[gmx-users] restarting from cpt and binary identical

[Mark Abraham]

On 4/21/2011 8:00 AM, Peter C. Lai wrote:
> I have a fairly stupid question about restart from a checkpoint and binary
> identical messages, probably because I can't understand the doco.
>
> When I restart (or extend) a double-precision run from a checkpoint involving
> either dlb or even a different decomposition count (say I changed the number of
> parallel threads), I get a message that says it restarted but the
> trajectory is not binary identical.
>
> My question is: I assume this trajectory will continue from the last set of
> double-precision coordinates and velocities in the cpt file? What is the
> loss in precision/error propagation of the resulting trajectory as compared
> to if I started the entire run from the begining? (say I am extending my
> trajectory from 1ns to 10ns). I guess I understand that both trajectories
> would be "equally valid" but not binary identical, but what does this mean
> in terms of things like ultimate RMSD measurements and so forth? I don't
> necessarily care that the two concatenated trajectory files will not cmp to
> zero.

These issues are discussed here 
http://www.gromacs.org/Documentation/Terminology/Reproducibility

There's no *loss* upon continuation with a new DD, it's just that it is 
almost guaranteed to be different from what would have resulted from 
continuing with the old DD. In the limit of an infinite situation, both 
would converge to the same ensembles.

Makr