[
https://issues.apache.org/jira/browse/MAPREDUCE-2266?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12981998#action_12981998
]
Todd Lipcon commented on MAPREDUCE-2266:
----------------------------------------
I took a peek at YDH and it looks like the solution used there is to defer the
sleep/SIGKILL into a new DelayedProcessKiller thread. But I couldn't find any
open source JIRA for this improvement. It would be appreciated if this patch
could be contributed - otherwise I'm happy to forward port for trunk.
> JvmManager sleeps between SIGTERM and SIGKILL while holding many TT locks
> -------------------------------------------------------------------------
>
> Key: MAPREDUCE-2266
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-2266
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: task-controller, tasktracker
> Affects Versions: 0.22.0
> Reporter: Todd Lipcon
> Priority: Blocker
> Fix For: 0.22.0
>
>
> Between sending a task SIGTERM and SIGKILL, the JvmManager will sleep for
> sleepTimeBeforeSigKill millis. But in many call heirarchies this is done
> while holding important locks like the TT lock and the JvmManagerForType
> lock. With the default 5 second sleep, this prevents other tasks from getting
> scheduled and reduces scheduling throughput.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
