1Tracy: a profiler you don’t want to miss Marcos Slomp Senior Research Engineer IIWhy Tracy? Real-time workflow (analyze application while it is running) Precise measurements (nanosecond resolution) com/dfeneyrou/palanteer Orbit Profiler https://github.com/google/orbit SuperLuminal https://superluminal.eu Perfetto https://perfetto.dev MicroProfiler https://github.com/tyoma/micro-profiler EasyProfiler https://github https://github.com/yse/easy_profiler Oprofile https://oprofile.sourceforge.io/examples/ Coz https://github.com/plasma-umass/coz VerySleepy https://github.com/VerySleepy/verysleepy LukeStackwalker http://lukestackwalker

1Tracy: a profiler you don’t want to miss Marcos Slomp Senior Research Engineer IIWhy Tracy? Real-time workflow (analyze application while it is running) Precise measurements (nanosecond resolution) com/dfeneyrou/palanteer Orbit Profiler https://github.com/google/orbit SuperLuminal https://superluminal.eu Perfetto https://perfetto.dev MicroProfiler https://github.com/tyoma/micro-profiler EasyProfiler https://github https://github.com/yse/easy_profiler Oprofile https://oprofile.sourceforge.io/examples/ Coz https://github.com/plasma-umass/coz VerySleepy https://github.com/VerySleepy/verysleepy LukeStackwalker http://lukestackwalker

allocations and containers • Optimised, production applications in production environments • Dynamic Profiler • Can be used with no recompilation or code modification • APIsCode Generation Outline Source Reconstruction • Code Generation • Object Introspection as a Library • Object Introspection as a Profiler • Object Introspection Applied Type Reconstruction Object Introspection• (Re)construct an Reconstruction • Code Generation • Object Introspection as a Library • Object Introspection as a Profiler • Object Introspection Applied Type Reconstruction Object Introspection Measurement Code

AI-powered optimizations!AI-powered optimizations... COMING SOON!evolveevolve profiler that suggests optimizationsevolve profiler that suggests optimizations "closes loop" between tool and human's goalevolve efficient maybe, you figure it out "closes loop" between tool and human's goal profiler that suggests optimizationsevolve profiler that suggests optimizations "closes loop" between tool and human's goal here's some beer profiler knows why code is inefficient exploit nicheprofiler knows where code is inefficient I figured out where your code could be more efficient so you go get some beer profiler knows why

4445 Memory Profiler I couldn’t use standard memory profilers as the Linux kernel was old. Gave me a good understanding of the sizes and the quantities of allocations.MEMORY PROFILER 46 Memory Time 18 MByteMEMORY PROFILER 47 Memory Time 18 MByte 20 MByteMEMORY PROFILER 48 ⊡ Number of allocations per second. ⊡ Current and maximum number of allocations. ⊡ Current and maximum bytes allocated allocated. ⊡ Current and maximum allocations per size value. Interesting measurements:MEMORY PROFILER 49 ⊡ Number of allocations per second. ⊡ Current and maximum number of allocations. ⊡ Current

notesPROFILING • Many Profilers used in Game Development Tracy, Intel VTune, Microso� PIX, Custom • Need profiler with instrumentation • Need to see whole process and all threads Using Intel’s VTune in this presentation GameDev - custom or 3rd party - many integrated into engine Used to measure CPU, GPU, etc Need profiler with instrumentation Cannot rely on sampling - everything is in jobs Need to see all threads Looking dependencies Using VTune - best view for data building Presentation about Tracy profiler from CppCon 2023 An Introduction to Tracy Profiler in C++ CppCon 2023 Speaker notesTECHNIQUES 14KEEP THREADS BUSY Keep

mixed precision-like API that can be used seamlessly with your PyTorch code. ‣ NVIDIA Deep Learning Profiler (DLProf) v1.8, which was included in the 21.12 container, was the final release of DLProf. Starting mixed precision-like API that can be used seamlessly with your PyTorch code. ‣ NVIDIA Deep Learning Profiler (DLProf) v1.8, which was included in the 21.12 container, was the final release of DLProf. Starting mixed precision-like API that can be used seamlessly with your PyTorch code. ‣ NVIDIA Deep Learning Profiler (DLProf) v1.8, which was included in the 21.12 container, was the final release of DLProf. Starting

技术全景概览 技术细节探索 lPython Profiling的工具链和解决方案 ØCPU: cProfile、Profile、Pyinstrument、line-profiler，py-spy ØMemory : memory-profiler、memray Ø新兴Continuous 方案：Pyroscrope l实践 & 展示 背景概述 • 什么是Profiling？ • 定义： Prof yKernelz{|m}~b•€•stack trace - py-spy instead !"#$:%&'(-Memory wxe tracemalloc ‚ƒ psutil 典型： memory_profiler „ssnapshot Line …†->‡ˆ !"#$:%&'(-Memory ‰cProfile Š‹Œ• wxc-apiePyTracefunc 典型： Memray Python 持续 profiling • 只到function 粒度，函数内部未知 Pyinstrument 89Sampling#:;<= >?@AoverheadBnoise #CD line_profiler 89setTrace#: ;EFProfileG

0 码力 | 28 页 | 12.73 MB | 1 年前

3

分支」 有没有别的办法？ Tracing runtime 有 Profiler Label Go 1.9 is introducing profiler labels, a way to add arbitrary key-values to the samples collected by the CPU profiler. CPU profilers collect and output output hot spots where the CPU spent most time in when executing. A typical CPU profiler output is primarily reports the location of these spots as function name, source file/line, etc. By looking at the context.Context) { // Do some work... go update(ctx) // propagates labels in ctx. }) Profiler Label 思路： ● 在 SQL 解析器完成解析后，生成 SQL 的指纹信息（SQL Digest） ● 生成带着这个 Digest 的 Context ● 使用 pprof.SetGoroutineLables

● Branch predictors are quite complex ● Do not optimize until misprediction is confirmed by a profiler 5x slower 5x slowerBranchless Computing 39 BENCHMARK ● 01c ● with perfBranchless Computing reality: return cond ? x : y; ● Never optimize such code preemptively ● Optimize only if the profiler shows high misprediction rate ● Optimizations depend on the compiler!Branchless Computing 52 But very similar code may not be ● Never optimize such code preemptively ● Optimize only if the profiler shows high misprediction rateBranchless Computing 54 BENCHMARK ● 05a, b – optimization does