> Our problem is now that there are very few 21264 around, that it is
> unclear what kind of boards will there be in the near future and how
> they will perform, and how much they will cost in the end.
The XP1000 is Compaq's single-processor workstation. It is out now and
should have identical performance (and clock speeds) to the
single-processor OEM board (264SP), which should be out "real soon
now".
I can buy "bare" XP1000s for $4,300. I don't think pricing has been
announced yet for the 264SP.
> As the deadline for our project nears, I would appreciate any hints,
> tips, or comments. We would very much like to hear from people that
> use (or are going to) the 21264 for scientific computing cluster
> applications, especially with a high-speed interconnect like
> Myrinet.
There are several such clusters out there: CPLANT, mine, RWCP...
> Or maybe somebody knew whom to contact at Compaq/DEC about
> this?
I have contacts with the Digital Workstation group, if you like.
However, they are in the US.
> One point is here that we will need MPI directly over the netword
> card, without the TCP/IP stack. As far as we have learned, there is an
> implementation from RWCP/SCore, and a commercial one from ParaStation
> in Karlsruhe (Germany), and more maybe to come with VIA, or maybe not.
I would recommend trying the Myricom-supported GM driver and mpich/GM
software. There are similar mpich implementations from RWCP (mpich/PM)
and the BIP people in France. Myricom has a VIA interface on top of
GM, but there's no reason to use it if you're going to be using MPI
anyway.
-- g
