[gmx-developers] Implementing Replica Exchange

[Senthil Kandasamy]

Just as I suspected, it is not going to be too hard to implement this. I
will contact Prof. Tieleman to learn how he plans to implement REMD and
maybe I can chip in a little. I am not very familiar with Parallel
programming and so am a little confused about the best way to organize
the code. 

If I run all the replicas as separate processes, at exchange time, all I
really need to exchange will be potential energies for the metropolis
trial and a Temperature if the trial is accepted. I think transferring
temperatures will be easier than transferring coordinates. But
transferring coordinates between replicas might yield options for a more
generalized parallel tempering algorithm.

Senthil




On Fri, 2003-09-19 at 23:18, Erik Lindahl wrote:
> Hi Senthil,
> 
> It's trivial since the individual simulations are essentially 
> disconnected.
> 
> Just run the simulations in parallel for N steps. Check the results, 
> exchange if necessary, and start new simulations.
> 
> Actually, I would probably recommend implementing it as a higher-level
> script that calls the gromacs programs, so you can parallelize each of
> the underlying replica simulations independently. Otherwise you'll get
> a complicated two-layer situation if you are running e.g. 8
simulations 
> in parallel where each of the simulation wants to use MPI
independently 
> of the others. .
> 

actually it is mainly implemented (except for the replica bit).
mdrun -multi 
will start N simulations with topologies topXXX.tpr and rename all the
other output likewise. It is currently only useful in special cases but
if you implement the communication of coordinates (for which there are
routines already) it should be quite straightforward.

> Cheers,
> 
> Erik
> 
> 
> On Friday, September 19, 2003, at 02:07 PM, Senthil Kandasamy wrote:
> 
> > Hi Developers,
> >
> > I am interested in doing some replica exchange md simulations of  
> > lipid/peptide systems.
> > How easy or difficult do you think will it be to implement it in 
> > GROMACS?
> > I would probably have 8 replicas, each at  a different temperature, 
> > exchanging configurations for enhanced sampling.
> > Ideally, there would be one replica per processor, with
communication 
> > between processors happening every n (
> > ~100) steps, to attempt exchanges.
> > So communication is rather limited and it should not affect the 
> > efficiency.
> > I know of people who do REMD using both CHARMM and AMBER.
> > So I guess implementing it in GROMACS should not be THAT difficult.
> > Any ideas? If the developers think that it is not that difficult, I 
> > can devote most of my time to doing this.
> > Or else, I will just switch to AMBER for this particular problem, I 
> > think. (sigh!!)
> >
> > _______________________________________________
> > 
-- 
Groeten, David.