Dear Marios:
AMBER 5 can run on the native shared memory of SGI. You can give it a try.
Many factors can contribute to the performance of the parallel code,
including the load of the computer itself since parallel code requires
certain level of OS support and more so than typical serial code. For
instance, it is not very efficient to have a few CPUs running ahead of
everybody else. It is even less efficient if one or two CPUs run behind
everybody else. In your case, to my limited knowledge, it sounds like that
you were running the program when somebody else was also running
his/her favorite program. I'd be surprised if this was related to the
issues of MPI/OpenMP/shared memory and would love to know your source of
information.
We've also tested the MPI version against the shared memory version on
SGI. Our conclusion was that they scale and perform almost exactly the
same.
yong
Department of Pharmaceutical Chemistry Phone: (415) 476-8291
University of California Fax: (415) 502-1411
San Francisco, CA 94143
On Mon, 13 Mar 2000, Marios Philippopoulos wrote:
> Hi everyone,
>
> I have been told that MPI parallelization is not optimal on SGI
> machines, whereas OpenMP is. From my own recent experience, I have found
> that my jobs take about twice as long to complete as suggested by the
> actual cpu time reported in the output file. I just checked the AMBER6
> manual, and in the INSTALL.parallel file (in the "Shared Memory"
> section) it is stated that:
>
> "Under development is an OpenMP version which should be portable to a
> variety
> of shared memory multiprocessor machines."
>
> Would anyone know when the OpenMP implementation might be ready for use?
> Also, any comments on any of the above are welcome.
>
> Thanks,
>
> Marios
>
>
> --
> Marios Philippopoulos, PhD
> Structural Biology and Biochemistry Programme
> Hospital for Sick Children
> 555 University Avenue
> Toronto Ontario Canada M5G 1X8
>
> mphilip_at_sickkids.on.ca
>
>
>
Received on Mon Mar 13 2000 - 12:21:07 PST
