[jira] [Commented] (TEZ-221) Deadlock in DAGSchedulerMRR when setting slow start to 0.5f on a small cluster

["Bikas Saha (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/TEZ-221?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13741635#comment-13741635 ] 

Bikas Saha commented on TEZ-221:
--------------------------------

While a bunch of these changes helped resolve deadlocks I dont see much value added by this over the natural order scheduler. Both support slow start as maps complete and provide space for reduces. DAGMRRScheduler allows a stronger reduce resource guarantees that allow reducers to run even before some mappers even get to run. While this may be useful for some hand tuned scenarios, it requires a tricky deadlock-prone dance with priorities and IMO should be left as an expert level option. Also, making MRR logic work in a DAG gets even trickier. I plan to keep this scheduler in the code but default to using natural order scheduler. If needed this MRR scheduler may be plugged in via to-be-added API/config.
                
> Deadlock in DAGSchedulerMRR when setting slow start to 0.5f on a small cluster
> ------------------------------------------------------------------------------
>
>                 Key: TEZ-221
>                 URL: https://issues.apache.org/jira/browse/TEZ-221
>             Project: Apache Tez
>          Issue Type: Bug
>            Reporter: yeshavora
>            Assignee: Bikas Saha
>              Labels: TEZ-0.2.0
>         Attachments: TEZ-221.1.patch, TEZ-221.2.patch, TEZ-221.3.patch, TEZ-221.4.preemption.patch, TEZ-221.5.patch
>
>
> generate 10GB data using randomwriter and run sort job using tez-mapreduce-example.jar on single node cluster. The job runs very slow. 

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira