Hi Anselm,
Yes, interesting that network communications indeterminacy + rounding error
drags us into this realm; I think it is something to think about and expect
any time computation is done in parallel using floating point. Note also
that you would see differences if you were using a different random seed to
start a trajectory (of course). The "macroscopic effect" here, ie., why it
is so dramatic, is that at least one atom has a sufficiently different
trajectory that at some point it acquires coordinates that are so large they
cannot be represented in the fixed format of the restart file (f12.7); this
is not that unusual in trajectories that have been run for a while - suck
the amber manual into a pdf reader and look for "iwrap" to see what you can
do about it (basically, do a run with iwrap 1 to get coordinate wrapping,
but note that you lose some trajectory info at this point - ie., you can't
calculate diffusion through this operation). I would expect your other runs
to fail pretty soon too (sorry I did not mention this part earlier) if you
don't use iwrap to bring all the atoms back inside the unit cell, so to
speak.
Regards - Bob
----- Original Message -----
From: "Anselm Horn" <Anselm.Horn.biochem.uni-erlangen.de>
To: <amber.scripps.edu>
Sent: Tuesday, December 18, 2007 4:10 AM
Subject: Re: AMBER: pmemd: same rst, different result?
> Dear Bob,
>
> thanks a lot for your extensive answer and the detailed analysis of the
> underlying computer numerics.
>
> That numerics do matter I had learned from the Numerical Recipies and
> some other Fortran books. However, I had always thought of that problem
> as something microscopic having little impact on everyday-work and thus
> was astonished to see the macroscopic effect (=pmemd does not start
> off): The computational scientist's variant of Schroedinger's cat.
>
> Thank you also for pointing me to the reference of Goldberg, which I
> promise to read soon.
>
> Best regards,
>
> Anselm
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber.scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Wed Dec 19 2007 - 06:07:29 PST
