Cables

[Florent Calvayrac]

Keith Murphy wrote:

> Like you I do not want to generalize.  Of course if cheaper hardware can do
> the job it should be used.  However many Beowulf projects could certainly
> use a faster interconnect and SCI (or Myrinet) will improve their
> performance.
>
> If you use 2D Torus you will not need any switches, there is a 96 node 192
> server SCI system running in Paderborn Germany with no switches rated at
> 86.4 GigaFlops.
>
> >> In the meantime it is the
> >> fastest interface available today and makes an ideal Beowulf interface.
> >
> >Please avoid generalizations. It isn't ideal if much cheaper hardware
> >can do the same job -- maybe your application doesn't need that much
> >network, or isn't sensitive to latency? And you still can't buy huge
> >SCI switches, which makes it inferior to Myrinet for large systems.
> >
> >Linux drivers, or better yet, open-source drivers will be a big step
> >forward for SCI and cluster computing. But you still have to look at
> >price/performance.
> >
>

I had the occasion thanks to our German colleagues to test
the 32 processors SCI cluster in Paderborn, and I had a very sobering
experience.
I have developed a parallel Density Functional  program under MPI where
the wavefunctions are distributed among the processors, and to
a good approximation the parallel work amounts to repetitively
summing up the density on the discretization grids. It seems (but it might be
wrong)
that the corresponding MPI_ALLREDUCE are very defavorable
under Scampi, and indeed I get a better performance with a TCP/IP
Fast Ethernet network, because it seems that the reductions/distributions
are way better with such a communication network.

Any comments ?

What is BIP by the way ?







--
Florent Calvayrac                          | Tel : 02 43 83 32 72
Laboratoire de Physique de l'Etat Condense | Fax : 02 43 83 35 18
UPRESA-CNRS 6087                           |
Universite du Maine-Faculte des Sciences   |
72085 Le Mans Cedex 9