On Wed, Aug 5, 2015 at 7:34 PM, Adrian Roitberg <roitberg.ufl.edu> wrote:
> Unless I am reading this wrongly, you did
>
> 2 fs dt, nstlim=2000000 --> 4 ns/10.6 hours
> 4fs dt , nstlim=2000000 --> 8 ns/11.5 hours
>
> So, basically in the same wallclock time, you DOUBLED the simulation
> time. That is exactly what is supposed to happen, right ?
>
​Just to add in a little bit here -- you would naively expect the wallclock
time to stay the same if all you're changing is the time step. But keep in
mind that with a 4 fs timestep, particles move farther (about twice as far)
as with a 2 fs timestep. As a result, particles diffuse farther *per
timestep*, meaning that the pairlist will need to be rebuilt significantly
more often. Since pairlist builds are fairly expensive, this will account
for the extra ~10% slowdown in the apparent per-timestep efficiency.
But as Adrian pointed out, this 10% slowdown is more than countered by
doubling the timestep, such that the total ns/day is slightly less than
double the ns/day for a 2 fs timestep.
All the best,
Jason
--
Jason M. Swails
BioMaPS,
Rutgers University
Postdoctoral Researcher
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Aug 05 2015 - 17:30:02 PDT
