[gmx-developers] replica exchange -- coordinates versus energy

[Mark Abraham]

servaas michielssens wrote:
> hi,
>  
> Again a question about the replica exchange code. So we start in de do_md,
> energies are calculated in do_force and after this coordinates are updated.
> e.g.:
> step=0, energie at step0, update the coordinates
> So at the end of step 0 you have the energy of step 0 but the 
> coordinates of
> step1.
> But when you come to the replica_exchange() (e.g., step=200) you pass the
> potential energy of step200 but the coordinates of step201, is this the 
> case? Or what am I missing?

Yes, that is what is happening.

Ordinarily in GROMACS, forces and energies are calculated in the same 
non-bonded inner loops. It would be wasteful (i.e. poorly load-balanced) 
to calculate them, test for replica-exchange, and then have to calculate 
them again for the replicas that exchanged successfully.

You could perhaps eliminate some of the waste by only calculating the 
energies before the exchange test (since there are energy-only versions 
of the inner loops - but this assumes that there are no hidden 
dependencies of the energy on the force calculation, and I don't know if 
there are). However, since (IIRC) there are no force-only inner loops 
for the non-exchanged replicas, and they'll have to sit and wait for the 
energy+force processors to finish anyway, then assuming each replica 
takes the same amount of time for an energy+force evaluation, you'd 
probably implement an algorithm to do an energy evaluation, do replica 
exchange, then do full energy+force evaluation for all replicas. Thus 
there's a wasteful repeated energy evaluation for (typically) most of 
your replicas.

In a perfect world you could re-distribute the repeat energy evaluations 
over your whole processor array, and then just continue with the 
force-only evaluations but this costs a lot of overhead and will not 
produce numerically-identical results.

... or you can implement the approximate version described above.

Mark