Re: [AMBER] AMBER/sander parallel performance problem

From: case <case.biomaps.rutgers.edu>
Date: Thu, 28 Jul 2011 20:18:43 -0400

On Fri, Jul 29, 2011, Jyh-Shyong wrote:

> I built parallel version of sander with both mvapich2-1.5 and
> openmpi-1.4.3.

The parallel performance of pmemd is *much* better than that for sander,
especially as the number of threads increases. Everything depends a lot on
what type of job you are running, but I'm not surprised to see really bad
performance in going from 48 to 192 cores. Pmemd should do much better
(again, depending somewhat on the details of the job.)

Also, I think that you need mvapich to get good performance across nodes with
infiniband; and the mvapich needs to be infiniband-aware.

...dac


_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Jul 28 2011 - 17:30:05 PDT
Custom Search