Re: [AMBER] running simulation using increased number of cores

From: Jason Swails <>
Date: Sat, 14 Sep 2013 09:53:47 -0400

On Sat, Sep 14, 2013 at 12:13 AM, onetwo <> wrote:

> Dear All,
> We have just installed RACK server in our lab, in which we have installed
> AMBER software in
> parallel. Server shows that it has 160 cores. When I run simulation job, I
> give np 160,
> equilibration job crashed. Then I tried using np 8, then it worked, then
> again I increased np to
> 32, it is running fine, but for 4ns production run for 321 amino acid, it
> is taking 72 hours to
> finish. When I used to run it on 8 cores on our previous workstation, it
> used to take same time.
> I tried on server np 80 also, still it shows more than 70 hours to finish.

A couple comments here. First, 'Amber' is a suite of programs, not a
single one. You need to tell us _which_ program you were using.
 Typically, pmemd.MPI gives better parallel scaling than sander.MPI (and it
gives better single-processor performance, as well). If you are not
using pmemd.MPI, you should start using that program instead.

Second, the interconnect between the nodes is very important -- unless you
have very fast infiniband, you will not be able to use multiple nodes (as
you've seen in the lack of speed-up).

> Please suggest, if I am missing something. and how to utilize all the
> cores for faster simulation

You may not be able to use all cores to speed up your simulation. This
will be true of all MD codes.


Jason M. Swails
Rutgers University
Postdoctoral Researcher
AMBER mailing list
Received on Sat Sep 14 2013 - 07:00:02 PDT
Custom Search