feat: enforce per-node move limit rather than per-node requeue limit#51
Merged
adsharma merged 1 commit intoJul 2, 2026
Merged
Conversation
Contributor
|
Agreed! |
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
2 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
An extension to #47. In that PR, a per-node requeue limit was implemented. Note the difference between a per-node move limit: a node can be requeued many times without successfully moving. Further benchmarking show that a per-node requeue limit has high variance in its runtime, which may be due to the randomness of node scheduling. Since the variance does seem concerning (i.e., that there exists a scheduling path that may avoid a lot of useful work), I propose we switch to a per-node move limit (originally suggested by @adsharma). The below graph shows that variance is less with this change, achieved CPM is higher, but runtime is higher. Although runtime is higher, it is still considerably bounded (see the "no limit" data point):
In some cases with per-node requeue limit, we saw that runtime can differ by a factor of 2x-3x. A preliminary look at the quality of each produced cluster suggest that cluster quality is degraded in the low-runtime case (e.g., min-cut is lower).
Summary: A per-node move limit is more reliable than a per-node requeue limit, since it prevents cases where a lot of work is skipped (through the randomness of node scheduling). In cases where work is skipped, we believe cluster quality is worse.