[
https://issues.apache.org/jira/browse/YARN-3809?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14591806#comment-14591806
]
Jason Lowe commented on YARN-3809:
----------------------------------
I agree the thread pool should be configurable and possibly larger by default,
but note that the thread pool size has little to do with the number of AMs
running on a node. The pool size needs to be as large as the worst-case number
of AMs being launched during the worst-case retry duration to avoid any
unnecessary delays. For a 15 minute retry delay on a cluster launching
multiple apps per second on average, that's an unreasonable thread pool size.
I agree with [~kasha] that we need to lower the retry timeout as part of this
fix. As it is today we will expire the NM due to lack of heartbeat before we
will give up on an AM launch retry which makes no sense.
We can update ipc.client.connect.max.retries.on.timeouts and potentially
ipc.client.connect.timeout for the conf passed to the NM proxy we create,
although we need to make sure we make a copy of the config to avoid polluting
other proxies.
> Failed to launch new attempts because ApplicationMasterLauncher's threads all
> hang
> ----------------------------------------------------------------------------------
>
> Key: YARN-3809
> URL: https://issues.apache.org/jira/browse/YARN-3809
> Project: Hadoop YARN
> Issue Type: Improvement
> Components: resourcemanager
> Reporter: Jun Gong
> Assignee: Jun Gong
> Attachments: YARN-3809.01.patch
>
>
> ApplicationMasterLauncher create a thread pool whose size is 10 to deal with
> AMLauncherEventType(LAUNCH and CLEANUP).
> In our cluster, there was many NM with 10+ AM running on it, and one shut
> down for some reason. After RM found the NM LOST, it cleaned up AMs running
> on it. Then ApplicationMasterLauncher need handle these 10+ CLEANUP event.
> ApplicationMasterLauncher's thread pool would be filled up, and they all hang
> in the code containerMgrProxy.stopContainers(stopRequest) because NM was
> down, the default RPC time out is 15 mins. It means that in 15 mins
> ApplicationMasterLauncher could not handle new event such as LAUNCH, then new
> attempts will fails to launch because of time out.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
