[
https://issues.apache.org/jira/browse/MAPREDUCE-2011?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12898481#action_12898481
]
Koji Noguchi commented on MAPREDUCE-2011:
-----------------------------------------
When a task is initialized, it calls getFileStatus for every distributed cache
file/archive entries it has (_dfsFileStamp_) and compare it with task's
timestamp specified in the config (_confFileStamp_).
This makes sure that tasks fail *at start up* if distributed cache files were
changed after the job was submitted and before the task started.
(It still doesn't guarantee that job would fail reliably since all the tasks
could have been started before the modification.)
Now asking if we can change this logic to,
If exact localized cache exists ('lcacheStatus.mtime == confFileStamp ') on the
TaskTracker, use that and do not call getFileStatus(_dfsFileStamp_).
With this, no getFileStatus calls are made if TaskTracker already has the
localized cache with the same timestamp. This should reduce the amount of
getFileStatus calls significantly when people submit jobs using the same
distributed cache files.
This still makes sure that all the tasks use the same dist cache files
specified at the job startup. (corectness)
But with this change, tasks that would have failed at start-up due to
(_dfsFileStamp_ != _confFileStamp_) can now succeed.
> Reduce number of getFileStatus call made from every
> task(TaskDistributedCache) setup
> ------------------------------------------------------------------------------------
>
> Key: MAPREDUCE-2011
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-2011
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Components: distributed-cache
> Reporter: Koji Noguchi
>
> On our cluster, we had jobs with 20 dist cache and very short-lived tasks
> resulting in 500 map tasks launched per second resulting in 10,000
> getFileStatus calls to the namenode. Namenode can handle this but asking to
> see if we can reduce this somehow.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
