[
https://issues.apache.org/jira/browse/YARN-3136?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14306500#comment-14306500
]
Sunil G commented on YARN-3136:
-------------------------------
1. As you mentioned a full lock on Scheduler is grabbed for getting
liveContainers (except am container) from FicaSchedulerApp. This comes from
below map.
{code}
protected Map<ApplicationId, SchedulerApplication<T>> applications;
{code}
And I could see that, most of the methods are grabbing a Scheduler lock while
accessing this map.
2. *rmContext.getRMApps()*. Yes, this is a concurrent map and i feel this need
not to be surrounded by a lock in scheduler level.
I feel, we could improve getTransferredContainers like below:
A getter can be used under scheduler lock to get the FicaSchedulerApp object
within *getTransferredContainers* . Also we can remove the scheduler lock from
getTransferredContainers. This must be safe enough.
[~varun_saxena] As i have done some ground work, is it ok if i assign this
ticket to myself :).
[~jlowe] pls share your thoughts on this approach. Also I would like to take
this jira, pls feel free to takeover otherwise :).
> getTransferredContainers can be a bottleneck during AM registration
> -------------------------------------------------------------------
>
> Key: YARN-3136
> URL: https://issues.apache.org/jira/browse/YARN-3136
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: scheduler
> Affects Versions: 2.6.0
> Reporter: Jason Lowe
> Assignee: Varun Saxena
>
> While examining RM stack traces on a busy cluster I noticed a pattern of AMs
> stuck waiting for the scheduler lock trying to call getTransferredContainers.
> The scheduler lock is highly contended, especially on a large cluster with
> many nodes heartbeating, and it would be nice if we could find a way to
> eliminate the need to grab this lock during this call. We've already done
> similar work during AM allocate calls to make sure they don't needlessly grab
> the scheduler lock, and it would be good to do so here as well, if possible.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
