I use pppm a lot and I deal with CG models involving charged lennard jones
particles. I observe that the speed of a simulation is dependent on the
cutoff that you specify in your pair potential.(i.e. the cutoff you
specify in pair_style lj/cut/coul/long). The rule of thumb is that the
denser your system is (assuming density of your box is uniform) the
shorter is the cutoff. I suggest that you benchmark your system with
different cutoffs using a single processor. Once you find the optimum
cutoff you can proceed benchmarking parallel runs.
Jan-Michael Carrillo
PS. I use gigabit Ethernet and when I switched to openmpi from mpich
(Which Axel adviced and I thanked him for that) I got better performance
I use pppm a lot and I deal with CG models involving charged lennard jones
particles. I observe that the speed of a simulation is dependent on the
cutoff that you specify in your pair potential.(i.e. the cutoff you
specify in pair_style lj/cut/coul/long). The rule of thumb is that the
denser your system is (assuming density of your box is uniform) the
shorter is the cutoff. I suggest that you benchmark your system with
different cutoffs using a single processor. Once you find the optimum
cutoff you can proceed benchmarking parallel runs.
it is probably advisable to issue a word of warning in this context.
with lj/cut/coul/long you still have a plain cutoff scheme in the
lennard-jones part, so you cannot choose the _global_ cutoff arbitrarily
short, as that would introduce an increasingly large error in the
lennard jones part. so that cutoff has to be kept constant, but
you can "play" a bit with the coulomb cutoff (LAMMPS allows to set
them individually) to balance the amount of work that you spend
on the real space and the kspace part of the coulomb interactions.
but also there you have to consider that moving the calculation towards
more kspace will introduce some bias to the system to behave more
"crystal-like" by enhancing dipolar fluctuations that will favor a
collinear arrangement of dipoles. this is quite subtle and usually not
so much of a problem, but one should be aware of it and check if a
new "feature" is not an artefact of the treatment of electrostatics.
on the subject of CG with charges, we locally made the experience
that the estimator formula for the PPPM gridsize works best
for homogeneous charge distributions. some of our systems have
only few (oppositely) charged atoms but in relatively close
vicinity and for these systems the grids usually have to
be increased. with "normal" all-atom force field setups,
that have charged sites throughout, there was never a problem.