LAMMPS doesn't run in parallel

[m-barbosa]

Hello @xiaochendu,

I was trying to run the code on multiple processors but couldn't find a way to make it work. I tried exporting the environment variables in my .bashrc:

export LAMMPS_COMMAND="/mypath/lmp_mpi"
export ASE_LAMMPSRUN_COMMAND="mpirun --np 32 $LAMMPS_COMMAND"

and

os.environ["LAMMPS_COMMAND"]="/mypath/lmp_mpi"
os.environ["ASE_LAMMPSRUN_COMMAND"]="mpirun -np 4 /mypath/lmp_mpi"

just like on the ASE page, but it doesn't seem to work. Is there a way to run LAMMPS in parallel using surface-sampling?

[xiaochendu]

I haven't tried running the code using mpirun. I believe running on multiple processors requires 1) a lammps binary that supports mpirun and 2) a potential that is compatible as well. I think it is only faster if the simulation cell is very big (thousands or millions of atoms). For my use case, I haven't done more than 300 or so atoms.

EDIT: To clarify, using mpirun only parallelizes the lammps energy calculation, but not the sampling.

[m-barbosa]

I have a LAMMPS binary compiled with MPI support. Some potentials are very slow, like ReaxFF and other reactive potentials. For that reason, when running a CG algorithm with a couple of thousand steps (for thousands of sweeps), parallelization comes in handy.

The problem is, I don't know if I'm doing something wrong, but the environment variables don't seem to be working. As I mentioned before, the ASE documentation specifies that I can use something like this:

export ASE_LAMMPSRUN_COMMAND="/path/to/mpirun --np 4 lmp_binary"

However, for some reason, this doesn't work for parallelizing both energy and optimization calculations. I believe the only LAMMPS binary that the software recognizes is the one that comes with conda install -c conda-forge lammps, but I'm not quite sure about it.

[xiaochendu]

I see what you mean now. ASE_LAMMPSRUN_COMMAND is only used for the copper example in example.ipynb, which uses the ASE built-in lammpsrun module. For the GaN and Si examples, I created a separate LAMMPSSurfCalc that runs directly using the python lammps module. (Apologies if the imports make it confusing.) You might ask why I chose to do that. Because I couldn't find a way using the ASE module to compute relaxations in one LAMMPS call. By interfacing with the lammps module, I can do that in one step, which I believe you are doing.

To run using MPI, I think you can modify LAMMMPSCalc, perhaps around this line:

surface-sampling/mcmc/calculators.py

Line 340 in 2b5f8a1

lmp = lammps(cmdargs=["-log", "none", "-screen", "none", "-nocite"])

using the information over here:
https://docs.lammps.org/Python_launch.html#running-lammps-and-python-in-parallel-with-mpi

[m-barbosa]

I have followed the instructions on the link you sent, adding something like comm = MPI.COMM_WORLD next to the lmp line, but I've started facing other problems, sometimes with the MPI and other times after few sweeps with the LAMMPS (usually 3-6), where I faced the message Exception: ERROR: Energy was not tallied on needed timestep (src/compute_pe.cpp:88). One thing I've noticed is that the code was much faster with MPI, but I couldn't solve the problem with the compute, which is inexistent in a serial run.

I alse tested an unofficial LAMMPS implementation called pylammpsmpi, which allows running LAMMPS in parallel from a Python serial code. It worked, but was slower than a serial run.

[xiaochendu]

Hi @m-barbosa, apologies, I was busy last week. I don't have much experience parallelizing LAMMPS. But if you're willing to send over the modified code, scripts, potential, and a test system (doesn't have to be exactly what you're doing), I can try out some runs myself.