[
https://issues.apache.org/jira/browse/YARN-1724?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13905039#comment-13905039
]
Junping Du commented on YARN-1724:
----------------------------------
I could miss that JIRA and discussion. My thought here is: given container
requests are coming randomly and continuous scheduling are looping forever, it
didn't give node with more resources more chance to get assigned containers.
The only differences for beginner node of the loop is it gives 5 ms sleep
window (from last iteration). If this is the real reason to cause imbalanced,
may be we should try to remove sleep to achieve more balanced scheduling?
Locking whole scheduler seems expensive to me.
> Race condition in Fair Scheduler when continuous scheduling is turned on
> -------------------------------------------------------------------------
>
> Key: YARN-1724
> URL: https://issues.apache.org/jira/browse/YARN-1724
> Project: Hadoop YARN
> Issue Type: Bug
> Components: scheduler
> Reporter: Sandy Ryza
> Assignee: Sandy Ryza
> Priority: Critical
> Fix For: 2.4.0
>
> Attachments: YARN-1724-1.patch, YARN-1724.patch
>
>
> If nodes resource allocations change during
> Collections.sort(nodeIdList, nodeAvailableResourceComparator);
> we'll hit:
> java.lang.IllegalArgumentException: Comparison method violates its general
> contract!
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
