[gmx-users] HREX simulation

Mark Abraham mark.j.abraham at gmail.com
Fri Aug 28 17:19:17 CEST 2015


Hi,

On Fri, Aug 28, 2015 at 5:02 PM Michael Shirts <mrshirts at gmail.com> wrote:

> > I am new to doing replica exchange simulations.
> > I am trying to understand the empirical transition matrix printed in the
> > log files. I tried exchange intervals 2 ps and 0.2 ps. However, later I
> > read gromacs recommends exchange interval not to be under 1ps.
>
> I don't know that there are firm rules for this.  It seems like faster
> exchange is generally better, up to a point (if there are specific
> code reasons one might thing this is not the case, anyone else on the
> list, speak up).
>

The exchange attempt forces replicas to synchronize, and this forces the
whole simulation system to pay the cost of any localized slow-down (e.g.
network traffic congestion, internal simulation load imbalance from
diffusion, even throttling of core clocks from heating effects) at each
synchronization point. Longer periods between exchange attempts use
resources more efficiently if the observed variations are approx random.

That ought to be a problem, but in fact the replica-exchange implementation
already synchronizes much more frequently than it should (
https://gerrit.gromacs.org/#/c/4312/ needs more testing before this is
solved).

Mark

> the system is protein in water and I am only showing parts of the matrix:
> >
> > Exchange interval: 0.2 ps
> > Repl            1            2            3            4             5
> >
> > Repl  0.9528  0.0472  0.0000  0.0000  0.0000
> > Repl  0.0472  0.9017  0.0511  0.0000  0.0000
> > Repl  0.0000  0.0511  0.8946  0.0543  0.0000
> > Repl  0.0000  0.0000  0.0543  0.8898  0.0558
> >
> > 1. If I look at line 3 ""Repl  0.0472  0.9017  0.0511" does this mean
> > exchange between replica 1, 2 is accepted 4.72 %  and exchange between
> 2,3
> > is accepted 5.11 % ?
>
> Yes.
>
> > Repl            1            2            3            4             5
> >
> >
> > Repl  0.9504  0.0496   0.0000  0.0000  0.0000
> >
> > Repl  0.0496  0.8982  0.0522   0.0000  0.0000
> >
> > Repl  0.0000  0.0522  0.8942  0.0536   0.0000
> >
> > Repl  0.0000  0.0000  0.0536  0.8888  0.0576
> >
> > 2. With exchange interval 2 ps, I am getting slightly better acceptance
> > rate. For example, again looking at line 3 "Repl  0.0496  0.8982  0.0522"
> > the acceptance rate for exchange 1 to 2 is 4.96 %, and he acceptance rate
> > for exchange 2 to 3 is 5.22 %
>
> Those are statistically insignificant results.  The success rate
> exchange should not change as the interval changes, which is what you
> are seeing, which is good.
>
> You probably want to make your spacing narrower in order to improve
> the exchange rate; otherwise, replicas move too slowly in state space.
>
> The optimal rate depends on a few factors, but generally you want to
> keep the diagonal around 0.60 to 0.70. not 0.9.  One minus the second
> smallest eigenvalue of the transition matrix expressed the rate at
> which the system moves in state space (see
> http://scitation.aip.org/content/aip/journal/jcp/135/19/10.1063/1.3660669
> ).
> --
> Gromacs Users mailing list
>
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> send a mail to gmx-users-request at gromacs.org.
>


More information about the gromacs.org_gmx-users mailing list