[
https://issues.apache.org/jira/browse/MAPREDUCE-1495?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12833834#action_12833834
]
Rajesh Balamohan commented on MAPREDUCE-1495:
---------------------------------------------
As of now, its implemented as follows in JobTracker
public synchronized TaskCompletionEvent[] getTaskCompletionEvents(
JobID jobid, int fromEventId, int maxEvents) throws IOException{
synchronized (this) {
JobInProgress job = this.jobs.get(jobid);
if (null != job) {
if (job.inited()) {
return job.getTaskCompletionEvents(fromEventId, maxEvents);
} else {
return EMPTY_EVENTS;
}
}
}
return completedJobStatusStore.readJobTaskCompletionEvents(jobid,
fromEventId, maxEvents);
}
Where, jobs is TreeMap<JobID, JobInProgress>().
It is possible to reduce to contention in 2 ways.
1. Reduce the synch to only JobInProgress job = this.jobs.get(jobid); Rest of
the code is independent of the synch block (asaik).
2. Change datastructure of jobs to ConcurrentHashMap<JobID, JobInProgress>().
This way, we can jobs.get(jobid) automatically becomes threadsafe and the total
synchornization itself can be eliminated. If its mandatory to maintain the
order, I have to try the 1st one.
> Reduce locking contention on JobTracker.getTaskCompletionEvents()
> -----------------------------------------------------------------
>
> Key: MAPREDUCE-1495
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-1495
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Affects Versions: 0.20.1
> Reporter: Rajesh Balamohan
>
> While profiling JT for slow performance with small-jobs, it was observed that
> JobTracker.getTaskCompletionEvents() is attributing to 40% of lock contention
> on JT.
> This JIRA ticket is created to explore the possibilities of reducing the
> sychronized code block in this method.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
