Trajopt Profiling #47
Conversation
…for shooting methods
…th static goal and cost matrices
…ta in case we need it later
| "torch>=1.13.1", | ||
| "tensorboard>=2.15.1", | ||
| # "torch>=1.13.1", | ||
| "tensorboard>=2.13.0", # [Dec. 4, 2023] https://github.com/tensorflow/tensorflow/issues/62075#issuecomment-1808652131 |
There was a problem hiding this comment.
[NOTE] As it turns out, if I pin the version to 2.13.0, the code check workflow fails. However, we know that things are broken with the most recent tensorflow version, so we should probably not merge this yet.
| "tensorflow>=2.13.0", # [Dec. 4, 2023] https://github.com/tensorflow/tensorflow/issues/62075#issuecomment-1808652131 | ||
| "tensorboard-plugin-profile>=2.13.0", |
There was a problem hiding this comment.
See above comment about pinning versions.
pculbertson
left a comment
pculbertson
left a comment
There was a problem hiding this comment.
LGTM overall -- have a couple ideas for minor improvements to documentation but we can do this later. Bigger issue is the compatibility stuff you already noted -- not sure how to fix this but would be nice to have profiling functional eventually
| fn1() | ||
| # stuff we do want to specifically profile | ||
| with jax.named_scope("name_of_your_choice"): |
There was a problem hiding this comment.
Quick q: I know TraceAnnotation didn't work here but it's maybe good to note this here since the jax profiling tool recommends it
| ``` | ||
| tensorboard --logdir=/dir/to/profiling/results --port <port> | ||
| ``` | ||
| where `--port` should be some open port like `8008`. In the top right dropdown menu which should say "Inactive," scroll down and select "Profile." Select the run you'd like to analyze and under tools, the most useful tab will usually be "trace_viewer." |
There was a problem hiding this comment.
You might consider adding some screenshots for folks who are unfamiliar with flame charts
Tips on profiling
tensorboard-related dependencies topyproject.tomlProfiling results:
Takeaways from the above results:
stepis simply to use fewer solver iterations and line search iterations._positionsubroutine of the forward dynamics takes about 26% of the total runtime!_accelerationsubroutine takes about 11% of the total runtimeEssentially 100% of the runtime of
forwardis captured by the above operations, which should give some insight into code optimization prioritizes on themjxside if we want to give some feedback. We should remember that the contact engine is DISABLED for this benchmark.