On Tue, Dec 25, 2012 at 5:41 PM, Amparo Garcia Lopez <
Amparo.GarciaLopez.unige.ch> wrote:
> Oh, that sounds worrying... I know other people used our machines in the
> past for parallel computing. I'll investigate your suggestion, though, and
> will confirm if that's the case. Thanks very much again!
>
This is quite common among most (all?) MD codes. Even the high-performance
MD codes that boast scaling to many thousands of CPUs only do so for
infiniband or faster connections. I don't know of any MD codes that scale
beyond a handful of separate nodes (e.g., 2 to 4 nodes) that are only
connected by gigabit ethernet (and the more CPUs are present on each node,
the more the interconnect must be shared, potentially reducing parallel
scaling even more).
This only holds true for standard molecular dynamics, though. pmemd and
sander both have highly scalable methods you could use to minimize
communication between nodes (and get good scaling as a result). For
example, replica exchange methods involve separate replica simulations that
exchange information between each other only occasionally (this process of
information exchange leads to more rapid convergence of equilibrium
properties if done 'correctly'). As a result, you can run one replica on
each node to optimize performance with gigabit ethernet interconnect.
There are examples of other methods that allow you to run a large number of
relatively short calculations, which are more efficient on clusters with
slow interconnect (e.g., umbrella sampling and steered MD). GPUs are also
a cost-efficient way to increase simulation speed.
HTH (and Happy Holidays),
Jason
--
Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
352-392-4032
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Dec 25 2012 - 18:30:02 PST
