[AURON #2422] make RSS spill memory triggers configurable#2423
Open
cxzl25 wants to merge 1 commit into
Open
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Which issue does this PR close?
Closes #2422
Rationale for this change
Configurable spill trigger (rss_sort_repartitioner.rs). Replace the hardcoded
mem_used_percent() > 0.4with two tunable thresholds; spill fires when either is exceeded (whichever first):spark.auron.rss.spill.memoryFraction(relative, default 0.4 = previous behavior)spark.auron.rss.spill.memorySize(absolute bytes, default 0 = disabled; acts as a deterministic cap on the per-spill burst regardless of executor memory)Both are read once per executor via OnceCell (no per-batch JNI conf reads).
Adds
MemConsumer::mem_used()(absolute bytes accessor) to support the absolute threshold.What changes are included in this PR?
Auron's rss shuffle buffers records in native memory and spills to celeborn in bursts. At a shuffle-stage peak, many tasks spill simultaneously; if each spill is large, the aggregate burst can overrun a celeborn worker's inbound capacity and push the worker into a push-pause: it stops reading, holds inbound buffers, and only recovers after the idle-connection timeout / connection close (observed as a multi-minute stall — tasks survive via celeborn's client-side revive but throughput degrades).
The previous hardcoded 0.4 fraction made the per-spill burst size depend on executor memory and not directly tunable against this.
Are there any user-facing changes?
No
How was this patch tested?
Production environment verification
Was this patch authored or co-authored using generative AI tooling?