this is pretty far from beowulfry, but...
I was wondering what are the bandwidth enhancements or limitations from CPU
to cache and system memory in HP's Superdomes. I mean, what should be better
for very intensive memory access based applications, taking a Superdome with
64 CPUs or using mx2 modules and 128 CPUs?
mx2 just doubles the number of cpus without changing anything else; a bit
like going dual-core. as such, it has the same drawback as dual-core:
cpus see half the effective memory bandwidth (same peak/aggregate, just
divided more ways.)
so, unless "memory intensive" somehow also means "cache friendly",
mx2 is not what you want. more cell boards bring both more memory bandwidth
and more memory capacity. mx2 is basically a way to economize on cell boards...
IMO, you should be planning to abandon the big-SMP approach, and looking
at how your application can run on a cluster instead.
regards, mark hahn.
_______________________________________________
Beowulf mailing list, [email protected]
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
