[
https://issues.apache.org/jira/browse/MAPREDUCE-3404?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13185986#comment-13185986
]
Vinod Kumar Vavilapalli commented on MAPREDUCE-3404:
----------------------------------------------------
bq. How do we make sure that if mapreduce.job.maps.speculative=false and
mapreduce.job.reduce.speculative=true, the maps dont get speculated and the
reduces gets speculated?
The speculator handles map and reduce speculation separately. I just looked at
the patch, and it achieves the above by not sending any map events to the
speculative when map-speculation is disabled. The speculator doesn't seem to
find any maps to speculates (as it doesn't know about any maps at all) and so
only speculates reduces. Works (IMO) a convoluted way but can live with that.
+1 for the patch. Pushing this in.
> Speculative Execution: speculative map tasks launched even if
> -Dmapreduce.map.speculative=false
> -----------------------------------------------------------------------------------------------
>
> Key: MAPREDUCE-3404
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-3404
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: job submission, mrv2
> Affects Versions: 0.23.0
> Environment: Hadoop version is: Hadoop 0.23.0.1110031628
> 10 node test cluster
> Reporter: patrick white
> Assignee: Eric Payne
> Priority: Critical
> Fix For: 0.23.0, 0.23.1, 0.24.0
>
> Attachments: MAPREDUCE-3404.1.txt, MAPREDUCE-3404.2.txt
>
>
> When forcing a mapper to take significantly longer than other map tasks,
> speculative map tasks are
> launched even if the mapreduce.job.maps.speculative.execution parameter is
> set to 'false'.
> Testcase: ran default WordCount job with spec execution set to false for both
> map and reduce but still saw a fifth mapper
> task launch, ran job as follows:
> hadoop --config <config> jar /tmp/testphw/wordcount.jar WordCount
> -Dmapreduce.job.maps.speculative.execution=false
> -Dmapreduce.job.reduces.speculative.execution=false
> /tmp/test_file_of_words* /tmp/file_of_words.out
> Input data was 4 text files >hdfs blocksize, with same word pattern plus one
> diff text line in each file, fourth
> file was 4 times as large as others:
> hadoop --config <config> fs -ls /tmp
> Found 5 items
> drwxr-xr-x - user hdfs 0 2011-10-20 16:17 /tmp/file_of_words.out
> -rw-r--r-- 3 user hdfs 62800021 2011-10-20 14:45 /tmp/test_file_of_words1
> -rw-r--r-- 3 user hdfs 62800024 2011-10-20 14:46 /tmp/test_file_of_words2
> -rw-r--r-- 3 user hdfs 62800024 2011-10-20 14:46 /tmp/test_file_of_words3
> -rw-r--r-- 3 user hdfs 271708312 2011-10-20 15:50 /tmp/test_file_of_words4
> Job launched 5 mappers despite spec exec set to false, output snippet:
> org.apache.hadoop.mapreduce.JobCounter
> NUM_FAILED_MAPS=1
> TOTAL_LAUNCHED_MAPS=5
> TOTAL_LAUNCHED_REDUCES=1
> RACK_LOCAL_MAPS=5
> SLOTS_MILLIS_MAPS=273540
> SLOTS_MILLIS_REDUCES=212876
> Reran same case as above only set both spec exec params to 'true', same
> results only this time the fifth task being
> launched is expected since spec exec = true.
> job run:
> hadoop --config <config> jar /tmp/testphw/wordcount.jar WordCount
> -Dmapreduce.job.maps.speculative.execution=true
> -Dmapreduce.job.reduces.speculative.execution=true
> /tmp/test_file_of_words* /tmp/file_of_words.out
> output snippet:
> org.apache.hadoop.mapreduce.JobCounter
> NUM_FAILED_MAPS=1
> TOTAL_LAUNCHED_MAPS=5
> TOTAL_LAUNCHED_REDUCES=1
> RACK_LOCAL_MAPS=5
> SLOTS_MILLIS_MAPS=279653
> SLOTS_MILLIS_REDUCES=211474
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
