[jira] [Commented] (TEZ-3452) Auto-reduce parallelism calculation can overflow with large inputs

["Jonathan Eagles (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/TEZ-3452?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15545549#comment-15545549 ] 

Jonathan Eagles commented on TEZ-3452:
--------------------------------------

[~hitesh], [~rajesh.balamohan], [~bikassaha], [~sseth].
Total source output size calculation now is only approximate and could cause some confusion. Let me know if you prefer and I can use exact method and fallback if overflow is detected.

> Auto-reduce parallelism calculation can overflow with large inputs
> ------------------------------------------------------------------
>
>                 Key: TEZ-3452
>                 URL: https://issues.apache.org/jira/browse/TEZ-3452
>             Project: Apache Tez
>          Issue Type: Bug
>            Reporter: Jonathan Eagles
>            Assignee: Jonathan Eagles
>         Attachments: TEZ-3452.1.patch, TEZ-3452.2.patch
>
>
> Overflow can occur when the numTasks is high (say 45000) and outputSize is high (say 311TB) and slow start is set to 1.0. 
> {code:title=ShuffleVertexManager}
>     for (Map.Entry<String, SourceVertexInfo> vInfo : getBipartiteInfo()) {
>       SourceVertexInfo srcInfo = vInfo.getValue();
>       if (srcInfo.numTasks > 0 && srcInfo.numVMEventsReceived > 0) {
>         // this assumes that 1 vmEvent is received per completed task - TEZ-2961
>         expectedTotalSourceTasksOutputSize += 
>             (srcInfo.numTasks * srcInfo.outputSize) / srcInfo.numVMEventsReceived;
>       }
>     }
> {code}



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