[Beowulf] "dual" Quad solution from Tyan

[Mark Hahn]

>   I'm responsible for the aquisition of my labs new cluster. I already had 

what's the purpose of the cluster?  without a well-defined purpose,
there's no way to evaluate the options...

> here, in Portugal. The target machine I was looking for was 8 node with 
> dual processor Opteron (16 CPU). The machine that was at the presentation 

OK, so 8 dual-socket nodes will deliver an aggregate memory bandwidth
of 16*6.4 GB/s, and a capacity of 64 dimms.

> http://www.cybernex.co.uk/tyan_vx50_quad_opteron_servers.htm

afaikt, this is a ladder design: 

c-c-c-c
| | | |
c-c-c-c

which is 5 hops corner-to-corner, with a bisection of 3.2 GB/s per node.
half the aggregate memory bandwidth and capacity.

> which means 8 CPU, dual core. And all are connect with hypertransport 
> which meand I got ready of latency I figure. So...

well, HT latency is better than gigabit, if that's what you mean.
I've never read a good evaluation of MPI implementations on this kind 
of hardware.

don't forget that dual-core is lower clocked.

> The professor head of the lab wants me to know "if that is so good why 
> isn't everyone buying one?"

who says it's so great?  the HT fabric is pretty far from a "perfect"
low-latency 16-port crossbar, for instance.  also, is it really cheaper?
how about if you compare to four 2-socket DC nodes?