[
https://issues.apache.org/jira/browse/YARN-6081?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15816990#comment-15816990
]
Wangda Tan commented on YARN-6081:
----------------------------------
This is the test case to reproduce the problem:
{code}
@Test
public void testPreemptionNotHappenForSingleReservedQueue() {
Logger rootLogger = LogManager.getRootLogger();
rootLogger.setLevel(Level.DEBUG);
int[][] qData = new int[][]{
// / A B C
{ 100, 40, 40, 20 }, // abs
{ 100, 100, 100, 100 }, // maxCap
{ 100, 70, 0, 0 }, // used
{ 10, 30, 0, 0 }, // pending
{ 0, 50, 0, 0 }, // reserved
{ 1, 1, 0, 0 }, // apps
{ -1, 1, 1, 1 }, // req granularity
{ 3, 0, 0, 0 }, // subqueues
};
ProportionalCapacityPreemptionPolicy policy = buildPolicy(qData);
policy.editSchedule();
// ensure all pending rsrc from A get preempted from other queues
verify(mDisp, times(0)).handle(argThat(new IsPreemptionRequestFor(appA)));
}
{code}
Please note that there's only one active queue. But preemption policy still
preempt container from it.
> LeafQueue#getTotalPendingResourcesConsideringUserLimit should deduct reserved
> from pending to avoid unnecessary preemption of reserved container
> ------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: YARN-6081
> URL: https://issues.apache.org/jira/browse/YARN-6081
> Project: Hadoop YARN
> Issue Type: Bug
> Reporter: Wangda Tan
> Assignee: Wangda Tan
> Priority: Critical
>
> While doing YARN-5864 tests, found an issue when a queue's reserved >
> pending. PreemptionResourceCalculator will preempt reserved container even if
> there's only one active queue in the cluster.
> To fix the problem, we need to deduct reserved from pending when getting
> total-pending resource for LeafQueue.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
