[gmx-users] Myrinet vs SCALI

[Erik Lindahl]

Bert de Groot wrote:

> Erik Lindahl wrote:
>
> > in the meantime
> > you can increase the grid spacing and compensate by using a higher interpolation order. The
> > grid memory size is proportional to the third power of the spacing, so this will improve
> > scaling considerably.
> >
>
> thanx for the hint. Can you provide an estimate for an appropriate grid spacing/interpolation
> order that gives a good scaling AND reasonable accuracy?

Not very scientifically, but a reasonable start is to change pme_order from 4 (default) to 6
in the mdp file, and increase the spacing about 50% too, say from 0.12 to 0.18. I have NOT tried

this a lot, so these are probably not the most optimal values :-)

You can check the accuracy by comparing the forces on charged particles
(set fout to 1 and run a step) with your previous run, or even make a 'reference' run with
very small spacing and pme_order=6. (The interpolation cost increases as the third power of
pme_order).

Cheers,

Erik