At 09:24 AM 1/8/2004, Peter Shenkin wrote:
>Happy New Year, folks!
>
>This was the subject of some discussion here before.
>
>IIRC, Tim Prince asserted that MPI executables not specially
>compiled for Opteron would not run on Opteron on more than a
>single processor.
>
>An outside vendor who is pushing Opteron Linux clusters has been
>trying to get this to work with our MPI executable (Jaguar), which
>was compiled on Intel under RH 7.3, using gcc and Absoft Fortran.
>
>When I last reported here, I asserted that he had been able
>to get it to run under MPI (one or) two processors. However,
>he has since succeeded in getting it to run on multiple processors.
>
>I asked:
>:: Were you ever able to get MPI Jaguar to run on Opteron on
>:: more than two processors?
>
>He responded:
>: Yes, I was able to get it running. Schrodinger x86 versions
>: of Jaguar and mpich were used (you normally don't distribute
>: mpich with your releases). The "secret" was to use the mpich
>: version(1.2.5) that [Schrodinger developer] provided on your
>: systems. I think the problems were caused by Jaguar/mpich being
>: built on RH 7.x, but trying to run it with a native mpich for
>: SLES 8 (Opteron cluster OS).
>:
>: Last time [he] and I talked, [he] told me of plans to switch to
>: using dynamic shared libraries for mpich - that should allow
>: local versions of mpich to be used and reduce dependence on
>: the version of Linux running on the cluster.
>
>Hope this is helpful,
>-P.
Peter,
Yes, part of the problem (or solution) is in finding a compatible mpi
implementation for multiple similar but not totally compatible OS
implementations. We have run into the dynamic library requirement as
well. It's an annoyance, having to distribute all the shared libraries
which belong to the compilers which were used to build mpich and the
application which runs on it.
Tim Prince
|