[gmx-users] Trying to use cutoff electrostatics with MARTINI

[Szilárd Páll]

On Tue, Apr 11, 2017 at 1:03 PM, Mark Abraham <mark.j.abraham at gmail.com>
wrote:

> Hi,
>
> On Tue, Apr 11, 2017 at 12:09 PM Graham J.A. <J.A.Graham at soton.ac.uk>
> wrote:
>
> > Hi,
> >
> > I'm preparing to do some benchmarking and wanted to look at the
> > performance differences between reaction-field and cutoffs for the
> > electrostatic model with the MARTINI forcefield.
> >
>
> You'd expect that to be small - one can implement plain cutoff as RF with a
> zero parameter.
>
>
> > I've downloaded the recommended MDPs from the MARTINI website, but when
> > I specify cutoff in the MDP I still get "NxN RF Elec. + LJ" in the
> > performance tables at the end of the log.  The section at the top of
> > the log shows that GROMACS has recognised that I'm asking for cutoffs.
> >
> > This is happening with both GROMACS 2016.3 and 5.1.4.  Am I
> > misunderstanding the log output, or is something going wrong?
> >
>
> Particularly on hardware that supports fused multiply-addition (most of the
> current relevant ones), the extra operation is ~free, and the only overhead
> is the load of the constant parameter, which happens outside the inner
> loop, and perhaps some register pressure. So we don't bother with a special
> kernel for this case, and your observations are expected.
>

That's indeed true for all CPU SIMD kernels, but the GPU kernels (which
seem to be used here) do have a plain cut-off flavor which is slightly
faster than RF. The difference is not only in the single extra Flop you
mention, but also in the exclusion force treatment. Hence, the flop table
in the log is not accurate in this case.

Cheers,
--
Szilárd


> Mark
> --
> 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.
>