On Sat, Jan 30, 2021, Charo del Genio wrote:
>
>I strongly suspect that this behaviour is due to the REMD engine computing the energies at step numbers that are not multiples of nrespa, thereby getting meaningless results.
>So, is this a bug, or is RESPA not supposed to be used in REMD runs?
Thanks for your investigation here. It sounds like a bug.
But there is a bigger point worth discussing here. Is RESPA really worth
it in Amber simulations, whether or not they use replica exchange? Amber
uses the "original" RESPA idea, which is know to have stability problems
for outer time steps somewhere around 8 fs (for the types of systems we
simulate). I think most Amber developers use hydrogen mass repartitioning
and SHAKE, which allows a 4 fs time step. In that circumstance, even an
nrespa of 2 might be dangerous, and in any event would provide only modest
speedup. So, I guessing that the combination of REMD + RESPA has not been
tested much.
(REMD people: please chime in here!)
So: we should either fix the REMD problem, or at least prevent people from
using REMD + RESPA (assuming that your analysis is indeed correct). You
could see how much of a performance hit you get by turning off RESPA -- that
might help us decide what course to take.
As an aside: there are very promising "isokinetic" integrators that can
avoid the resonance problems with RESPA, and allow quite long outer time
steps. We might want to consider implementing those. But the code
complexity can be hard: the forces in the outer time step have to be both
smoothly varying and expensive to compute. (The long-range Lennard-Jones
contributions come to mind here, but other codes concentrate on splitting
"direct" non-bonded interactions into buckets, to reduce the pair-list size
for the inner time steps.)
...thanks...dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sat Jan 30 2021 - 07:30:02 PST
