[Hadoop-common-user] What can cause: Map output copy failure

Hi,
Can you please let us know your system configuration running hadoop?
The error you see is when the reducer is copying its respective map output into memory. The parameter mapred.job.shuffle.input.buffer.percent can be manipulated for this ( a bunch of others will also help you optimize sort later ), but I would say 200M is far too less memory allocated for hadoop application jvms :)

Amogh


On 1/8/10 2:46 AM, "Mayuran Yogarajah" wrote:

I'm seeing this error when a job runs:

Shuffling 35338524 bytes (35338524 raw bytes) into RAM from attempt_201001051549_0036_m_000003_0
Map output copy failure: java.lang.OutOfMemoryError: Java heap space
at org.apache.hadoop.mapred.ReduceTask$ReduceCopier$MapOutputCopier.shuffleInMemory(ReduceTask.java:1198)


I originally had mapred.child.java.opts set to 200M. If I boost this up
to 512M the error goes away.
I'm trying to understand whats going on though. Can anyone explain?
Also are there any other parameters that
I should be tweaking to help with this?

thank you very much,
M

Hi Amogh,

We're using a 3 node cluster, all are quad cores (intel x3220) with 4
gigs of ram.
They are running Centos 5.3, and Hadoop 0.18.3.

After looking at the source code I (possibly mistakenly) thought that
fs.inmemory.size.mb might
have something to do with this. I had bumped it up to 200 (default is
75), but the heap was left
at 200M. I think when I configured the cluster initially I had
mistakenly thought that 200M for
heap was enough, but it wasn't.

I was able to make the error go away by:
1) increasing mapred.child.java.opts
2) decreasing fs.inmemory.size.mb

Do you know of any other parameters that I should be tweaking ?

thanks,
M