Getting access to profiling data.

[Michael Isaev]

Hello, 

I am working on a ParaGraph framework - it is a tool that I started developing while interning at Google and that would allow you to interface TensorFlow applications via HLO with various computer systems/network event driven simulators. To compare its modeling against actual hardware, I am trying to get access to OpStats protobuf (I used to get access to it while at Google through xprof session). I see that it is likely to be used to generate data that profiler currently dumps, i.e. tf_stats.pb, op_profile.json, but OpStats has more useful information for me.

I would imagine there could be a couple ways to get OpStats dumped. First, via writing some app that sends request to profiler server and parses response similar to how it’s done in ConvertXSpaceToProfileResponse. Second, by hacking somewhere along profiler/tensorboard/profiler_plugin lines so more things get dumped. Ideally, I would like to write some tutorial that uses Colab or free GCP instance to demonstrate how my tool could be used to extract trace information from TF models (including running on TPUs) to provide them to computer architecture simulators.

Could you please advice me the best way that could be achieved? I would really appreciate your help.

Thanks,
Michael