hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Devaraj Das (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HADOOP-2980) slow reduce copies - map output locations not being fetched even when map complete
Date Wed, 12 Mar 2008 09:12:46 GMT

    [ https://issues.apache.org/jira/browse/HADOOP-2980?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12577767#action_12577767
] 

Devaraj Das commented on HADOOP-2980:
-------------------------------------

The '50' can be explained since the default probe size (for mapcompletionevents) is 50. The
probe size increases with the number of copier threads via the calculation: _max(numCopiers*5,
50)_ So one thing that you should do is to increase the number of copier threads (mapred.reduce.parallel.copies
to something like 20).
But I am not able to clearly explain the 1 minute interval between the queries. This is one
scenario when it can happen - discussed in HADOOP-1719 (https://issues.apache.org/jira/browse/HADOOP-1719?focusedCommentId=12522954#action_12522954).
The other scenario where this can happen is when there is a failure to fetch output from a
host and the task backs off between 1 & 5 minutes from that host. But you say that you
don't see any exceptions... 
BTW the above issues are fixed in 0.16.

> slow reduce copies - map output locations not being fetched even when map complete
> ----------------------------------------------------------------------------------
>
>                 Key: HADOOP-2980
>                 URL: https://issues.apache.org/jira/browse/HADOOP-2980
>             Project: Hadoop Core
>          Issue Type: Bug
>    Affects Versions: 0.15.3
>            Reporter: Joydeep Sen Sarma
>
> maps are long finished. reduces are stuck looking for map locations. they make progress
- but slowly. it almost seems like they get new map locations every minute or so:
> 2008-03-07 18:50:52,737 INFO org.apache.hadoop.mapred.ReduceTask: task_200803041231_3586_r_000021_0
done copying task_200803041231_3586_m_004620_0 output from hadoop082.sf2p.facebook.com..
> 2008-03-07 18:50:53,733 INFO org.apache.hadoop.mapred.ReduceTask: task_200803041231_3586_r_000021_0:
Got 0 new map-outputs & 0 obsolete map-outputs from tasktracker and 0 map-outputs from
previous failures
> 2008-03-07 18:50:53,733 INFO org.apache.hadoop.mapred.ReduceTask: task_200803041231_3586_r_000021_0
Got 0 known map output location(s); scheduling...
> ...
> 2008-03-07 18:51:49,767 INFO org.apache.hadoop.mapred.ReduceTask: task_200803041231_3586_r_000021_0
Got 50 known map output location(s); scheduling...
> 2008-03-07 18:51:49,767 INFO org.apache.hadoop.mapred.ReduceTask: task_200803041231_3586_r_000021_0
Scheduled 41 of 50 known outputs (0 slow hosts and 9 dup hosts)
> they get about 50 locations at a time and this 1 minute delay pattern is surprisingly
common ..

-- 
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