hadoop-mapreduce-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Allen Wittenauer (JIRA)" <j...@apache.org>
Subject [jira] [Resolved] (MAPREDUCE-1939) split reduce compute phase into two threads one for reading and another for computing
Date Tue, 10 Mar 2015 02:54:39 GMT

     [ https://issues.apache.org/jira/browse/MAPREDUCE-1939?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Allen Wittenauer resolved MAPREDUCE-1939.
-----------------------------------------
    Resolution: Won't Fix

stale

> split reduce compute phase into two threads one for reading and another for computing
> -------------------------------------------------------------------------------------
>
>                 Key: MAPREDUCE-1939
>                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-1939
>             Project: Hadoop Map/Reduce
>          Issue Type: Improvement
>          Components: task
>    Affects Versions: 0.20.2
>            Reporter: wangxiaowei
>
> it is known that  reduce task is made up of three phases: shuffle , sort and reduce.
During reduce phase, a reduce function will read a record from disk or memory first and process
it to write to hdfs finally. To convert this serial progress to parallel progress , I split
the reduce phase into two threads called producer and consumer individually. producer is used
to read record from disk and consumer to process the records read by the first one. I use
two buffer, if  producer is writing one buffer consumer will read from another buffer.  Theoretically
 there will be a overlap between this two phases so we can reduce the whole reduce time.
> I wonder why hadoop does not implement it originally? Is there some potential problems
for such ideas ?
> I have already implemmented a prototypy. The producer just reads bytes from the disk
and leaves the work of transformation to real key and value objects to consumer. The results
is not good only a improvement of 13%  for time. I think it has someting with the buffer size
and the time spending on different threads.Maybe the tiem spend by consumer thread is too
long and the producer has to wait until the next buffer is available.



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

Mime
View raw message