mcdc 0.9.1

Monte Carlo / Dynamic Code, a pure python high performance Monte Carlo neutronics package

MC/DC: Monte Carlo Dynamic Code

MC/DC is a performant, scalable, and machine-portable Python-based Monte Carlo neutron transport software currently developed in the Center for Exascale Monte Carlo Neutron Transport (CEMeNT).

Our documentation on installation, contribution, and a brief user guide is on Read the Docs.

Installation

We recommend using conda or some other environment manager to manage the MC/DC installation. This avoids the need for admin access when installing MC/DC's dependencies and allows greater configurability for developers. For most users working on a single machine of which they are administrators, MC/DC can be installed via pip:

pip install mcdc

For developers or users on HPC machines, we recommend that you not use the pip distribution and instead install MC/DC and its dependencies via the included install script, which builds mpi4py from source and uses conda to manage your environment. This is the most reliable way to install and configure MC/DC. It also takes care of the Numba patch and can configure the continuous energy data library, if you have access.

Common issues with mpi4py

The pip mpi4py distribution commonly has errors when building due to incompatible local MPI dependencies it builds off of. While pip does have some remedy for this, we recommend the following:

• Mac users: we recommend openmpi is installed via homebrew (note that more reliable mpi4py distribution can also be found on homebrew), alternatively you can use conda if you don't have admin privileges;
• Linux users: we recommend openmpi is installed via a root package manager if possible (e.g. sudo apt install openmpi) or a conda distribution (e.g. conda install openmpi)
• HPC users and developers on any system: On HPC systems in particular, mpi4py must be built using the system's existing mpi installation. Installing MC/DC using the install script we've included will handle that for you by installing dependencies using conda rather than pip. It also takes care of the Numba patch and can configure the continuous energy data library, if you have access.

Numba Config

Running MC/DC performantly in Numba mode requires a patch to a single Numba file. If you installed MC/DC with the install script, this patch has already been taken care of. If you installed via pip, we have a patch script will make the necessary changes for you:

1. Download the patch.sh file here (If you've cloned MC/DC's GitHub repository, you already have this file in your MCDC/ directory).
2. In your active conda environment, run bash patch_numba.sh. If you manage your environment with conda, you will not need admin privileges.

visualizer Config

MC/DC has a visualizer built from the netgen package. This is not included in the base install of MC/DC due to the size of the dependencies (~300MB). To install these dependencies use

pip install mcdc['viz']

Running

MC/DC can be executed in different modes: via pure python or via a jit compiled version (Numba mode). Both modes have their use cases; in general, running in Numba mode is faster but more restrictive than via pure python.

Pure Python

To run a hypothetical input deck (for example this slab wall problem) in pure python mode run:

python input.py

Simulation output files are saved to the directory that contains input.py.

Numba mode

MC/DC supports transport kernel acceleration via Numba's Just-in-Time compilation (currently only the CPU implementation). The overhead time for compilation when running in Numba mode is about 15 to 80 seconds, depending on the physics and features simulated. Once compiled, the simulation runs MUCH faster than in Python mode.

To run in Numba mode:

python input.py --mode=numba

Running in parallel

MC/DC supports parallel simulation via MPI4Py. As an example, to run on 36 processes in Numba mode with SLURM:

srun -n 36 python input.py --mode=numba

For systems that do not use SLURM (i.e., a local system) try mpiexec or mpirun in its stead.

Contributions

We welcome any contributions to this code base. Please keep in mind that we do take our code of conduct seriously. Our development structure is fork-based: a developer makes a personal fork of this repo, commits contributions to their personal fork, then opens a pull request when they're ready to merge their changes into the main code base. Their contributions will then be reviewed by the primary developers. For more information on how to do this, see our contribution guide.

Bugs and Issues

Our documentation is in the early stages of development, so thank you for bearing with us while we bring it up to snuff. If you find a novel bug or anything else you feel we should be aware of, feel free to open an issue.

Testing

MC/DC uses continuous integration (CI) to run its unit and regression test suite. MC/DC also includes verification and performance tests, which are built and run nightly on internal systems. You can find specifics on how to run these tests locally here.

Cite

To provide proper attribution to MC/DC, please cite

@inproceedings{var_mc23_mcdc,
    Booktitle = {International Conference on Mathematics and Computational Methods Applied to Nuclear Science and Engineering},
    title = {Development of {MC/DC}: a performant, scalable, and portable Python-based {M}onte {C}arlo neutron transport code},
    year = {2023},
    author = {Ilham Variansyah and Joanna Piper Morgan and Kyle E. Niemeyer and Ryan G. McClarren},
    address = {Niagara Falls, Ontario, Canada},
    doi={10.48550/arXiv.2305.07636},
}

which should render something like this

Variansyah, Ilham, J. P. Morgan, K. E. Niemeyer, and R. G. McClarren. 2023. “Development of MC/DC: a performant, scalable, and portable Python-based Monte Carlo neutron transport code.” In International Conference on Mathematics and Computational Methods Applied to Nuclear Science and Engineering, Niagara Falls, Ontario, Canada. DOI. 10.48550/arXiv.2305.07636

License

MC/DC is licensed under a BSD-3 clause license. We believe in open source software.