[gmx-users] Effect of pressure coupling frequency on gpu simulations

[Mark Abraham]

On Wed, May 22, 2013 at 6:32 AM, Trayder <trayder.thomas at monash.edu> wrote:

> Hi all,
> I've been running 5fs timestep simulations successfully without gpus
> (united-atom, HEAVYH). When continuing the same simulations on a gpu
> cluster
> utilising the verlet cutoff-scheme they crash within 20 steps. Reducing the
> timestep to 2fs runs smoothly, however I noticed the message:
>
>
>
> Making this change manually led to crashing simulations as nstcalclr,
> nsttcouple and nstpcouple default to the value of nstlist. After defining
> them all separately I was able to determine that the simulation exploding
> was dependent entirely on nstpcouple and by lowering it to 5 (from the
> default 10) I was able to run simulations at a 5fs timestep.
>
> So, my questions: Is lowering nstpcouple a legitimate solution or just a
> bandaid?
>

P-R does not cope well with situations where the box size changes enough
(e.g. you should normally avoid it during equilibration). nstpcouple != 1
means that you simulate on an NVE manifold for a period of time (maybe with
some T changes if nsttcouple != nstpcouple), and I'd suppose the longer
that interval the bigger the chance of a build-up of pressure that P-R will
then try to relieve by changing the box size. Larger nstlist and dt will
exacerbate this, of course. I would recommend you experiment and see how
far you can push things and keep statistics that still resemble those with
small nstpcouple. Larger nstpcouple helps reduce the frequency with which
global communication occurs, and that affects your simulation rate... life
is complex!

It would be nice if we were able to compute heuristics so that mdrun could
anticipate such a problem and warn you, but off-hand that seems a tricky
problem...

The simulation runs with nstcalclr and nsttcouple set to 50 along with
>

nstcalclr should have no effect - it works only with the group scheme,
which does not work on GPUs.


> nstlist. Is nstlist the only setting that should be increased when
> utilising
> gpus?
>

Yes, AFAIK. The point is that nstlist is the interval between neighbour
searches, and (at the moment at least) that's only done on the CPU. The
Verlet kernels cheerfully compute lots of zero-strength interactions
outside the cutoff (by design), and depending on the relative performance
of your hardware it is normally more efficient to bump nstlist up (and
rlist accordingly, to provide a larger buffer for diffusion of particles)
and compute more zeroes than it is to search for neighbours more often.

Mark


>
> Thanks in advance,
> -Trayder
>
> P.S. The working mdp file:
>
>
>
>
>
>
>
> --
> View this message in context:
> http://gromacs.5086.x6.nabble.com/Effect-of-pressure-coupling-frequency-on-gpu-simulations-tp5008439.html
> Sent from the GROMACS Users Forum mailing list archive at Nabble.com.
> --
> gmx-users mailing list    gmx-users at gromacs.org
> http://lists.gromacs.org/mailman/listinfo/gmx-users
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
> * Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-users-request at gromacs.org.
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>