[gmx-users] issue in replica exchange

[Mark Abraham]

On Thu, May 2, 2013 at 12:58 PM, XAvier Periole <x.periole at rug.nl> wrote:

>
> I saw that redmine report, which could be related but it seems to happen
> only for runs done outside the domain and particle decompositions.
>
> I'll fill up a red mine.
>
> Anything I could do to help speeding the fix?
>

What'd be really nice is some thought on how one can demonstrate that the
implementation of the exchange matches what would be expected from the
theory. For T-exchange under NVT, it is sufficient to rescale velocities
and quantities derived from them by the correct factor. That includes
various things like T-coupling history and integrator half-step quantities
(and does REMD with leap-frog make sense anyway?). For NPT, there's
probably also some P-coupling quantities to scale, and the box to exchange.
Anything I've missed? Hopefully virial contributions don't matter either
way?

Perhaps a decent first step is to hack the code to do a "self exchange," by
clearing the entire state and rebuilding with what would/should be received
from an exchange with a hypothethetical replica in an identical
pre-exchange state. Only if the code can do that (i.e. mdrun -reprod
produces a trajectory indistinguishable from a run that does not attempt
this self exchange) is it worth considering proper state exchanges, and the
process of making the code do the former should illustrate what is required
for the latter.

Mark