Beowulf & Fluid Mechanics

[Josip Loncaric]

Patrick Geoffray wrote:
> 
> It's maybe another difference : MPI/Pro provides a shared memory device
> for local communications and i don't think that MVICH does.
> 
> I do not know if MPI/Pro provides the shared memory device on Linux (the
> last time i have heard about that, it was only supported for Windows)
> and if it was enabled for these tests...

The tests were done with MPI/Pro 1.5b7-2tv and MVICH 0.03, both using
VIA transport over Giganet network fabric.  The MVICH version was built
with the MPICH 1.1.2 ch_p4 device only, but I am not sure if the MPI/Pro
version used shared memory.  

Our experience with LAM over Fast Ethernet indicates that shared memory
may have great benchmarks but need not help much with actual
applications.  Sometimes (e.g. in certain parallel computer graphics
applications, see http://www.icase.edu/coral/mpl/Nov15.99.html) it is
actually better to use the TCP-only flavor of LAM.   TCP within a box
can be very fast, perhaps fast enough so that further improvements
resulting from shared memory are not always obvious.  Finally, even
*with* shared memory (either usysv or sysv flavor), LAM 6.3.2 has about
25% performance penalty on SMP machines using Fast Ethernet.  This
suggests that the shared memory device is not a relevant factor.

Sincerely,
Josip

-- 
Dr. Josip Loncaric, Senior Staff Scientist        mailto:josip at icase.edu
ICASE, Mail Stop 132C           PGP key at http://www.icase.edu./~josip/
NASA Langley Research Center             mailto:j.loncaric at larc.nasa.gov
Hampton, VA 23681-2199, USA    Tel. +1 757 864-2192  Fax +1 757 864-6134