Re: [AMBER] Scale sander.MPI to 64 ncpus - Amber11

From: case <>
Date: Thu, 17 Feb 2011 15:47:20 -0500

On Fri, Feb 18, 2011, Leong Wye Kit Damien wrote:
> I'm helping an amber ( Amber11 ) user to scale his sander.MPI
> computation from 32 ncpus to 64 ncpus. Everything seems to work fine
> when we use 32 ncpus, once we try to scale up to 64 ncpus ( eventually,
> if possible 128 npcus ), the results did not compute. I believe there is
> a limit to 256 ncpus in sander.MPI ( please correct me if I'm wrong ).

I don't have any good response to the specific question about why sander hangs
at a certain number of processors, but do have some general comments:

1. Don't test with minimization. Sander uses a completely different method
of parallelization for minimization than for MD. Since minimization is
generally only need for a relatively few steps at the beginning, it has not
been exercised very much a high processor counts. (Of course, if there is a
bug it should get fixed....)

2. Even when you find the problem, you will almost certainly be disappointed
with the results. There is probably no hardware combination where sander will
work well on an explicit sovlent simulation at such high processor counts.
Really try to see if you can't use pmemd, instead.

3. You don't mention the interconnect on your system. If it is not a
high-performance one (e.g. infiniband or similar) even pmemd will work very
poorly at 64 threads.

...good luck...dac

AMBER mailing list
Received on Thu Feb 17 2011 - 13:00:06 PST
Custom Search