[jira] Commented: (HADOOP-3961) resource estimation works badly in some cases

["Xing Shi (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HADOOP-3961?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12664038#action_12664038 ] 

Xing Shi commented on HADOOP-3961:
----------------------------------

Hi, I have meet a special resource estimateion work. Several of the map's input files is zero, but there is also several bytes created in map's output. So :

double blowupOnThisTask = ts.getOutputSize() / 
        ((double) tip.getMapInputSize() + 1);

blowupOnThisTask  will be increased large, and the reduces can't be allocated the reduce tasks from the JobTracker.

> resource estimation works badly in some cases
> ---------------------------------------------
>
>                 Key: HADOOP-3961
>                 URL: https://issues.apache.org/jira/browse/HADOOP-3961
>             Project: Hadoop Core
>          Issue Type: Bug
>          Components: mapred
>    Affects Versions: 0.19.0
>            Reporter: Ari Rabkin
>            Assignee: Ari Rabkin
>            Priority: Blocker
>             Fix For: 0.19.0
>
>         Attachments: resourceEstFix.patch
>
>
> The disk space estimation introduced in HADOOP-657 performs badly under some circumstances.  In particular, if maps have very small output, the initial estimate of output size equals input size throws everything off.
> Fix is to use a smaller initial estimate, and to update it appropriately.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.