Date: Fri, 4 Oct 2002 09:26:07 -0700
From: "David A. Case"
Subject: Re: Different numbers on different architectures.

On Fri, Oct 04, 2002, David Smith wrote:
>
> I also have a test job of my own I have been using. This is a
> perturbation in a small box of water with periodic boundary conditions
> in the NTP ensemble. I use internal constraints via intr=1 and itor=2. I
> equilibrate this box using gibbs at lambda=1 by setting NSTMEQ >
> NSTLIM. I then use the restart file from this run to start the
> perturbation (51 2ps steps with electrostatic decoupling and
> Thermodynamic Integration, total 102ps for the electrostatic part)...

>
> In summary:
>
> Water box.
> compiler/architecture F_energy
>
> g77/i686 0.40692
> pgf77/i686 0.39719
> compaq/alpha 0.41389
> mips/sgi 0.37039
>
> I also have an 50ps FEP (double wide) run of a protein in a droplet
> which I didn't run under pgf yet but I do get:
>
> Protein.
> complier/architectuere DG(forward) DG(reverse)
>
> g77/i686 0.12555 -0.16087
> compaq/alpha 0.17262 -0.21105
> mips/sgi 0.02955 -0.07174
>
>
> I know that both of the runs are quite short but I was expecting a
> little better agreement.

I think the variation you see is quite normal.

MD of these systems is "chaotic" (in a loose sense): trajectories compiled
in different ways
(even starting at the same point) will diverge due to accumluated
roundoff errors, and will soon bear no specific relation to one another,
although they are all exploring the same potential energy surface.
Hence, you should only expect that only averages should agree, and you
are seeing results that agree within 0.1 kcal/mol or so. I suspect if you
ran a second 102 ps simulation with the same compiler/machine, but with
ever so slightly different starting conditions, you would get results that
are equally far away from the first simulation.

>
> In general, I am extending my runs to times long
> enough to see convergence.

The amount of sampling you would need to do to get convergence and
reproducibility below 0.1 kcal/mol is surprisingly large. I certainly
encourage you to explore these issues for yourself, but don't expect this
to be easy. Very few published studies on proteins or peptides have errors
less that 0.5 kcal/mol.

..hope this helps...dac