Re: recent papers on MD gotchas with PRNG seeds

From: Peter Freddolino (petefred_at_ks.uiuc.edu)
Date: Sat May 09 2009 - 17:22:43 CDT

Hi Brian,

Brian Bennion wrote:
> I know that NAMD is less vulnerable here than other packages but I don't
> fully understand.
>

Which part don't you fully understand? In brief, NAMD's resistance comes
from having suitable defaults and from its patch structure, which tends
to reduce the extent to which numbers occur repeatedly because if atoms
migrate from patch to patch between restarts, the random number stream
is not, in fact, the same. I can send more details if this is what
you're after.

> Cerutti, D. S.; Duke, R.; Freddolino, P. L.; Fan, H.; Lybrand, T. P. A
> Vulnerability in Popular Molecular Dynamics Packages Concerning Langevin and
> Andersen Dynamics J. Chem. Theory Comput. 2008, 4, 1669-1680
> http://pubs.acs.org/doi/abs/10.1021/ct8002173
>
> Second, this paper shows synchronization of supposedly separate trajectories
> if seeds are not chosen properly and has definite consequences in REMD.
> Even programs that use the clock as a source for the seed could be affected
> by this problem if seed "requests" occur faster than the clock can propagate
> to the next time point.
>
> http://pubs.acs.org/doi/full/10.1021/ct800573m
> Bad Seeds Sprout Perilous Dynamics: Stochastic Thermostat Induced Trajectory
> Synchronization in Biomolecules
>
> Any comments from developers/users would be helpful
>

This could in principle be a problem (although here again in namd the
patch structure will help a bit by breaking synchronization between the
replicas as they experience different patch migration -- one would need
to test it in order to see how strong the effect is in namd). The clock
time used for a seed is in seconds, so whether or not this is an issue
depends on how quickly all of your replica exchange jobs run. If you
have a previous job that you're worried about I'd recommend checking the
seeds used in all the individual replica exchange runs. In the future, I
would **highly** recommend that people add something like
wait 2
to the section of the replica exchange script that spawns the individual
replica jobs to make sure that they all end up with different seeds;
this will cost a minute or so on startup but will remove the possibility
of synchronization. I'll see about putting this into the repex scripts
that are distributed with namd.

Best,
Peter

> Thank you
> Brian Bennion
>

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:50:55 CST