please *always* reply to the mailing list.
Hello Axel,
Thanks for the advices. This input is just for obtaining a better molecular
configuration, i will further simulate a box containing more molecules. In
that still doesn't make it a good idea to use GPU acceleration with
multiple GPUs. GPUs are only faster than CPUs, when you have a lot of
work units.
now, if you have a large (and thus mostly empty) volume and a few
atoms in them, then i can see how you see a GPU acceleration, but
mostly because you basically are computing a lot of stuff that is
essentially zero or close to zero. since you seem to be far from
equilibrium, it may be even faster to turn off kspace completely until
you get close to equilibrium. after all, we are talking about a state
function here, so it doesn't matter how you get there as long as you
do get there.
fact this was the reason why i'm kind of distressed with this problem,
because in the way we compiled things here the greatest advantage of running
lammps with gpu is the kspace acceleration. But I think the problem can
well, what kind of performance do you get on the CPU on its own? also,
there are quite a few things that can be tweaked to make "unusual"
systems run more efficiently.
reside in the fact, which you have pointed out, that the system is just to
small, because we were able to done some simulations with the pppm/gpu
before. I'll report back if you have any interest in the information.
while my personal interest is limited, it is considered good practice
to report the solution of a problem to the mailing list, so that other
people that run into similar issues, can later dig out the discussion
from the mailing list archives and get help without having to ask the
same question (well, in theory. in practice many still ask, but this
way we can just tell them to search the archives).
axel.