mridulm commented on code in PR #2119:
URL:
https://github.com/apache/incubator-celeborn/pull/2119#discussion_r1413334606
##########
master/src/main/scala/org/apache/celeborn/service/deploy/master/Master.scala:
##########
@@ -666,23 +666,32 @@ private[celeborn] class Master(
val shuffleKey = Utils.makeShuffleKey(requestSlots.applicationId,
requestSlots.shuffleId)
val availableWorkers = workersAvailable()
+ // reply false if all workers are unavailable
+ if (availableWorkers.isEmpty) {
+ logError(
+ s"Non available workers, offer slots for $numReducers reducers of
$shuffleKey failed!")
+ context.reply(RequestSlotsResponse(StatusCode.SLOT_NOT_AVAILABLE, new
WorkerResource()))
+ return
+ }
+
val numAvailableWorkers = availableWorkers.size()
val numWorkers = Math.min(
Math.max(
if (requestSlots.shouldReplicate) 2 else 1,
if (requestSlots.maxWorkers <= 0) slotsAssignMaxWorkers
else Math.min(slotsAssignMaxWorkers, requestSlots.maxWorkers)),
numAvailableWorkers)
+
+ // We treated availableWorkers as a Circular Queue here.
val startIndex = Random.nextInt(numAvailableWorkers)
+ val endIndex = startIndex + numWorkers
+
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))
+ for (index <- startIndex until endIndex) {
+ val realIndex = index % numAvailableWorkers
+ selectedWorkers.add(availableWorkers.get(realIndex))
}
+
Review Comment:
> so that the very small performance gap is acceptable from my side, and
the focus of this patch is to improve readability.
While readability and maintainability are very important, there is nothing
inherently unmaintainable about using `subList`.
IMO performance is also important - atleast for our deployment, where we are
running millions of shuffles a day for PB's of data, it is :-)
> we can see that there is a similar algorithm in the
`Master#handleRequestSlots`
There the subList implementation would not work - it is inserting a tuple
into slots map based on the worker, while here it is a bulk addition.
--
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]
