[jira] [Commented] (REEF-276) Requesting 100 Containers on YARN causes injection errors in submitting tasks on REEF.NET on HDInsight

["Andrew Chung (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/REEF-276?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14512687#comment-14512687 ] 

Andrew Chung commented on REEF-276:
-----------------------------------

Thanks Markus, we discovered that the call to SubmitContext was taking 35 seconds on its first call (probably due to .NET JIT compiling), and 2-5 seconds on each subsequent call. Previously we were locking on the call, but now that we've moved it out, the performance has noticeably improved, though it still takes 5 minutes to allocate 100 containers. The problem we are running into now is after we've successfully received the containers, we fail to inject containers when the request is large. We know it works with 50 containers. The job is exited and marked as FINISHED-SUCCEEDED on YARN when we request 100 containers. More investigation is needed on what is causing the behavior.

> Requesting 100 Containers on YARN causes injection errors in submitting tasks on REEF.NET on HDInsight
> ------------------------------------------------------------------------------------------------------
>
>                 Key: REEF-276
>                 URL: https://issues.apache.org/jira/browse/REEF-276
>             Project: REEF
>          Issue Type: Bug
>          Components: REEF Bridge, REEF.NET
>    Affects Versions: 0.10
>         Environment: REEF.NET on HDInsight
>            Reporter: Andrew Chung
>              Labels: performance
>
> Requesting 100 Containers on YARN causes injection errors in submitting tasks on REEF.NET on HDInsight. Requesting 50 containers, however, is fine. More investigation on what is causing this is required, but we might need some help.



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