Scyld: disabling mpi processes on master

Art Edwards edwards at icantbelieveimdoingthis.com
Wed Jun 27 07:35:42 PDT 2001


On Tue, Jun 26, 2001 at 06:03:40PM -0500, Jarrod Smith wrote:
> I don't want to use the master as a compute node for MPI runs (by design,
> it has slower processors than the rest of the cluster).  I've read some
> things in the archive which hint at the fact that it is possible, but
> nothing I've tried works.  There doesn't appear to be a -nolocal
> option (despite posts from Scyld that it is in the mpirun
> man page), and using bpsh to spawn the job doesn't work, because
> there is no /var on the slaves, and the root process seems to want
> to want to write some temp files there.
> 
> So...can it be done?  If so, exactly how?
Thanks to Keith Underwood, I now can run jobs off of the head node using the following command.

NO_INLINE_MPIRUN=true bpsh n application -p4pg file

The p4p4 switch indicates that you will use a p4pg file to indicate which
nodes you will use for the calculation. I attach, below, a sample p4pg file.
application is the executable. n is the node on which the p4pg file resides.

I am making this post because I have stated before that I was unable to 
open files on slave nodes unless the head node was part of the run. It turns
out that, at 4:30 in the morning, I had made a dumb mistake in the call 
above so that I couldn't open the files. I have since been successful in
opening files in the /tmp directory of the slave nodes. 

Here is the sample file:

0 0 /home/edwardsa/QUEST/lcao2p46a/lcao_nebgga.x
1 1 /home/edwardsa/QUEST/lcao2p46a/lcao_nebgga.x
2 1 /home/edwardsa/QUEST/lcao2p46a/lcao_nebgga.x
3 1 /home/edwardsa/QUEST/lcao2p46a/lcao_nebgga.x
4 1 /home/edwardsa/QUEST/lcao2p46a/lcao_nebgga.x

Note that the application here has been the same as the application in the
call above.

I hope this helps.

Art Edwards
> 
> TIA
> 
> -- 
> Jarrod A. Smith
> Research Asst. Professor, Biochemistry
> Asst. Director, Center for Structural Biology
> Computation and Molecular Graphics
> Vanderbilt University
> 
> jsmith at structbio.vanderbilt.edu
> 
> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf




More information about the Beowulf mailing list