[
https://issues.apache.org/jira/browse/MAPREDUCE-257?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12730568#action_12730568
]
Hong Tang commented on MAPREDUCE-257:
-------------------------------------
I am pretty sure we track memory by walking down a process tree. Of course, a
user can defeat this by fork->fork->exec (aka daemonize). We should also be
able to (if not already so) first collect the set of processes derived from the
child task (inclusive), and then kill them all. But we may still miss some
processes that are created between the checking and killing.
> Preventing node from swapping
> -----------------------------
>
> Key: MAPREDUCE-257
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-257
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Reporter: Hong Tang
>
> When a node swaps, it slows everything: maps running on that node, reducers
> fetching output from the node, and DFS clients reading from the DN. We should
> just treat it the same way as if OS exhausts memory and kill some tasks to
> free up memory.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
