Use multi_get for store that has extended API support. #408
Conversation
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
Summary: All credit goes to original author XilunWu. I am just landing the code to unblock large Ads jobs. D45740631 reduces gloo rendezvous cost for TCP backend by eliminating duplicate address publishing to TCPStore. Ben suggested "have seq_number == global_rank" to further get rid of `seq_number` exchange and Shawn reported why this didn't work. This diff serves a starting point for benchmarking the benefit of doing so ([testbed record](https://docs.google.com/document/d/1_p390fx0IiaZWbt-Dkdvp9jSgCKebiuG8_a6BVjHtMU/edit) shows 2x speedup: 46 min ProcessGroupGloo init time on 8k ranks -> 20 min). The feature will be enabled via env variable (GLOO_ENABLE_RANK_AS_SEQUENCE_NUMBER) disabled by default that will be controlled by justKnobs. Differential Revision: D48130088
|
This pull request was exported from Phabricator. Differential Revision: D52083376 |
fduwjj
approved these changes
Feb 11, 2025
Summary: Pull Request resolved: pytorch#408 The TCP store has API v2 support we can reduce the network overhead of Gloo rendezvous significantly by fetching a batch of key instead of doing them one by one. Initial testing shows ~15X improvement for 4k jobs. Gloo process group init: Baseline ( fbcode trunk): 2k job (https://fburl.com/mlhub/x1prxu89) : ~82sec (~1.4 min) 4k job (https://fburl.com/mlhub/v1djk4n5) : ~393 sec (~6.6min) 8k job (https://fburl.com/mlhub/cagqrs7m): (~55mins) With optimizations (D48130088 + D52083376): 2k job (https://fburl.com/mlhub/x0cskdag) : ~18 sec ( ~5x faster) 4k job (https://fburl.com/mlhub/xzmvkm4j) : ~ 25 sec (~15x faster) 8k job (https://fburl.com/mlhub/gdyeizv9) : ~ 85 sec (~35x faster) Reviewed By: xunnanxu Differential Revision: D52083376
|
This pull request was exported from Phabricator. Differential Revision: D52083376 |
XilunWu
force-pushed
the
export-D52083376
branch
from
b223c31 to
06f24f5
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
Summary:
The TCP store has API v2 support we can reduce the network overhead of Gloo rendezvous significantly by fetching a batch of key instead of doing them one by one.
Initial testing shows ~15X improvement for 4k jobs.
Gloo process group init:
Baseline ( fbcode trunk):
2k job (https://fburl.com/mlhub/x1prxu89) : ~82sec (~1.4 min)
4k job (https://fburl.com/mlhub/v1djk4n5) : ~393 sec (~6.6min)
8k job (https://fburl.com/mlhub/cagqrs7m): (~55mins)
With optimizations (D48130088 + D52083376):
2k job (https://fburl.com/mlhub/x0cskdag) : ~18 sec ( ~5x faster)
4k job (https://fburl.com/mlhub/xzmvkm4j) : ~ 25 sec (~15x faster)
8k job (https://fburl.com/mlhub/gdyeizv9) : ~ 85 sec (~35x faster)
Reviewed By: xunnanxu
Differential Revision:
D52083376
Privacy Context Container: L1156430