Multiple node groups in Scyld

[Ray Jones]

Karl Kosack <kosack at hbar.wustl.edu> writes:

> I know in other MPI  implementations, it is possible to create a
> file of machine names which specifies which nodes to use for a particular
> job, but I can't seem to find that capability in the Scyld distribution.
> I imagine there must be a simple way to do this, since a large cluster
> wouldn't be very useful without it.  Can anyone help me?

In /usr/mpi-beowulf/bin there's a version of mpirun from the MPICH
implementation.  It takes a -machinefile argument.  However, I'm not
sure how well supported and integrated this version of mpirun is.  At
least some parts (e.g., -nolocal) don't work.

The issue of running jobs on specific nodes has come up several
times.  It should probably be made a FAQ in the Scyld documentation.

It would be nice to see an answer from someone at Scyld, either in the
form "<foo> is the supported way to run jobs on specific nodes" or
"we're working on a solution for the next release."

Ray Jones