[Core][Distributed] use device group for all broadcast #5320
Conversation
This file contains 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
|
@youkaichao Could you add some performance numbers on H100? I'm wondering how this affects the performance. |
|
see https://docs.google.com/spreadsheets/d/1c9xgR0fGvm6SROfk7vrjwOZdYnKQk9oOafWK4_KgOyo/edit#gid=593626425 for reference. in particular, check the |
|
close as #5399 will be better. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
A fix to #4444
That PR was originally tested on A100, and it showed some speedup.
However, later it seems to slow down on H100 and other machines. The hypothesis is that
glooperforms poorly while nvlink is better in these high-end machines.Before we find a good solution, we can just use device group for all the broadcast.
TODO:
investigate if it is possible and beneficial to only communicate cpu data, using mechanisms such as message queue.