[CELEBORN-1811] Update default value for `celeborn.master.slot.assign…

….extraSlots` ### What changes were proposed in this pull request? To avoid possible worker load skew for the stages with tiny reducer numbers. ### Why are the changes needed? If a stage has tiny reducers and skewed partitions, The default value will lead to serious worker load imbalance cause some workers unable to handle shuffle data. ### Does this PR introduce _any_ user-facing change? Yes ### How was this patch tested? GA and cluster test. Closes #3039 from FMX/1811. Authored-by: mingji <[email protected]> Signed-off-by: SteNicholas <[email protected]>
apache · Dec 31, 2024 · 4ec0228 · 4ec0228
1 parent 56019c7
commit 4ec0228
Show file tree

Hide file tree

Showing 3 changed files with 4 additions and 2 deletions.
diff --git a/common/src/main/scala/org/apache/celeborn/common/CelebornConf.scala b/common/src/main/scala/org/apache/celeborn/common/CelebornConf.scala
@@ -2883,7 +2883,7 @@ object CelebornConf extends Logging {
       .version("0.3.0")
       .doc("Extra slots number when master assign slots.")
       .intConf
-      .createWithDefault(2)
+      .createWithDefault(100)
 
   val MASTER_SLOT_ASSIGN_MAX_WORKERS: ConfigEntry[Int] =
     buildConf("celeborn.master.slot.assign.maxWorkers")

diff --git a/docs/configuration/master.md b/docs/configuration/master.md
@@ -72,7 +72,7 @@ license: |
 | celeborn.master.port | 9097 | false | Port for master to bind. | 0.2.0 |  | 
 | celeborn.master.rackResolver.refresh.interval | 30s | false | Interval for refreshing the node rack information periodically. | 0.5.0 |  | 
 | celeborn.master.send.applicationMeta.threads | 8 | false | Number of threads used by the Master to send ApplicationMeta to Workers. | 0.5.0 |  | 
-| celeborn.master.slot.assign.extraSlots | 2 | false | Extra slots number when master assign slots. | 0.3.0 | celeborn.slots.assign.extraSlots | 
+| celeborn.master.slot.assign.extraSlots | 100 | false | Extra slots number when master assign slots. | 0.3.0 | celeborn.slots.assign.extraSlots | 
 | celeborn.master.slot.assign.loadAware.diskGroupGradient | 0.1 | false | This value means how many more workload will be placed into a faster disk group than a slower group. | 0.3.0 | celeborn.slots.assign.loadAware.diskGroupGradient | 
 | celeborn.master.slot.assign.loadAware.fetchTimeWeight | 1.0 | false | Weight of average fetch time when calculating ordering in load-aware assignment strategy | 0.3.0 | celeborn.slots.assign.loadAware.fetchTimeWeight | 
 | celeborn.master.slot.assign.loadAware.flushTimeWeight | 0.0 | false | Weight of average flush time when calculating ordering in load-aware assignment strategy | 0.3.0 | celeborn.slots.assign.loadAware.flushTimeWeight | 

diff --git a/docs/migration.md b/docs/migration.md
@@ -23,6 +23,8 @@ license: |
 
 # Upgrading from 0.5 to 0.6
 
+- Since 0.6.0, Celeborn changed the default value of `celeborn.master.slot.assign.extraSlots` from `2` to `100`, which means Celeborn will involve more workers in offering slots.
+
 - Since 0.6.0, Celeborn deprecate `celeborn.worker.congestionControl.low.watermark`. Please use `celeborn.worker.congestionControl.diskBuffer.low.watermark` instead.
 
 - Since 0.6.0, Celeborn deprecate `celeborn.worker.congestionControl.high.watermark`. Please use `celeborn.worker.congestionControl.diskBuffer.high.watermark` instead.