Skip to content

coll: update the json selection of MPIR_Bcast_intra_scatter_ring_allgather#7332

Draft
hzhou wants to merge 2 commits intopmodels:mainfrom
hzhou:2503_csel_perproc
Draft

coll: update the json selection of MPIR_Bcast_intra_scatter_ring_allgather#7332
hzhou wants to merge 2 commits intopmodels:mainfrom
hzhou:2503_csel_perproc

Conversation

@hzhou
Copy link
Copy Markdown
Contributor

@hzhou hzhou commented Mar 12, 2025

Pull Request Description

The MPIR_Bcast_intra_scatter_ring_allgather won't perform if the
per_proc_msg_size (chunk size) is too small, which accumulates latency
in each round.

Fixes #7330
[skip warnings]

Author Checklist

  • Provide Description
    Particularly focus on why, not what. Reference background, issues, test failures, xfail entries, etc.
  • Commits Follow Good Practice
    Commits are self-contained and do not do two things at once.
    Commit message is of the form: module: short description
    Commit message explains what's in the commit.
  • Passes All Tests
    Whitespace checker. Warnings test. Additional tests via comments.
  • Contribution Agreement
    For non-Argonne authors, check contribution agreement.
    If necessary, request an explicit comment from your companies PR approval manager.

hzhou added 2 commits March 12, 2025 17:23
For some algorithm, e.g. MPIR_Bcast_intra_scatter_ring_allgather, what
matters is not message size, but message size per process, since that
determines whether each round is bandwidth-bound.
The MPIR_Bcast_intra_scatter_ring_allgather won't perform if the
per_proc_msg_size (chunk size) is too small, which accumulates latency
in each round.
@hzhou hzhou requested a review from mjwilkins18 March 12, 2025 22:51
@mjwilkins18
Copy link
Copy Markdown
Contributor

Did you do a performance test with this, or are these changes just based on the conversation earlier this week? Its on my TODOs to do some performance runs on Aurora, so I can try testing this too

@hzhou
Copy link
Copy Markdown
Contributor Author

hzhou commented Mar 13, 2025

Did you do a performance test with this, or are these changes just based on the conversation earlier this week? Its on my TODOs to do some performance runs on Aurora, so I can try testing this too

Thanks for volunteering! :)

The patch is to address the obvious issue so it don't perform outrageously bad. Yes, we should use tests to finetune the threshold.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

perf: MPI_Bcast show significant variance (default is very bad)

2 participants