Running OpenMC in CAPI with arguments

User Support
1

How can I pass run arguments like ‘cwd’ and ‘threads’ when running OpenMC in the C API through openmc.lib?
I tried the context manager openmc.lib.run_in_memory but I couldn’t figure out the syntax of sending arguments correctly.

For example, I want to run openmc in memory batch-by-batch in a different directory using cwd=‘myfolder/’. Maybe something like:

with openmc.lib.run_in_memory(kwargs={'cwd':'myfolder/'}):

Thanks!

2

Hi @mkreher13. The arguments that are passed to init via run_in_memory are the same as those that one could specify on the command line. To see what’s possible, run openmc --help from a command line, which gives you:

Usage: openmc [options] [directory]

Options:
  -c, --volume           Run in stochastic volume calculation mode
  -g, --geometry-debug   Run with geometry debugging on
  -n, --particles        Number of particles per generation
  -p, --plot             Run in plotting mode
  -r, --restart          Restart a previous run from a state point
                         or a particle restart file
  -s, --threads          Number of OpenMP threads
  -t, --track            Write tracks for all particles
  -e, --event            Run using event-based parallelism
  -v, --version          Show version information
  -h, --help             Show this message

If you want to change the number of threads and the working directory, this could be accomplished then with:

with openmc.lib.run_in_memory(args=['--threads', '16', 'myfolder']):
    ...
3

Follow-up question: how about MPI? I tried a few different things:

openmc.lib.init(args=[‘mpiexec’,’-n’,‘4’])

and

run_kwargs = {‘mpi_args’: [‘mpiexec’,’-n’,‘4’]}
openmc.lib.init(**run_kwargs)

But the first gave an MPI error, and the second gave an unexpected argument error for ‘mpi_args’. I’m further confused by the option to send an MPI intracommunicator like mpi4py. How does this come into play?

I am able to use MPI with regular OpenMC runs without the CAPI, so I know my parallel installation is configured correctly. I’d like to tag Shikhar Kumar but I’m not sure he is on this discourse page.

Thanks,
Miriam

4

Sorry, I know this is a little confusing. Because openmc.run just calls the openmc executable under the hood, it’s very easy to instead call mpiexec ... openmc which is why openmc.run has an mpi_args argument. For openmc.lib.init(), we’re not calling a subprocess so there’s no way to instantiate new MPI processes (at least in a way that MPI can handle cleanly). Instead, you need to call your Python script with mpiexec. When doing so, you also should use mpi4py to get the communicator to pass to openmc.lib.init. So, you could have a script called “run_openmc.py” with:

import openmc.lib
from mpi4py import MPI

openmc.lib.init(intracomm=MPI.COMM_WORLD)
openmc.lib.run()
openmc.lib.finalize()

and then call it with mpiexec -n N python run_openmc.py.

5

Ok thanks! That works with the downside that the entire python code must run in parallel instead of just the Monte Carlo run.