Run benchmark with OpenMP only

LAMMPS LAMMPS Beginners
1

Dear all,
I’m interested in running benchmarks of Lammps on our heterogeneous memory processor (DRAM +NVM ). The setup I would like to use is only with OpenMP (no MPI, no GPU) with gnu.
I installed and built everything with cmake (cmake …, make -j ) and ran a benchmark with lj.in input.
However when monitoring cpu usage I can see that only 1 CPU is used during execution.
I did 2 tests:

  1. building lmp without any MPI libs and run with OMP_NUM_THREADS=16 + ./lmp -in in.lj
  2. building lmp with MPI (OpenMPI) libs and run with OMP_NUM_THREADS=16 + mpirun -np 1 lmp -in.lj

Log files show that nb of omp threads was taken into acount but this is not what I observe with htop. (100% on one CPU)
Also the in.lj input is only one I can use. When running the others (chain, chute, rhodo and eam) I get:
ERROR: Unrecognized atom style ‘bond’ is part of the MOLECULE package which is not enabled in this LAMMPS binary (chain example)

Did I miss something in terms of cmake options?
Is it possible to only use OpenMP?
I was going through the doc and this forum but could not find an answer to my question.
Thanks in advance,
Pierre

2

If you only want OpenMP parallelization, you could just download and use the precompiled packages attached to the GitHub release pages
E.g. for Linux x86_64: https://github.com/lammps/lammps/releases/download/patch_7Feb2024/lammps-linux-x86_64-7Feb2024.tar.gz

The other packages include the LAMMPS GUI, but also the console executable lmp.

That would be the expected behavior.

Have you read this part of the documentation?
https://docs.lammps.org/Speed_omp.html

Bottom line, multi-thread support requires modified compute kernels and those are in the OPENMP package and those require different style names, or more conveniently use a suffix flag setting.

3

That would be the expected behavior.

I meant 1 cpu among 16. (sorry for the typo)

In the end, one needs to cmake with -D PKG_OPENMP=yes, export OMP_NUM_THREADS=16, run with -sf omp arguments to effectively use the 16 threads.