NIXL/UCX threadpool - #573
Merged
Merged
Conversation
|
👋 Hi iyastreb! Thank you for contributing to ai-dynamo/nixl. Your PR reviewers will review your contribution then trigger the CI to test your changes. 🚀 |
mkhazraee
reviewed
Jul 15, 2025
mkhazraee
reviewed
Jul 15, 2025
Pyngon
reviewed
Jul 16, 2025
Pyngon
reviewed
Jul 17, 2025
iyastreb
force-pushed
the
nixl-ucx-workers-pool
branch
from
July 18, 2025 04:52
9ede329 to
264962e
Compare
Contributor
Author
|
Tests fail because of using UCX 1.18, should be fixed by #673 |
tvegas1
reviewed
Aug 11, 2025
tvegas1
reviewed
Aug 11, 2025
mkhazraee
reviewed
Aug 11, 2025
mkhazraee
previously approved these changes
Aug 12, 2025
brminich
reviewed
Aug 13, 2025
rakhmets
reviewed
Aug 14, 2025
brminich
previously approved these changes
Aug 15, 2025
Signed-off-by: Ilia Yastrebov <iyastrebov@nvidia.com>
Contributor
|
/build |
brminich
approved these changes
Aug 15, 2025
Contributor
|
/build |
1 task
This branch was previously deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
7 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What?
Enable threadpool in NIXL/UCX to parallelise posting of large requests (in terms of batch size) .
To enable this feature client needs to set "num_threads" in UCX backend params (close to "num_workers").
This essentially specifies the number of dedicated workers and threads (1 worker per thread) that are used to post requests and progress.
Progress thread logic remains the same as before: if configured, a single progress thread is used for all shared workers. There must be at least one shared worker configured in any use case.
Why?
Clients report long posting time with large batch sizes (> 1000).
For example, posting of 32k batch size with cuda_ipc protocol being used takes up to 0.25s.
With this fix we can drastically reduce the time needed to post the request.
How?
This PR depends on refactoring PR: #606 (merged)
It depends also on PR #673, which bumps UCX from 1.18 to 1.19
The performance evaluation
Summary:
With communication over IB devices, using multiple threads for posting+progress makes a huge difference in terms of request posting time, and also improves BW.
E.g. on 64k message size, the posting time is reduced by 8x, and BW is 10% better
With cuda_ipc communication I see not so huge decrease of posting time, it's basically around 30%, but BW is also 30% better.