I apologize for the long time away from this. Last semester was a bit
insane here with my teaching load. I finally got some things cleared up and
can start looking at this again. I gave this a try on a machine here and
got mixed results. Part of the problem is that our local implementation of
the submission system seems to be ignoring the host and hostfile
specification and allocating nodes as it sees fit.
I will try to find another machine to test this on. I tried it on my
desktop machine playing around with localhost versus the fully qualified
domain name, but openmpi seems to be outsmarting me and recognizing that it
is all the same ip address.
Kelly Black Phone: (315) 600-8334
Clarkson University Fax: (315) 268-2371
Department of Math. & C.S.
PO Box 5815
Potsdam, NY 13699-5815
On Fri, Oct 11, 2013 at 5:05 PM, Jeff Squyres (jsquyres) <jsquyres_at_[hidden]
> On Oct 11, 2013, at 4:31 PM, Kelly Black <kjblack_at_[hidden]> wrote:
> > Yeah, it never occurred to me to use both, but when I was double
> checking things on the Open MPI website I came across a blurb that said
> that when used together the behavior is different. The documentation said
> that if the names in the --host command are not contained in the host file
> then the mpirun command fails.
> Correct. We basically use the --host as a filter on the hostfile. If the
> --host argument isn't a proper subset of the hostfile, OMPI errors.
> > It was not clear if you do not use a fully qualified name in one but not
> the other how it would be handled.
> I'm not sure. Like Ralph said, I *think* it works. I *think* OMPI may
> actually resolve all names first, and then do comparisons/filtering with IP
> You might want to give it a whirl and see what happens...?
> Jeff Squyres
> For corporate legal information go to: