The OpenMX installations are generally maintained by Peter Larsson (firstname.lastname@example.org).
Typically, you want to run OpenMX in hybrid MPI/OpenMP mode. This requires some care when writing the batch script. For example, to run on 256 cores using 128 MPI ranks and 2 OpenMP threads per rank, we request 16 nodes (16x16=256 cores) and ask for two cores to be assigned to each task (=rank):
#!/bin/bash #SBATCH -J jobname #SBATCH -A SNIC-xxx-yyy #SBATCH -N 16 #SBATCH --cpus-per-task=2 export OMP_NUM_THREADS=2 mpprun /software/apps/openmx/3.7.6/default/openmx input.dat -nt 2
The “pseudopotentials” folder, called
DFT_DATA13 is installed here:
You need to specify the path in the input file, like this:
The DFT_DATA folder may not be regularly updated, or checked for correctness, so I advise users to download and maintain there own pseudopotential library.
Parallel benchmarking is ongoing; so far I can only say that MPI/OpenMP hybrid mode seems faster than pure MPI running on the same number of cores. I would expect that you need to run at least 2 MPI ranks per node (one per NUMA zone) to get good performance, though.