Re: [ESPResSo] Hardware Requirements for Espresso and Other Things

[Torsten Stuehn]

Hi Lorenzo,

just one more comment on what Olaf wrote:

Olaf Lenz wrote:
> If you want to do really huge simulations that require more than the
> available number of cores, you will also need a fast network between
> them (Infiniband or the like). However, this very much increases the
> cost of a cluster.

Many modern mainboards that are used in clusters today already have an 
infiniband
network card onboard. This makes it easy to connect two nodes via a simple 
infiniband
cable, without the need of an infiniband switch (which is expensive).
So you can easily build a system with 16 CPU Cores ( 2 nodes a 2 CPUs a 4 Cores 
)
for approx. 10T EUR. For many scenarios ESPResSo will run very well on such a 
system.

> The even more expensive big shared memory machines are not really worth
> it, as ESPResSo only uses MPI parallelization, so the shared memory can
> not be effectively used.

I do not agree to that. The native MPI implementations of the big expensive
shared memory machines (e.g. IBM Regatta Clusters) are usually very well
optimized for the underlying architecture.
ESPResSo does indeed profit from this. We did benchmarks and simulations on such
systems with up to 1 mio particles and 1024 CPUs. For many scenarios ESPResSo
scales very well on these machines (there are exceptions, though).


Greetings,
Torsten