Q: Node selection with mpirun under Scyld

[Daniel Ridge]

> I have a question about running mpi programs under the Scyld distribution. I
> have been unable to find in an answer in the documentation.
> 
> How do I specify which nodes to run an mpi program on? Or equivalently, how
> do I avoid my job being send to specific nodes (if, for instance, some nodes
> have too little memory for the job)?
> 
> I have some idea from the following:

You can try the 'npr' command that we shipped with the 27BZ-7 release.

The documentation for this system is not complete -- but you can try
'npr --help' or 'npr --man' to get human and nroff freindly info
respectively out of the tool.

The short form is that you can say:
npr --np 12 --exclude 4 --map 4:2 mpi-mandel

in place of 'mpirun' where you wanted a total of 12 processes, none on
node 4 and wanted rank 4 on node 2.

We're putting a lot of work into future releases of NPR to make it an
easy-to-use and easy-to-extend system for job co-scheduling on our
platform.

On the permissions front, you can see who owns Scyld nodes via:
bpstat

You can change the ownership via:
/usr/sbin/bpctl -S <nodenum> -u <uid>
	or
/usr/sbin/bpctl -S <nodenum> -g <gid>

Regards,
	Dan Ridge
	Scyld Computing Corporation