[MPICH] RE: [MPICH] MPICH2 doesn´t distribute jobs when running applications
Matthew Chambers
matthew.chambers at vanderbilt.edu
Tue May 15 09:28:29 CDT 2007
Im still quite confused about what youre doing. I dont know whether or
not youre using the --ncpus option because you havent posted the way in
which you are starting your MPDs. Personally, I start my nodes MPDs at
boot time in the /etc/rc.local script. I always start my jobs from the head
node (which is running the main MPD) without a machine file because I dont
care which process ranks are running on which nodes as long as all the nodes
get the correct load. In your case, you should start your MPDs with the
--ncpus option depending on how many cores the machine has, and then start a
job WITHOUT a machine file and using -np <total number of cores>. That
should create a job that properly distributes your process over every core
(i.e. it runs 8 processes on each machine with 8 cores, 4 processes on each
machine with 4 cores, and 2 processes on each machine with 2 cores). The
caveat is that if you run less than -np <total number of cores> the end
result wont be balanced (i.e. your 2 core machines would be unused if they
were the last nodes in your MPD ring and you ran -np <total number of cores
- 4>.
Hope this helps,
Matt
_____
From: Christian M. Probst [mailto:cmacprobst at gmail.com]
Sent: Monday, May 14, 2007 9:35 PM
To: Matthew Chambers
Subject: Re: [MPICH] MPICH2 doesn´t distribute jobs when running
applications
I thought that machine file would be a good place to equilibrate
resources... I have 2 servers with 8 cores, 3 with 4 and 2 with 2... So, in
the machine file, I inform how many processes could be running in the MPD
ring...
As you probably saw, I am not using the --ncpus options, although now I
wonder if informing the processor number in the machine file would be
equivalnet?
Thanks.
Christian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20070515/ddaac02b/attachment.htm>
More information about the mpich-discuss
mailing list