Keep in mind that the Petsc faq mentions:
A fast, low-latency interconnect; any ethernet, even 10 gigE cannot provide
the needed performance.
which seems foolish on the face of it, since 10G can be fairly low latency.
for instance, Myrinet quotes 2.3 us for mx-over-myrinet on their 10G nic,
and 2.63 for mx-over-ethernet (same nic and host, fujitsu eth switch).
that's not the lowest latency interconnect, but it's a far cry from
50 us Gb latency...
I wouldn't assume that your workload will scale with gigE to 16 nodes.
I haven't scrutinized petsc docs, but I would expect the answer to
depend on what you're doing and also how big a chunk you can put onto
a single node. standard volume/surface argument. from 10,000 feet,
it _looks_ like petsc could be used in a pretty high-work-per-communication
manner.
tested DDR3, but all the tests I've seen show minimal performance improvements
at substantial price increases.
my understanding is that ddr3-ddr2 changes are largely system-level
engineering (lower voltage/power, some changes in signal handling to
improve snr and scalability.)
_______________________________________________
Beowulf mailing list, [email protected]
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
