[
https://issues.apache.org/jira/browse/YARN-9697?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16956962#comment-16956962
]
Bibin Chundatt commented on YARN-9697:
--------------------------------------
Thank you [~abmodi] for updating patch
Few comments and suggestion
# OpportunisticContainerAllocatorAMService -> NodeQueueLoadMonitor init could
be moved to AbstractService#serviceinit
# NodeQueueLoadMonitor ScheduledExecutorService#scheduledExecutor shutdown not
done
# NodeQueueLoadMonitor#nodeIdsByRack do we need the NodeIds to be sorted ??
# Thoughts on replacing NodeQueueLoadMonitor#addIntoNodeIdsByRack as follows
{code}
private void addIntoNodeIdsByRack(RMNode addedNode) {
nodeIdsByRack.compute(addedNode.getRackName(), (k, v) -> v == null ?
new ConcurrentHashMap<String, NodeId>().newKeySet() :
v).add(addedNode.getNodeID());
}
{code}
# We could think of replacing NodeQueueLoadMonitor#removeFromNodeIdsByRack too
with computeifPresent
Not related to patch
# OpportunisticSchedulerMetrics shouldn't we be having a destroy() method to
reset the counters. During switch over i think we should reset the counters ?
> Efficient allocation of Opportunistic containers.
> -------------------------------------------------
>
> Key: YARN-9697
> URL: https://issues.apache.org/jira/browse/YARN-9697
> Project: Hadoop YARN
> Issue Type: Sub-task
> Reporter: Abhishek Modi
> Assignee: Abhishek Modi
> Priority: Major
> Attachments: YARN-9697.001.patch, YARN-9697.002.patch,
> YARN-9697.003.patch, YARN-9697.004.patch, YARN-9697.005.patch,
> YARN-9697.006.patch, YARN-9697.007.patch, YARN-9697.ut.patch,
> YARN-9697.ut2.patch, YARN-9697.wip1.patch, YARN-9697.wip2.patch
>
>
> In the current implementation, opportunistic containers are allocated based
> on the number of queued opportunistic container information received in node
> heartbeat. This information becomes stale as soon as more opportunistic
> containers are allocated on that node.
> Allocation of opportunistic containers happens on the same heartbeat in which
> AM asks for the containers. When multiple applications request for
> Opportunistic containers, containers might get allocated on the same set of
> nodes as already allocated containers on the node are not considered while
> serving requests from different applications. This can lead to uneven
> allocation of Opportunistic containers across the cluster leading to
> increased queuing time
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
