Re: [AMBER] Required time for a typical MD run

From: David A Case <>
Date: Mon, 22 Jun 2015 20:53:44 -0400

On Mon, Jun 22, 2015, Morteza Chehel Amirani wrote:
> I'm wondering how long a typical MD run will be using Amber. It seems
> something is wrong in my relatively short-size model as it takes a long
> time to be simulate for 50 ns. Even if I split the simulations into 10
> stages, it is beyond my expectations.

As already noted, it would be helpful to know what your "expectations" are
here; also how many cores you are running on.

Since the advent of GPUs, we have not kept up so much on CPU benchmarks, but
some good hints can be gleaned from the GPU benchmark page:

Your 10,000 atom system, with a 12 Ang. cutoff, is very roughly the same
as a 23,000 JAC system with an 8 Ang. cutoff. You should look at the
third DHFR example benchmark (with a 2 fs time step, like you have). That
benchmark shows pmemd.MPI running on 20 (modern) cores achieves 17 ns/day,
or three days to get to 50 ns.

If you are using a different number of cores, you would need to make
adjustments; if you have older CPU's, that can also make a significant
difference. But this should give you a *very rough* idea of whether or not
the performance you see is in line with what others are getting.

If you indeed have hardware that is a bit older, the benchmarks here might
be of interest:

These show that, even on supercomputers of a half-generation ago, taking a
week to get to 50 ns is rather common.)

(The more recent, 2012, benchmarks from Daresbury, which are cited on the
Amber web site, seem to have gone missing. Does anyone on the list know where
they can be found?)

...hope this helps...dac

AMBER mailing list
Received on Mon Jun 22 2015 - 18:00:02 PDT
Custom Search