I must have missed the original message.
Chris Samuel wrote:
On Wednesday 04 October 2006 00:37, Douglas Eadline wrote:
This is a non-obvious result many find hard to believe.
That is, MPI on the same node maybe faster than some shared/threaded
mode. (of course it all depends on the application etc.)
We believe we have seen that with LS-Dyna comparing SMP versus MPP (MPI)
variants on a single node.
Actually (my apologies to Doug for not seeing this before and
commenting), the Cray folks have noted this for a number of years,
specifically when they moved MPI code from T3D/T3E back onto Vector
boxes, it behaved better.
The rationale for this is that MPI does in fact impose a very stringent
restriction on sharing in that you must explicitly share data. This
*forces* you into spatial localization, and to a degree, temporal
localization, as you have to get the most out of your data.
OpenMP (and please don't beat me up on this: I like the OpenMP
syntax/usage model better than MPI) doesn't force you to explicitly
share items. The compiler does it for you. If you forget, well, thats
ok, it will default to doing the right thing.
This also means that you may not work as hard to spatially localize your
computation as you would in MPI. MPI requires a more rigid discipline,
and if you have ever debugged an MPI program, it also requires lots of
antacid. In the age of gui debuggers, I find myself using print
statements, and barriers like MPI_Barrier... (yeah I know about Etnus,
but I don't have enough need to justify the purchase).
I don't have enough experience with PGAS languages to see whether or not
they provide the ease of use of OpenMP with the rigid sharing and
localization of MPI. I still think Chapel looks neat, would love to
play with a compiler for it. I think UPC might even run on some
clusters, last I looked you had to compile it remotely (some sort of web
page thingy).
Joe
cheers,
Chris
------------------------------------------------------------------------
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
--
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: [EMAIL PROTECTED]
web : http://www.scalableinformatics.com
phone: +1 734 786 8423
fax : +1 734 786 8452
cell : +1 734 612 4615
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
