[
https://issues.apache.org/jira/browse/MAPREDUCE-2634?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13058214#comment-13058214
]
Owen O'Malley commented on MAPREDUCE-2634:
------------------------------------------
Proposal #1 breaks the architecture by making a loop in the communication
graph. That means you need to work very hard to ensure there aren't any
distributed deadlocks. (We also need some way to minimize the chance that
someone introduces one later when working on the code.) I don't see how you'll
get enough performance out of that to risk the potential for deadlocks.
Proposal #2 violates the security model of never running user code in the
servers. That is pretty much a non-starter since it would require security
managers in all of the servers, which would impose a huge performance
degradation based on everything I've read about them.
Proposal #3 has been there for a long time.
If proposal #4 isn't already done, we should do it. A fast heartbeat for
committing the output makes sense.
> MapReduce Performance Improvements using forced heartbeat
> ----------------------------------------------------------
>
> Key: MAPREDUCE-2634
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-2634
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Reporter: Abhijit Suresh Shingate
> Priority: Minor
> Original Estimate: 168h
> Remaining Estimate: 168h
>
> Following are the proposals which would cause some performance optimizations
> over MapReduce
> *1.Notify TaskTracker to send heartbeat when a new Job is submitted*
> a) Presently when new Job is submitted to JobTracker, the tasks are
> assigned to TaskTracker only when the TaskTracker sends heartbeat to
> JobTracker
> b) Proposal:
> - JobTracker will notify all TaskTrackers to send heartbeat to
> JobTracker whenever a new Job is submitted to JobTracker. So that the Tasks
> of the new Job can be immediately assigned to all TaskTrackers.
> *2. Execute Job Setup and Cleanup on JobTracker JVM*
> a) Presently Job Setup and Cleanup is carried out as a separated task on
> TaskTracker
> b) Launching a new JVM for Setup and Cleanup of the Job introduces some
> amount of overhead. It takes generally about 0.7 - 1.5 seconds.
> c) Proposal:
> - JobTracker will execute the Job Setup and Cleanup tasks on the
> JobTracker JVM only.
> *3. Request TaskTracker to send heartbeat when the Map Task is completed.*
> a) Presently TaskTracker reports status of completed Map Tasks as part of
> heartbeat at a regular interval.
> b) Proposal:
> - Map Task requests TaskTracker to send heartbeat to JobTracker when
> Map Task is completed. So that Reduce task can quickly know which map task is
> finished and copy map outputs to local.
> *4. Request JobTracker to trigger committing of Reduce output when Reduce
> Task has finished.*
> a) Presently JobTracker will ask the Reduce Task to commit its output to
> HDFS through heartbeat response.
> b) Proposal:
> - Reduce Task requests TaskTracker to send heartbeat to JobTracker
> whenever Reduce Task is completed.
> These optimizations might work on small clusters but on big clusters it may
> be overhead.
> Please let us know your views.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
