[
https://issues.apache.org/jira/browse/MAPREDUCE-64?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12767009#action_12767009
]
Todd Lipcon commented on MAPREDUCE-64:
--------------------------------------
Thanks for those changes.
Regarding diagrams, I had an idea - one simple way might be to simply add TRACE
level log messages at every collect() call with the current values of every
index plus the spill number in simple TSV format. Feeding those into gnuplot or
R to graph them should produce something reasonably informative without much
manual diagramming. If you want, I can try to do this and post a patch - just
let me know. My inspiration here was seekwatcher:
http://oss.oracle.com/~mason/seekwatcher/ext3_vs_btrfs_vs_xfs.png
Also, I remember when we were first looking at this last August, I ran a simple
test where I was running a sort of 10 byte records, and it turned out that the
"optimal" io.sort.record.percent caused my job to be significantly slower. It
was the case then that a small number of large spills actually ran slower than
a large number of small spills. Did we ever determine what that issue was? I
think we should try to understand why the theory isn't agreeing with
observations here.
> Map-side sort is hampered by io.sort.record.percent
> ---------------------------------------------------
>
> Key: MAPREDUCE-64
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-64
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Reporter: Arun C Murthy
> Assignee: Chris Douglas
> Attachments: M64-0.patch, M64-1.patch, M64-2.patch, M64-3.patch
>
>
> Currently io.sort.record.percent is a fairly obscure, per-job configurable,
> expert-level parameter which controls how much accounting space is available
> for records in the map-side sort buffer (io.sort.mb). Typically values for
> io.sort.mb (100) and io.sort.record.percent (0.05) imply that we can store
> ~350,000 records in the buffer before necessitating a sort/combine/spill.
> However for many applications which deal with small records e.g. the
> world-famous wordcount and it's family this implies we can only use 5-10% of
> io.sort.mb i.e. (5-10M) before we spill inspite of having _much_ more memory
> available in the sort-buffer. The word-count for e.g. results in ~12 spills
> (given hdfs block size of 64M). The presence of a combiner exacerbates the
> problem by piling serialization/deserialization of records too...
> Sure, jobs can configure io.sort.record.percent, but it's tedious and
> obscure; we really can do better by getting the framework to automagically
> pick it by using all available memory (upto io.sort.mb) for either the data
> or accounting.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
