[ https://issues.apache.org/jira/browse/HADOOP-1433?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12499171
]
Doug Cutting commented on HADOOP-1433:
--------------------------------------
The answer we've come to in the past is that an elaborate scheduler is beyond the immediate
scope of Hadoop. So I don't think anyone's currently working on this, although you are free
to if you like.
Some workarounds that have been proposed are:
1. Run two sets of MapReduce daemons on the same cluster, one niced, the other not. Since
HADOOP-468, Hadoop has supporting setting HADOOP_NICENESS in conf/hadoop-env.sh. Then high-priority,
short-running jobs may be submitted to the un-niced cluster while low-priority, long-running
jobs may be submitted to the niced cluster. I've heard this works well, that when a high-priority
job is submitted, the low-priority job slows nearly to a halt, giving nearly all the resources
to the high-priority jobs.
2. Use HOD (HADOOP-719) or Amazon EC2 to dynamically allocate MapReduce clusters per job or
job-sequence.
> Add job priority
> ----------------
>
> Key: HADOOP-1433
> URL: https://issues.apache.org/jira/browse/HADOOP-1433
> Project: Hadoop
> Issue Type: New Feature
> Components: mapred
> Reporter: Johan Oskarsson
> Priority: Minor
>
> As more and more developers start using our cluster we run into problems where big low
priority jobs block smaller high priority ones.
> A simple way of specifying the job priority in the JobConf and perhaps even change it
during runtime via the jobtracker web ui would help a lot.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
|