Kayla Jay a écrit :
I'm trying to figure out why I need to use HOD vs. trying to run multiple
jobs at the same time on the same set of resources. Is it possible to run
> multiple hadoop jobs at the same time on the same set of input data?
I tried to run different jobs on the same set of data at the same time,
> but it takes a while (way while) and almost appears as if it queues up
> and the next job has to wait and so forth before completing.
So, I tried moving onto HOD. It's not very apparent why one would want
> to use HOD to run on different nodes at the same time for different
> jobs that access the same set of input data.
Can anyone provide any feedback on running multiple jobs at the same
> time on the same set of data? HOD use? Why would I have to run HOD
> and schedule running multiple jobs at the same time on the same
> set of data, but within their own set of resources in the cluster?
Hi !
I just contributed a new implementation of the scheduler that adds an
option called "mapred.jobtracker.scheduler.maxRunningTasksPerJob"
allowing you to limit the number of nodes allocated to a Job (and so not
to use HOD).
This limit is a hint and if some nodes have nothing to do, they will be
allocated anyway.
If you want to test it, the patch is available in the bug #3412
http://issues.apache.org/jira/browse/HADOOP-3412
It applies on TRUNK but I can make a few modifications if you want it to
apply on a release.
Ant jar should be sufficient to build it, but please ask me if you have
more question.
I would really appreciate your feedback about the behavior of that
scheduler. I'm trying to solve precisely those problem resulting of
partitioned clusters, and I'll try to do something that suit better to
your needs if you can tell me more.
Brice
PS: Please excuse me for my English :-P