fix process staggering logic to smooth memory spikes #451
+22
−18
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
zfan3-sc
requested review from
kmontemayor2-sc,
mkolodner-sc,
nshah-sc,
svij-sc,
xgao4-sc and
yliu2-sc
as code owners
zfan3-sc
force-pushed
the
zfan3/stagger
branch
from
64a62ef to
e042757
Compare
Collaborator
kmontemayor2-sc
left a comment
kmontemayor2-sc
left a comment
There was a problem hiding this comment.
Hmm, for GraphStore (server-client) mode this may still be an issue but also unforuntately need to launch the samplers in lock-step.
Can you add a TODO (assigned to me) around python/gigl/distributed/distributed_neighborloader.py:340 to look into this?
Collaborator
Author
|
Added the TODO |
zfan3-sc
force-pushed
the
zfan3/stagger
branch
from
0127e87 to
83af1b2
Compare
zfan3-sc
force-pushed
the
zfan3/stagger
branch
from
83af1b2 to
77d52f4
Compare
kmontemayor2-sc
approved these changes
Jan 21, 2026
Collaborator
kmontemayor2-sc
left a comment
kmontemayor2-sc
left a comment
There was a problem hiding this comment.
LGTM, you'll need a second stamp as this is an OSS repo :)
mkolodner-sc
approved these changes
Jan 21, 2026
Collaborator
mkolodner-sc
left a comment
mkolodner-sc
left a comment
There was a problem hiding this comment.
Thanks a lot Zihao!
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Outdated suggestions cannot be applied.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.
Scope of work done
The current implementation for staggering processes during data loader initialization, which was intended to smooth out memory spikes, does not have the desired effect.
There are two main reasons:
DistSamplingProducerinitialization).torch.distributed.broadcast_object_listcall synchronizes all processes, effectively negating the staggered sleep.In this PR, we move the sleep logic out of init_neighbor_loader_worker and place it directly before
DistSamplingProducerinitialization, or before the parentDistLoaderclass’s__init__which invokes it. This ensures the stagger occurs at the correct point and is no longer nullified by distributed synchronization.Where is the documentation for this feature?: N/A
Did you add automated tests or write a test plan?
Updated Changelog.md? NO
Ready for code review?: NO