DL profiling 도구들

도구 리스트

1. PyTorch Profiler
2. NVIDIA Nsight System
3. DLProf
4. PyProf

Detail

1. PyTorch Profiler

• Typical steps to optimize your model
  

• PyTorch Profiler의 profiling 결과는 TensorBoard에서 확인할 수 있다.

• PyTorch Profiler의 버전 별 feature
  

• Memory view
  

• Overview
  

• Trace
  

  • SM Efficiency?: Streaming Multiprocessor Efficiency
  • Streaming Multiprocessor?: GPU를 구성하는 Processor

2. Nvidia Nsight System

• PyTorch, TensorRT 모두 지원하는 듯?

3. DLProf

• Nvidia에서 User Guide 제공
  https://docs.nvidia.com/deeplearning/frameworks/dlprof-user-guide/index.html
• PyTorch, TensorRT 모두 지원
• Nvidia Nsight System의 wrapper 인 듯?

4. PyProf

• 잘 안쓰이는 것 같음, Pass

References

• https://www.youtube.com/watch?v=8Rr_8jy1_GY
• https://on-demand.gputechconf.com/ai-conference-2019/skr9110.pdf
• https://docs.nvidia.com/deeplearning/frameworks/dlprof-user-guide/index.html
• https://tigress-web.princeton.edu/~jdh4/how_to_profile_with_dlprof_may_2021.pdf
• https://developer.nvidia.com/blog/profiling-and-optimizing-deep-neural-networks-with-dlprof-and-pyprof/
• https://guide.ncloud-docs.com/docs/ai-clova-nsml-high-3-1