[ 
https://issues.apache.org/jira/browse/FLINK-1683?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14355947#comment-14355947
 ] 

ASF GitHub Bot commented on FLINK-1683:
---------------------------------------

GitHub user fhueske opened a pull request:

    https://github.com/apache/flink/pull/476

    [FLINK-1683] [jobmanager] Fix scheduling preference choice for non-unary 
execution tasks

    Fix validated on a cluster setup.

You can merge this pull request into a Git repository by running:

    $ git pull https://github.com/fhueske/flink schedulerFix

Alternatively you can review and apply these changes as the patch at:

    https://github.com/apache/flink/pull/476.patch

To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:

    This closes #476
    
----

----


> Scheduling preferences for non-unary tasks are not correctly computed
> ---------------------------------------------------------------------
>
>                 Key: FLINK-1683
>                 URL: https://issues.apache.org/jira/browse/FLINK-1683
>             Project: Flink
>          Issue Type: Bug
>          Components: JobManager
>    Affects Versions: 0.9, 0.8.1
>            Reporter: Fabian Hueske
>            Assignee: Fabian Hueske
>             Fix For: 0.9, 0.8.2
>
>
> When computing scheduling preferences for an execution task, the JobManager 
> looks at the assigned instances of all its input execution tasks and returns 
> a preference only if not more than 8 instances have been found (if the input 
> of a tasks is distributed across more than 8 tasks, local scheduling won't 
> help a lot in any case).
> However, the JobManager treats all input execution tasks the same and does 
> not distinguish between different logical input. The effect is that a join 
> tasks with one broadcasted and one locally forwarded task is not locally 
> assigned towards its locally forwarded input.
> This can have a significant impact on the performance of tasks that have more 
> than one input and which rely on local forwarding and co-located task 
> scheduling.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to