[ 
https://issues.apache.org/jira/browse/YARN-8743?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Hu Ziqian updated YARN-8743:
----------------------------
    Description: 
capacity scheduler doesn't set node label when new a reserved container's 
RMContainerImpl. When allocate this container, leafQueue will treat it as a 
ignorePartitionExclusivityRMContainer.

It will cause preempt failure. When preempt happens, the preemption policy will 
try to preempt the reserved container while leafQueue doesn't remove it from 
ignorePartitionExclusivityRMContainers. In our cluster, we found that 
preemption policy will always try to preempt the reserved container and 
actually preempt nothing.

We set the node label information to  reserved container's RMContainerImpl and 
redo our test. The preemption performs as expected.

  was:
capacity scheduler doesn't set node label when new a reserved container's 
RMContainerImpl and when allocate this container, leafQueue will treat it as a 
ignorePartitionExclusivityRMContainer.

This bug will cause preempt failure. When preempt happens, the preemption 
policy will try to preempt the reserved container while leafQueue doesn't 
remove it from ignorePartitionExclusivityRMContainers. In our cluster, we found 
that preemption policy will always try to preempt the reserved container and 
actually preempt nothing.

We set the node label information to  reserved container's RMContainerImpl and 
redo our test. The preemption performs as expected.


> capacity scheduler doesn't set node label to reserved container
> ---------------------------------------------------------------
>
>                 Key: YARN-8743
>                 URL: https://issues.apache.org/jira/browse/YARN-8743
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: capacity scheduler, scheduler preemption
>            Reporter: Hu Ziqian
>            Priority: Major
>
> capacity scheduler doesn't set node label when new a reserved container's 
> RMContainerImpl. When allocate this container, leafQueue will treat it as a 
> ignorePartitionExclusivityRMContainer.
> It will cause preempt failure. When preempt happens, the preemption policy 
> will try to preempt the reserved container while leafQueue doesn't remove it 
> from ignorePartitionExclusivityRMContainers. In our cluster, we found that 
> preemption policy will always try to preempt the reserved container and 
> actually preempt nothing.
> We set the node label information to  reserved container's RMContainerImpl 
> and redo our test. The preemption performs as expected.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org

Reply via email to