Tim Bradshaw  <tfb@cley.com> wrote:
+---------------
| ...as I know no one really knows how to write efficient stuff on a
| cache-coherent shared-memory box which is also efficient on a
| shared-nothing cluster.
+---------------

Well, one data point: The SGI versions of the MPI libraries (I think we
call ours "MPT"?) are tuned to use local shared-memory mailboxes instead
of network calls if the target of the MPI operation is on the same system.
This means, for example, that you get a *lot* better MPI performance by
clustering a few "fat" (many-CPU) nodes, compared to a whole bunch of
small nodes, assuming the same total CPU count.

Ob. c.l.l.: I only wish I could find a "killer" Lisp application that
could make good use of a 256-CPU (or bigger) system. Most of the huge
HPC apps are still Fortran... (*sigh*)


-Rob

-----
Rob Warnock, 41L-955		rpw3@sgi.com
Applied Networking		http://reality.sgi.com/rpw3/
Silicon Graphics, Inc.		Phone: 650-933-1673
1600 Amphitheatre Pkwy.		PP-ASEL-IA
Mountain View, CA  94043