As I know, many research groups in China including us are using
LAMMPS code. But there is a shortcoming when we using LAMMPS : most
of us are using commodity SMP clusters which connected by Fast
Ethernet, whose bandwidth and latency are not as good as Myrinet and
InfiniBand. I wonder that weather I can do some thing to speed up
LAMMPS on SMP cluster. So I write this mail to ask your suggestion.
If you have MPI installed correctly on your SMP cluster, LAMMPS should work
OK. If you want to write a hybrid MPI/SMP version of LAMMPS, start
with the pair potentials. But I've never seen this out-perform MPI,
so good luck.