Exploring parallel execution of Coreform Cubit

Recommend creating a virtual environment. Below are instructions for Windows 11, Powershell.

python3.11.exe -m venv env
.\env\Scripts\Activate.ps1
python.exe -m pip install -r requirements.txt

There are three examples:

1. Array
   • src/array_example.py
   • This example builds an NxNxN array of bricks and hex-meshes them in parallel
2. Mechanical
   • src/mechanical_example.py
   • This example builds a tetmesh on a mechanical assembly from NIST
3. Nuclear
   • src/nuclear_example.py
   • This example builds a surface mesh on the ITER example from Paramak

To execute, recommend creating a separate working directory and running the script from that location:

mkdir path\to\workdir\array
cd path\to\workdir\array
python.exe path\to\array_example.py --num-proc 8 --array-size 8

mkdir path\to\workdir\mechanical
cd path\to\workdir\mechanical
python.exe path\to\mechanical_example.py --num-proc 8

mkdir path\to\workdir\nuclear
cd path\to\workdir\nuclear
python.exe path\to\nuclear_example.py --num-proc 8

When these execute they do the following:

1. Create a base model, either via CAD generation or importing a CAD file.
   • Saves this base model as base_model.cub5 in the working directory
2. Distributes bodies in num_proc lists of approximately quantity of bodies
   • Exports a temporary file for each of the num_proc lists to ./tmp/proc_{p}.cub5 -- where {p} is an integer for the worker id.
3. Uses the multiprocessing library to map each temporary file to a worker within a Pool.
   • Each worker meshes its temporary file and saves the temporary file, overwriting it.
4. After completion, the main process creates a new Cubit file and imports each of the meshed temporary files.
   • Saves this meshed model as base_model_meshed.cub5

The results above demonstrate that it is possible to use Python to perform parallel, process-based operations using Coreform Cubit's Python API, however the speedups in these examples are minimal to moderate. It may be that a more sophisticated approach to distributing entities in order to better load-balance the workers would improve performance in the mechanical and nuclear examples. It should also be noted that there is additional overhead that may not be encountered in a more traditional serial processing usage of Coreform Cubit: exporting multiple CUB5 files, reading them in and re-exporting after meshing, gathering all the partial files into a new monolithic file, etc. This overhead cost may be alleviated by skipping the final gather in Coreform Cubit, instead exporting Exodus mesh files and using SEACAS tools (e.g., ejoin) to combine them. Overhead may also be reduced if that parallel execution is performed across unique cases (e.g., a DoE sweep) wherein each case would need to be loaded separately even in the serial case.