-machinefile support

[Pete Wyckoff]

cdmaest at sandia.gov said on Tue, 28 Oct 2003 17:53 -0700:
> I don't know if it's in the todo list, but is it possible to have
> mpiexec include a "-machinefile " option?  People grab a set of nodes,
> then launch executables on different sets of nodes with different
> machinefiles.  I explained how to do it with the -conf option, but was
> wondering if the -machinefile option could be used, then the testers
> don't have to change nothing in their suite.

I thought about this a bit, and am leaning against it although it would
be pretty trivial to implement.  Others might argue me back to the other
side, though.

My main complaint is the potential for confusion.  PBS generates a list
of hosts and mpiexec can only run on those hosts, no matter what is in
the machinefile.  I'm afraid that this option will tempt users to try to
use mpiexec outside of PBS, or with hosts not provided in a given PBS
allocation.

And somehow you have to get this machinefile generated within the
context of the batch job too.  Do the testers have any mechanism to
do that?  Are the hosts sufficiently different that they need to name
them by name rather than using, say, "-np 4" to grab a subset?  If so,
shouldn't the batch scheduler be taking care of choosing the correct
hosts according to some desired attributes?

Your explanation of how to generate a subset of hosts with -conf is the
right way to go, in my opinion.  You could write a little script that
provides the interface your testers seem to want, though:

    mpidoit -machinefile mach a.out --args

    #!/bin/bash
    # parse $* into variables instead of hardcoded ones below
    ...
    cat mach | while read host ; do
	echo $host : a.out --args
    done > mach2
    mpiexec -conf mach2

		-- Pete