recent Workload Characterization Activities at NERSC Harvey Wasserman NERSC Science Driven System

Recent Workload Characterization Activities at NERSC Harvey Wasserman NERSC Science Driven System Architecture Group www.nersc.gov/projects/SDSA October 15, 2008 Los Alamos Computer Science Symposium Workshop on Performance Analysis of

538 views • 53 slides
Cross-Layer and Cross-Domain QoS Signalling Using BGP 8th Wrzburg Workshop on IP: Joint

Cross-Layer and Cross-Domain QoS Signalling Using BGP 8th Wrzburg Workshop on IP: Joint

Cross-Layer and Cross-Domain QoS Signalling Using BGP 8th Wrzburg Workshop on IP: Joint EuroNF, ITC, and ITG Workshop on &quot;Visions of Future Generation Networks&quot; (EuroView2008) Thomas Martin Knoll Chemnitz University of Technology

101 views • 7 slides
THE COBRA PROGRAMMING LANGUAGE At the SoCal Piggies, Feb 2008 cobra-language.com Chuck

THE COBRA PROGRAMMING LANGUAGE At the SoCal Piggies, Feb 2008 cobra-language.com Chuck

THE COBRA PROGRAMMING LANGUAGE At the SoCal Piggies, Feb 2008 cobra-language.com Chuck Esterbrook INTRO Cobra is a new language (sub 1.0) Object-oriented, imperative Embraces unit tests, contracts and more General purpose Runs on .NET

1.15k views • 29 slides
Develop your Embedded Applications Faster: Comparing C and Golang Marcin Pasinski Mender.io My

Develop your Embedded Applications Faster: Comparing C and Golang Marcin Pasinski Mender.io My

Develop your Embedded Applications Faster: Comparing C and Golang Marcin Pasinski Mender.io My view on C vs Go I think Go is great and very productive programming language It excels when developing networking code Im not

494 views • 34 slides
CSE 501: Compiler Construction Course outline Main focus: program analysis and transformation

CSE 501: Compiler Construction Course outline Main focus: program analysis and transformation

CSE 501: Compiler Construction Course outline Main focus: program analysis and transformation Models of compilation how to represent programs? how to analyze programs? what to analyze? Standard transformations how to transform

503 views • 6 slides
Introduction to C Programming Jan Faigl Department of Computer Science Faculty of Electrical

Introduction to C Programming Jan Faigl Department of Computer Science Faculty of Electrical

Introduction to C Programming Jan Faigl Department of Computer Science Faculty of Electrical Engineering Czech Technical University in Prague Lecture 01 BE5B99CPL C Programming Language Jan Faigl, 2017 BE5B99CPL Lecture 01:

1.39k views • 94 slides
Phases of Programming SI 413 Overview of compilation Programming Languages What does

Phases of Programming SI 413 Overview of compilation Programming Languages What does

Unit 1 Phases of Programming SI 413 Overview of compilation Programming Languages What does programming actually involve? About this course Write a program Execute the program Note: an interpreter essentially does compilation and

762 views • 4 slides
INSIDE PERL MOHAMMAD T. IRFAN 11/6/13 Origin Scripting language based on sh and awk

INSIDE PERL MOHAMMAD T. IRFAN 11/6/13 Origin Scripting language based on sh and awk

11/6/2013 CSCI-2325: INSIDE PERL MOHAMMAD T. IRFAN 11/6/13 Origin Scripting language based on sh and awk Larry Wall: People often argue over the difference between a scripting language and a programming language. If you ask me,

301 views • 12 slides
TIE-20306 Principles of Programming Languages Autumn 2015 Basic Course Info Lecturer: Matti

TIE-20306 Principles of Programming Languages Autumn 2015 Basic Course Info Lecturer: Matti

TIE-20306 Principles of Programming Languages Autumn 2015 Basic Course Info Lecturer: Matti Rintala Main assistant: Jyke Savia Assistant: Pertti Lehtinen Web page: http://www.cs.tut.fi/~popl/ Email: popl@cs.tut.fi Important

607 views • 14 slides
Lecture 21: Interpreters I Marvin Zhang 07/27/2016 Announcements Roadmap Introduction

Lecture 21: Interpreters I Marvin Zhang 07/27/2016 Announcements Roadmap Introduction

Lecture 21: Interpreters I Marvin Zhang 07/27/2016 Announcements Roadmap Introduction Functions This week (Interpretation), the goals are: Data To learn a new language, Scheme, in two days! Mutability To understand how

554 views • 18 slides

More recommend