Extend flamegraphs to show alloc diff #238
Conversation
LenkaNovak
force-pushed
the
ln/flamegraph-diff
branch
from
513b13b to
4275122
Compare
LenkaNovak
force-pushed
the
ln/flamegraph-diff
branch
from
0198c6f to
a772021
Compare
docs/src/performance.md
Outdated
| @@ -11,6 +11,7 @@ | |||
|  | |||
|
|
|||
| - each row along the y-axis represents a level of backtraces. In this case the lowermost level is at the top, and the top level represents what is directly being run on the CPU. The stacks in each level are sorted alphabetically (not chronologically, like flame _charts_). The column width is proportional to the presence in samples (related to allocations). The colors are grouped into runtime-dispatch, gc, compilation and default. The intensity is random. | |||
| - we also have a local beta version of flame graphs (in `perf/ProfileCanvasDiff.jl` and `perf/ProfileViewerDiff.js`), which plots the same flame graphs as above but with the color corresponding to whether the stack allocation has been reduced (blue) or increased (red) compared to the last staged runs. The color intensity is proportional to the fractional change, and black signifies untracked traces. | |||
There was a problem hiding this comment.
It could be helpful to explain how to run the version of flame graphs with colors showing stack allocation change vs showing allocation type
| @@ -0,0 +1,390 @@ | |||
| # temporarily copied and modified from https://github.com/pfitzseb/ProfileCanvas.jl | |||
There was a problem hiding this comment.
instead of copying the code over, could you use the ProfileCanvas package and extend any functions you need to behave differently?
There was a problem hiding this comment.
Yeah, totally. We need to decide if we're going to keep this format when we generalize this for the other repos. If so, then we'll definitely condense this like you suggest. I've made a separate issue for this final format #245 , but since this may take a while to figure out, it would be useful to merge this prototype ASAP, so we can start tracking staged runs.
transfer + modify from ClimaCore/ClimaAtmos add alloc benchmarks add buildkite driver rev allocs docs activate docs clean fix doc fix doc fix check in perf Manifest bkite flag order fix scope fix FT redef error fix fix stash fix toy example runs prep for buildkite staging test buildkite driver format bug plot in bkite fix fix module path bkite build path switch from toy to coupler switch from toy to coupler rebase test Manifest up prep dummy benchmark fix original ProfileCanvas, ProfileViewer files diff format diff race and docs race and docs permissions restart restart restart revs
LenkaNovak
force-pushed
the
ln/flamegraph-diff
branch
from
a772021 to
e10619e
Compare
|
Discussed and approved offline. Merging this for immediate tracking and will follow up on the interface refactor in #245. |
|
bors r+ |
Experimental PR that test the ease and usefulness of color coding flame graphs based on allocation changes compared to the previous commit, using a local testing bed.
To-do
Content
Note: For the functions that were directly copied from ProfileCanvas.jl and ProfileView.js, we haven't added any new doc strings/CI tests yet, because Simon and I are still figuring out if we're going to integrate this eventually to the original packages. This PR is mainly a proof of concept, and if we decide to keep it long term, we'll clean it up to be more accessible for users. In the meantime though, it would be good to have this merged and available ASAP for our own perf monitoring purposes.