thank your for giving me the advice.
and sorry for bothering you a lot.
Actually, I have my own input file which include "fix npt". it has bigger
number of atoms.(around 1 million atoms)
and I had same kind of error with this.
then, I used Lammps tutorial-input file to figure out whether program's
bug or not. this is what I showed you before.
yes, but using GPUs has additional constraints and restrictions as well
as other performance characteristics. a lot of calculations will work with
GPUs but not all of them will be very efficient. there are a number of ways
how performance is affected and how calculations can be done. this is all
explained in the LAMMPS documentation and the papers referred from it that
describe the implementation.
with the big number of atoms, does it take longer time if I use GPU or
OpenCL package insteed of using CUDA package?
and, do they work without the big problem when I use "fix npt"?
i don't own a crystal ball that can tell me things i don't know. the major
restriction for your system is likely going to be the construction of the
neighbor lists. and whether they will fit into the RAM on your GPU. which
of the two GPU support packages is faster depends on many details and is
easiest found out by a few suitable benchmarks. the USER-CUDA package
already told you that your force it to move data on and off the GPU during
the MD. that will incur an overhead beyond the normal cost of using
USER-CUDA. the GPU package also has overhead, but since it always moves
data on and off the GPU (but less and differently) it is less affected by
not having all fixes or computes ported.
current my calculation environment is
・OS: ubuntu 12.04LTS
・LAMMPS version: lammps-1Feb14
・GPU: Tesla C2075
・Cuda6.0
whenever things don't work, always also test with the latest version.
I rebuild Lammps with GPU package.
It worked without obvious problems, as using both tutorial input file & my
own one.
calculation speed also keeps high.