hadoop-mapreduce-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Raghu Angadi (JIRA)" <j...@apache.org>
Subject [jira] Created: (MAPREDUCE-2149) Distcp : setup with update is too slow when latency is high
Date Mon, 25 Oct 2010 16:15:21 GMT
Distcp : setup with update is too slow when latency is high
-----------------------------------------------------------

                 Key: MAPREDUCE-2149
                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-2149
             Project: Hadoop Map/Reduce
          Issue Type: Bug
          Components: distcp
    Affects Versions: 0.21.0, 0.20.2
            Reporter: Raghu Angadi
            Assignee: Raghu Angadi



If you run distcp with '-update' option, for _each of the files_ present on source cluster
setup invokes a separate RPC to destination cluster to fetch file info. 
Usually this overhead is not very noticeable when both cluster are geographically close to
each other. But if the latency is large, setup could take couple of orders of magnitude longer.

E.g. : source has 10k directories, each with about 10 files, round trip latency between source
and destination is 75 ms (typical for coast-to-coast clusters). 
If we run distcp on source cluster, set up would take about _2.5 hours_ irrespective of whether
destination has these files or not. '-lsr' on the same dest dir from source cluster would
take up to 12 min (depending on how many directories already exist on dest). 

  * A fairly simple fix to how setup() iterates should bring the set up time to same as '-lsr'.
I will have a patch for this.. (though 12 min is too large).
  * A more scalable option is to differ update check to mappers.



-- 
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