<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 3.2//EN">
<HTML>
<HEAD>
<META HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=iso-8859-1">
<META NAME="Generator" CONTENT="MS Exchange Server version 6.5.7652.24">
<TITLE>RE: [gmx-users] Free energy calculations and speed problems</TITLE>
</HEAD>
<BODY>
<!-- Converted from text/plain format -->
<P><FONT SIZE=2>Hi Justin,<BR>
<BR>
You shouldn't use PME in vacuo. Just use pbc=no and infinite (i.o.w.: no) cutoffs.<BR>
<BR>
I'm not entirely sure why PME takes such a long time, but I can give a very wild guess. It may have something to do with the fact that, if I remember correctly, the Fourier transform of one (or a few) large charge spike(s) can only be represented to a certain precision by taking very high frequencies into account. Therefore the FFT has to calculate ridiculously long to reach the desired accuracy.<BR>
<BR>
Hope it helps,<BR>
Jeroen<BR>
<BR>
<BR>
<BR>
Date: Mon, 19 May 2008 15:04:51 -0400<BR>
From: "Justin A. Lemkul" <jalemkul@vt.edu><BR>
Subject: [gmx-users] Free energy calculations and speed problems<BR>
To: gmx-users@gromacs.org<BR>
Message-ID: <1211223891.4831cf53f0a46@webmail.vt.edu><BR>
Content-Type: text/plain; charset=ISO-8859-1<BR>
<BR>
<BR>
Hi all,<BR>
<BR>
I'm running some free energy calculations (in vacuo and in solution) for a set<BR>
of molecules I am trying to parameterize. All the molecules are fairly small,<BR>
about 30-35 atoms in each. The solvated calculations go smoothly, and rather<BR>
quickly (~5 hr/ns on two nodes). I am running into a problem when I do the in<BR>
vacuo calculations. Since I am constraining all bonds, I am running things on<BR>
a local machine in the lab (single core AMD64 Opteron). I get as far as NVT<BR>
equilibration (removing charges only) before I give up and kill it, since it<BR>
takes so long. A 10-ps NVT run takes about 20 minutes, but that adds up to<BR>
over 5 days for my 5-ns production run! The machine is no slouch; I have used<BR>
it to run other small NVT/NPT/MD runs with good performance. I get similar<BR>
performance on another machine that is even faster (a single CPU of the 2-CPU<BR>
node on our supercomputer). The installations of Gromacs on both are fine, as<BR>
far as I can tell, and both have been used extensively.<BR>
<BR>
I am using Gromacs version 3.3.3, and I have attached my .mdp file and the tail<BR>
end of my log file for reference. If anyone notices anything obvious that I'm<BR>
doing wrong, I would be grateful if you pointed it out! I see in the log of<BR>
the slow run (in vacuo) that the 3D-FFT calculation is dominating the CPU<BR>
usage. Is that indicative of any problem?<BR>
<BR>
Thanks for your attention.<BR>
<BR>
-Justin</FONT>
</P>
</BODY>
</HTML>