sp at scali.com
Mon May 6 15:30:58 PDT 2002
On Mon, 6 May 2002, Greg Lindahl wrote:
> On Mon, May 06, 2002 at 04:35:35PM +0200, Joachim Worringen wrote:
> > > > AFAIK the PMB benchmark uses rank 0 and 1 when doing ping-ping and
> > > > ping-pong tests,
> > >
> > > So a "round robin" default breaks PMB if you aren't careful. Which is
> > > unfortunate.
> > Do you have a definition of "breaking PMB"?
> It is clear from the quote: PMB only reports shared memory bandwidths
> and latencies for the ping-pong case, and that's broken. This was
> noted by everyone else posting in this thread; I'm sorry you missed
> it. If you have other simple questions like this one, I'd suggest
> sending personal email so that you don't waste the list's time.
Let's not turn this one into a flamewar, ok ? The original poster
(Kevin Van Workum) asked a legitimate question and I think he got his
answer (so basically this stuff is now OT).
But just to make the record straight, the numbers for ping-pong Joachim
reported in the SCI-SMP file was _not_ intra node (i.e shm) communication,
but actual SCI communication between two different nodes (as Patrick
suspects). However, it's not the 2 process results that interest me, it's
the larger ones (which test the scaling possibilities of the MPI
implementation and the interconnect). How are the collective numbers
on commercial Myrinet MPI implementations (MPI/Pro ?) compared to
MPICH (someone care to give us a comparison) ?
As some of you may have noticed, I've dropped my "marketing" based
.signature to avoid "offending" someone.
More information about the Beowulf