Errors on LAMMPS-CPU implementation

[bapatist]

Hi all,
trying to run my model-lammps.pt model on LAMMPS, I run into the following problems:

1. Using ntasks=1 the Segmentation Fault error. (maybe this is something I don't want to use anyway)
   sub_ntasks1.txt
   out.log

2. With ntasks=72 and pair_style mace no_domain_decomposition, I get a RuntimeError (index is out of bound) error from the torschscript interpreter.
   sub.txt
   err.txt
   out.log

3. Using a plain pair_style mace with ntasks>1 runs for a bit longer before crashing with out-of-memory kill events.
   sub2.txt
   err.txt
   out.txt

And the LAMMPS input file looks like this with no_domain_decomposition included for the second case
lammps.txt

Please let me know if you see anything problematic, obvious or otherwise. Thanks!

[wcwitt]

Thanks for the detailed report! Let's stick with (1.) for now. Do you definitely need atom_style full?

[bapatist]

I don't necessarily need the full style. Retried with atomic style for the first case and I don't encounter a seg fault anymore and the simulation seems to be running. Trying out different options to improve speed now. I will write back here once I have more details. Thanks a lot!

[bapatist]

I was able to run NPT dynamics for 120 atoms (40 H2O) at 0.021 ns/day with no_domain_decomposition turned on. Attaching the log file here. Do you have suggestions to make this faster? log.txt

I was trying to use ntasks=72 with no_domain_decomposition turned off. Or is it not implemented to work like this yet? Attaching the error I get when I try to do this:

without no_domain_decomposition: err_no_nodomain.txt
with no_domain_decomposition: err_nodomain.txt

[ilyes319]

Using GPUs instead of CPUs would be the most straight forward to make this faster. Even a quite cheap GPU would provide order of magnitude speed up (as MACE is a GPU code).

[wcwitt]

@ilyes319's answer will give you the best speedup, I'm fairly certain. But I'll try to address the CPU questions anyway.

I was able to run NPT dynamics for 120 atoms (40 H2O) at 0.021 ns/day with no_domain_decomposition turned on. Attaching the log file here. Do you have suggestions to make this faster?

Make sure you are using ntask=1 with --exclusive, and then try different threading values using OMP_NUM_THREADS and MKL_NUM_THREADS. But it sounds like you were doing this already.

I was trying to use ntasks=72 with no_domain_decomposition turned off

The problem here is that it loads 72 separate models and you run out of memory quickly. You could experiment with a few tasks per node, adjusting the threading accordingly, but the GPU will be better if you can use it.

Please do let us know how things go! If I think of anything else I will tell you.

[bapatist]

Hi! Just wanted to update on this: we have GPU models running with LAMMPS and indeed these are pretty fast! Thanks a lot.