hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Chris Douglas (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HADOOP-3446) The reduce task should not flush the in memory file system before starting the reducer
Date Sat, 30 Aug 2008 00:43:44 GMT

    [ https://issues.apache.org/jira/browse/HADOOP-3446?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12627175#action_12627175
] 

Chris Douglas commented on HADOOP-3446:
---------------------------------------

I'll add unit tests/docs with the next patch.

As a benchmark, I tried RandomWriter on 19 TaskTrackers, 1GB/node, followed by several sort
runs. The max heap memory is set to 512MB, mapred.copy.inmem.percent to 0.8, dfs.replication
to 1. The times recorded are the min/max/avg time for the reduce from the end of the shuffle
to the end of the reduce.

Params are formatted as: {{io.sort.factor/mapred.inmem.merge.threshold/mapred.inmem.merge.usage/mapred.reduce.inmem.percent}}

|| Params || Min || Max || Avg || Notes ||
| 100/0/1.0/1.0 | 8.35 | 57.775 | 23.1603 | Never hits disk |
| 9/15/1.0/0.01 | 11.164 | 67.569 | 38.0216 | Spills several times, merges some in-memory
segments during intermediate merge | 
| 100/0/1.0/0.5 | 11.215 | 74.59 | 33.8571 | Spills some segments to disk before starting
reduce |
| 100/0/1.0/0.0 | 17.184 | 88.479 | 59.5489 | Spills all segments to disk before starting
reduce |



> The reduce task should not flush the in memory file system before starting the reducer
> --------------------------------------------------------------------------------------
>
>                 Key: HADOOP-3446
>                 URL: https://issues.apache.org/jira/browse/HADOOP-3446
>             Project: Hadoop Core
>          Issue Type: Improvement
>          Components: mapred
>            Reporter: Owen O'Malley
>            Assignee: Chris Douglas
>            Priority: Critical
>             Fix For: 0.19.0
>
>         Attachments: 3446-0.patch, 3446-1.patch, 3446-2.patch
>
>
> In the case where the entire reduce inputs fit in ram, we currently force the input to
disk and re-read it before giving it to the reducer. It would be much better if we merged
from the ramfs and any spills to feed the reducer its input.

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


Mime
View raw message