[ale] Syntax problem

[Joe Knapka]

Dow_Hurst <dhurst at mindspring.com> writes:

[snip]

> Just to pass on some good news: The cluster is Opteron 250 CPUs in
> dual CPU 1U configuration.  We have 19 slave nodes and one head
> node.  The switch we have is a SMC 8624T.  The special fast
> interconnect is the Ammasso GigE RDMA 1100 NIC.  Using the molecular
> dynamics package NAMD, I've found that with the NAMD benchmark input
> file of ~92K atoms, the apoa1 protein, that we can get using Ammasso
> MPI libraries a full nanosecond of simulation time in 0.96 days.  If
> I try to scale up using just standard TCP communications and not use
> the Ammasso MPI, we can do a full nanosecond in 1.31 days.  What
> this means is that with our smaller cluster of 19 nodes we are still
> benefitting very well using the fast interconnect from Ammasso.
> When you think that we will save 462 minutes per nanosecond
> simulation time and we need a lot of nanoseconds to actually see
> changes, the fast interconnect shows it's worth.

Holy cow. How does one test such a system? I mean, it would suck to
simulate a couple of hundred nanoseconds, only to find that you're
getting bogus results. I'm trying to imagine how the developers of
NAMD validated their code. Are there simpler problems that one
can simulate much more quickly?

Also, I wonder how well your performance will scale with additional
CPUs?

-- Joe Knapka