[
https://issues.apache.org/jira/browse/HADOOP-3864?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Arun C Murthy updated HADOOP-3864:
----------------------------------
Attachment: HADOOP-3864_0_20080830.patch
Here is the simplest patch to fix this, rather than the correct path which is
to go fix the whole synchronization mess in the JobTracker (see HADOOP-869).
For now, I just turn away the JobClient when the job is being initalized,
thereby avoiding the problem...
> JobTracker lockup due to JobInProgress.initTasks taking significant time for
> large jobs on large clusters
> ---------------------------------------------------------------------------------------------------------
>
> Key: HADOOP-3864
> URL: https://issues.apache.org/jira/browse/HADOOP-3864
> Project: Hadoop Core
> Issue Type: Bug
> Components: mapred
> Affects Versions: 0.18.0
> Reporter: Arun C Murthy
> Assignee: Arun C Murthy
> Priority: Critical
> Fix For: 0.19.0
>
> Attachments: HADOOP-3864_0_20080830.patch
>
>
> JobInProgress.initTasks takes significant amount of time on a large cluster
> for large jobs (55k maps * 3 splits), during which the JobInProgress object
> is locked up.
> Simultaneously the JobClient is calling JobTracker.getTaskCompletionEvents
> which locks the JobTracker & tries to lock the JobInProgress, there-by it
> starves all heartbeats which are trying to lock the JobTracker - resulting in
> a lockup.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
