xleoken commented on code in PR #2119:
URL:
https://github.com/apache/incubator-celeborn/pull/2119#discussion_r1458669556
##########
master/src/main/scala/org/apache/celeborn/service/deploy/master/Master.scala:
##########
@@ -678,16 +679,9 @@ private[celeborn] class Master(
if (requestSlots.maxWorkers <= 0) slotsAssignMaxWorkers
else Math.min(slotsAssignMaxWorkers, requestSlots.maxWorkers)),
numAvailableWorkers)
- val startIndex = Random.nextInt(numAvailableWorkers)
- val selectedWorkers = new util.ArrayList[WorkerInfo](numWorkers)
- selectedWorkers.addAll(availableWorkers.subList(
- startIndex,
- Math.min(numAvailableWorkers, startIndex + numWorkers)))
- if (startIndex + numWorkers > numAvailableWorkers) {
- selectedWorkers.addAll(availableWorkers.subList(
- 0,
- startIndex + numWorkers - numAvailableWorkers))
- }
+ Collections.rotate(availableWorkers, Random.nextInt(numAvailableWorkers))
Review Comment:
> Maybe also no need to do the rotate?
Yeah, if remove the rotate, the slots offering could not be balanced in all
workers.
E.g, if we have 10 workers, and each shuffle request just need 2 workers,
the top two workers will be choosed always.
##########
master/src/main/scala/org/apache/celeborn/service/deploy/master/Master.scala:
##########
@@ -678,16 +679,9 @@ private[celeborn] class Master(
if (requestSlots.maxWorkers <= 0) slotsAssignMaxWorkers
else Math.min(slotsAssignMaxWorkers, requestSlots.maxWorkers)),
numAvailableWorkers)
- val startIndex = Random.nextInt(numAvailableWorkers)
- val selectedWorkers = new util.ArrayList[WorkerInfo](numWorkers)
- selectedWorkers.addAll(availableWorkers.subList(
- startIndex,
- Math.min(numAvailableWorkers, startIndex + numWorkers)))
- if (startIndex + numWorkers > numAvailableWorkers) {
- selectedWorkers.addAll(availableWorkers.subList(
- 0,
- startIndex + numWorkers - numAvailableWorkers))
- }
+ Collections.rotate(availableWorkers, Random.nextInt(numAvailableWorkers))
Review Comment:
As @waitinfuture mentioned in
https://github.com/apache/incubator-celeborn/pull/1975, Master just randomly
picks a range of available workers.
##########
master/src/main/scala/org/apache/celeborn/service/deploy/master/Master.scala:
##########
@@ -678,16 +679,9 @@ private[celeborn] class Master(
if (requestSlots.maxWorkers <= 0) slotsAssignMaxWorkers
else Math.min(slotsAssignMaxWorkers, requestSlots.maxWorkers)),
numAvailableWorkers)
- val startIndex = Random.nextInt(numAvailableWorkers)
- val selectedWorkers = new util.ArrayList[WorkerInfo](numWorkers)
- selectedWorkers.addAll(availableWorkers.subList(
- startIndex,
- Math.min(numAvailableWorkers, startIndex + numWorkers)))
- if (startIndex + numWorkers > numAvailableWorkers) {
- selectedWorkers.addAll(availableWorkers.subList(
- 0,
- startIndex + numWorkers - numAvailableWorkers))
- }
+ Collections.rotate(availableWorkers, Random.nextInt(numAvailableWorkers))
Review Comment:
> When allocate, it will start from a random index of worker. IMO it's fine.
Refer to
https://github.com/apache/incubator-celeborn/blob/main/master/src/main/java/org/apache/celeborn/service/deploy/master/SlotsAllocator.java#L339
Right, the SlotsAllocator use a random index of workers, it start from
selected workers.
If the client set the num of `requestSlots.maxWorkers` less then the num of
all wokers, the master will always choose the top workers.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
