So...
New half-cell NL code is a 10% loss at 8A cutoff, break-even at 10A, and
seemingly ~2x faster at 12A. What does this mean? It means I can't get rid
of the old code, but I need to switch between new and old based on nonbond
cutoff plus skin. It's vastly more efficient memory-use wise, but at lower
cutoffs, the cell occupancy is ~11 atoms and that cuts performance in half,
landing at a ~10% slowdown. Getting really tired of Brand X benchmarks
playing games with skipping the Ewald Sum, weird timesteps, and higher
cutoffs. The original code was tuned to 8A because that's what mattered at
the time and now in the future, it's time to future proof.
I have some ideas beyond this, but I'm calling this a win and moving
towards new multi-GPU stuff on the way to adding back SoftCore TI inside
the much 30+% more efficient code.
Scott
On Sun, Jun 27, 2021 at 5:25 AM David A Case <david.case.rutgers.edu> wrote:
> On Fri, Jun 25, 2021, Scott Le Grand wrote:
>
> >What's the biggest nonbond cutoff plus skin in practical use currently?
>
> I generally use 10 + 2 = 12 for at least some calculations. Sounds from
> Stephan's comments like 10 + 3 = 13 might also be useful.
>
> If one is really being radical, we could think about splitting the LJ terms
> in a way that is similar to what is being done for electrostatics. That
> has a lot of support in the community for giving more physical realism,
> especially for anisotropic systems, like bilayers + water. That should
> probably allow an 8 + 2 = 10 limit.
>
> I hope others will chime in here. I've never thought about pushing the
> limits on non-bonded cutoffs.
>
> ....dac
>
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Tue Jun 29 2021 - 13:00:02 PDT
