[ 
https://issues.apache.org/jira/browse/HADOOP-2764?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12564889#action_12564889
 ] 

Doug Cutting commented on HADOOP-2764:
--------------------------------------

> best to keep heap setting low.

The JVM doesn't generally use more heap unless it has to.  Are you seeing 
datanodes and tasktrackers that use a lot of memory?  The purpose of limiting 
the heap size is to cause the JVM to promptly crash instead of making the 
entire machine slow down by paging.

Reducing the heap size limit for the tasktracker and datanode will only help if 
these are running amok, e.g., have a memory leak.  It would help us identify 
such problems sooner, but such problems would debilitate the node either way.

Also one can use a different hadoop-env.sh on the namenode and jobtracker from 
that on other nodes, but this can complicate deployments.

I think someone once proposed adding per-command config files, so that, if a 
file named 'hadoop-namenode-env.sh' existed in the config directory then then 
'bin/hadoop namenode' would include it after hadoop-env.sh.  That would be 
simple to add and would enable the feature desired here and a lot of other 
things.

> specify different heap size for namenode/jobtracker vs. tasktracker/datanodes
> -----------------------------------------------------------------------------
>
>                 Key: HADOOP-2764
>                 URL: https://issues.apache.org/jira/browse/HADOOP-2764
>             Project: Hadoop Core
>          Issue Type: New Feature
>    Affects Versions: 0.15.3
>            Reporter: Joydeep Sen Sarma
>            Priority: Minor
>
> tasktrackers/datanodes should be run with low memory settings. theres a lot 
> of competition for memory on slave nodes and these tasks don't need much 
> memory anyway and best to keep heap setting low.
> namenode needs higher memory and there's usually lots to spare on separate 
> box.
> hadoop-env.sh can provide different heap settings for central vs. slave 
> daemons.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.

Reply via email to