Structure generation with MPI parallelization

Oct. 21 2023, update

Random structure generation using MPI has been available since version 1.1.0 ( using CrySPY >= 1.2.3 is better). You need to install mpi4py in your Python environment for MPI parallelization. Of course, an MPI library such as Open MPI, Intel MPI, and MPICH is required for your workstation.

Info

Requirements:

  • CrySPY 1.1.0 1.2.3 or later
  • mpi4py
  • MPI library (Open MPI, Intel MPI, MPICH, etc.)
Warning

1.1.0 <= CrySPY <=1.2.2 has a bug. When you use bash (zsh) to run a job with MPI (e.g., jobcmd = zsh, jobfile = job_cryspy), the MPI job does not run. There is no problem when you use a job scheduler (qsub, sbatch). It has already fixed in version 1.2.3.

The figure below shows the relationship between elapsed time and the number of processes for 1000 structures of Si8 with the following setting:

[basic]
algo = RS
calc_code = soiap
tot_struc = 1000
nstage = 1
njob = 2
jobcmd = zsh
jobfile = job_cryspy

[structure]
natot = 8
atype = Si
nat = 8
mindist_1 = 2.2

The structure generation is taking a long time because of a slightly stricter setting like mindset_1 = 2.2. The structure generation was performed 10 times for each number of processes.

fig_MPI fig_MPI

Run

mpiexec -n 4 cryspy -p
Info

See also Tutorial > Random structure generation with MPI