plafrim
play

PlaFRIM Exploration The Roofline model Performance Methodology - PowerPoint PPT Presentation

Sep 30, 2022 •249 likes •918 views

PlaFRIM Court` es L., Ru e F. Introduction General PlaFRIM Exploration The Roofline model Performance Methodology Court` es L., Ru e F. November 8, 2019 Table of contents PlaFRIM Court` es L., Ru e F. Introduction

  1. PlaFRIM Court` es L., Ru´ e F. Introduction General PlaFRIM Exploration The Roofline model Performance Methodology Court` es L., Ru´ e F. November 8, 2019

  2. Table of contents PlaFRIM Court` es L., Ru´ e F. Introduction Introduction 1 General Exploration The Roofline General Exploration 2 model Performance Methodology The Roofline model 3 Performance Methodology 4

  3. The hard way PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration printf(”%i”,time(NULL)); The Roofline model Performance Methodology

  4. The hard way PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration printf(”%i”,time(NULL)); The Roofline model Performance � Methodology

  5. The optimization objectives PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration Improve the speed of execution The Roofline model Performance Methodology

  6. The optimization objectives PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration Improve the speed of execution The Roofline model Reduce memory footprint Performance Methodology

  7. The optimization objectives PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration Improve the speed of execution The Roofline model Reduce memory footprint Performance Reduce energy consumption Methodology

  8. The optimization objectives PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration Improve the speed of execution The Roofline model Reduce memory footprint Performance Reduce energy consumption Methodology Consume fewer resources

  9. The Process PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration Identify bottlenecks (Profiling) The Roofline model Choose better algorithms or improve implementation Performance Methodology (Optimization)

  10. How profilers do it PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration Call stack sampling The Roofline model Performance Methodology

  11. How profilers do it PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration Call stack sampling The Roofline model Optional function call instrumentation Performance Methodology

  12. How profilers do it PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration Call stack sampling The Roofline model Optional function call instrumentation Performance Hardware simulation Methodology

  13. How profilers do it PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration Call stack sampling The Roofline model Optional function call instrumentation Performance Hardware simulation Methodology Hardware counter

  14. Memory PlaFRIM Court` es L., Ru´ e F. Introduction Understanding memory locality General Exploration The Roofline model Performance Methodology

  15. General Exploration PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline model Optimization and granularity Performance Methodology

  16. The easiest way PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration Time command The Roofline model Real, user & sys time Performance Best way to evaluate scalability Methodology

  17. The easiest way PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration Time command The Roofline model Real, user & sys time Performance Best way to evaluate scalability Methodology Accuracy of the evaluation?

  18. Profiler PlaFRIM Court` es L., Ru´ e F. Introduction General static instrumentation - gprof Exploration The Roofline Sampling technique model no instrumentation needed Performance Methodology 2 types of view (flat profile and call graph)

  19. Profiler PlaFRIM Court` es L., Ru´ e F. Introduction General static instrumentation - gprof Exploration The Roofline Sampling technique model no instrumentation needed Performance Methodology 2 types of view (flat profile and call graph) Annotated code

  20. Profiler PlaFRIM static instrumentation - gprof Court` es L., Ru´ e F. use the -pg option to compile Introduction evaluate the output : gprof ’binary name’ gmon.out General Exploration The Roofline model Performance Methodology

  21. Profiler PlaFRIM static instrumentation - gprof Court` es L., use the -pg option to compile Ru´ e F. evaluate the output : gprof ’binary name’ gmon.out Introduction General Exploration The Roofline model Performance Methodology

  22. Profiler PlaFRIM static instrumentation - gprof Court` es L., Ru´ e F. gprof -A -l ’binary name’ gmon.out Introduction General Exploration The Roofline model Performance Methodology

  23. Profiler PlaFRIM static instrumentation - gprof Court` es L., Ru´ e F. gprof -A -l ’binary name’ gmon.out Introduction General Exploration The Roofline model Performance Methodology

  24. Profiler PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline model and for memory usage ? Performance Methodology

  25. Profiler PlaFRIM Court` es L., Ru´ e F. Introduction Dynamic instrumentation - valgrind General Exploration Done at execution time The Roofline model no instrumentation needed Performance different tools for differents analysis Methodology massif - heap profiler callgrind - call history among functions cachegrind - interactions with machine cache

  26. Profiler PlaFRIM Dynamic instrumentation - valgrind Court` es L., Ru´ e F. valgrind –tool=massif –time-unit=ms ./bin/wave0 5 5 5 Introduction 100 100 100 0.0005 50 General Exploration ms print massif.out.%pid The Roofline model Performance Methodology

  27. Profiler PlaFRIM Dynamic instrumentation - valgrind Court` es L., Ru´ e F. valgrind –tool=massif –time-unit=ms ./bin/wave0 5 5 5 Introduction 100 100 100 0.0005 50 General ms print massif.out.%pid Exploration The Roofline model Performance Methodology

  28. Profiler PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline model what kind of expertise ? Performance Methodology

  29. Profiler PlaFRIM Court` es L., what kind of image of your program do you need ? Ru´ e F. Introduction General Exploration The Roofline model Performance Methodology

  30. Profiler PlaFRIM Court` es L., what kind of image of your program do you need ? Ru´ e F. Introduction General Exploration The Roofline model Performance Methodology

  31. The Roofline model PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline model roofline Performance Methodology

  32. The model PlaFRIM cache aware roofline model Court` es L., Ru´ e F. Introduction General Exploration The Roofline model Performance Methodology Figure: IBM - ICSC 2014, Shanghai, China

  33. The model PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline model Performance Methodology Figure: PICSAR Project

  34. The model PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline model Performance Methodology Figure: Thomas Jefferson National Accelerator Facility

  35. The model PlaFRIM cache aware roofline model Court` es L., Ru´ e F. Introduction General Exploration The Roofline model Performance Methodology Figure: IBM - ICSC 2014, Shanghai, China

  36. The model PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline How to construct this model ? model Performance Methodology

  37. The model PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline How to construct this model ? model Performance How to evaluate your Arithmetic Intensity ? Methodology

  38. Roofline evaluation PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline model ... Performance Methodology

  39. Roofline evaluation PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline model evaluate the performance you can achieve Performance Methodology

  40. PlaFRIM Court` es L., Ru´ e F. Performance achievement Introduction Understanding memory locality General Exploration The Roofline model Performance Methodology Figure: Memory Bound Figure: Compute Bound

  41. Intel Advisor tool PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline model one tool to do that ... Performance Methodology

  42. Intel Advisor tool PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline model Performance Methodology Figure: The 3D stencil: its memory access pattern (a) and the data points it uses (b). - Raul de la Cruz, BSC

  43. Intel Advisor tool PlaFRIM Court` es L., Ru´ e F. Introduction General Exploration The Roofline model Performance Methodology Figure: Stencil 1 thread - roofline

  44. Intel Advisor tool PlaFRIM Court` es L., Ru´ e F. Introduction General module load compiler/gcc/9.1.0 compiler/intel/2019 update4 Exploration intel/vtune-advisor The Roofline model advixe-cl -collect roofline –project-dir=wave0 –ignore-checksums Performance ./bin/wave0 5 5 5 100 100 100 0.0005 500 Methodology advixe-gui

Recommend

More recommend