Massive trace file?

[indigoviolet]

Hi Kunal,

Thanks for this library.

I'm trying it out to trace some pytorch code after striking out on pytracing and functiontrace.

One thing I noticed was that it created a massive trace file: 5.1GB that gzipped to 300M. Is that expected? (the other libraries created much smaller files). My chrome://tracing failed to load that file...

If that's unavoidable, would it make sense to have some way of restricting the depth of the trace - for example I would care about user code, imported libraries, python stdlib, python core code in decreasing order - so if I could only trace the first two that might help?

Cheers.

[kunalb]

Thanks for trying it out! Hmm, is the code shareable to see what you're doing? -- I would have expected similar order of magnitude compared to other tracers. The main additional thing it'll be capturing is connections between async function calls.

chrome://tracing will start dieing around ~1g of trace.

I have filtering planned, but no real urgency for it -- I'll move that up and implement it so that you can choose which frames show up in traces.

In the meantime, is it feasible for you to use @Probe instead to only instrument functions you explicitly care for / or use a context manager to isolate what you're digging into?

[kunalb]

(Another thing I've done in the past when I really needed to see a massive trace is to split it into multiple traces by just cutting it at arbitrary points, then loading the intermediate pieces. But this is a very painful workflow.)

[indigoviolet]

I can generate and share a subset of the trace if you think that'll be helpful - the code itself isn't particularly clever or unusual, a modified version of https://detectron2.readthedocs.io/ . I'm tracing the model training for ~20 iterations. No threading or asyncio or stuff like that.

Re @Probe, I've had some success using that same approach with https://region-profiler.readthedocs.io/en/latest/, but it's obviously more work to do it upfront.

[kunalb]

If you pull master you should be able to do something like:

with AsyncioTracer(trace, skip=stdlib()):
    ...

or

with AsyncioTracer(trace, skip=not_(module_equals("yourmodulename"))):
   ...

[kunalb]

PS. how are you finding all these tracers? I spent quite a bit of time before starting panopticon, and found less than half of the ones you have.

[indigoviolet]

Awesome, thank you so much for working on this so quickly. I'll give it a shot soon and let you know.

PS. how are you finding all these tracers?

haha - just googling I guess. python chrome tracing etc - in fact I found one of your comments on a Functiontrace HN thread that lead me here, though I had earlier seen an FB post about it.