Optimization of Monte-Carlo NBI NUBEAM code for GPU: success and challenges

NUBEAM is a comprehensive software package designed to model fast ion heating in tokamaks. Widely trusted and utilized in major tokamaks worldwide, NUBEAM serves as a reliable tool for plasma beam and heating modeling. This code efficiently calculates various parameters, including heating profiles, torque, current driven profiles, particle sources, and neutron emission resulting from neutral beam injection and fusion products. Its validation is robust, encompassing comprehensive physics that accounts for phenomena like Alfvénic and radio-frequency wave-particle interactions.

Driven by the need for rapid analysis and experimental planning in control rooms, this work draws inspiration from the present-day requirements of quick turnaround between shots. The Monte Carlo beam and fusion product code, NUBEAM, has undergone modifications to leverage the power of GPU architecture, thereby accelerating the simulation of fast ion population deposition, orbiting, and slowing down in tokamak plasmas. The newly developed code has been rigorously tested on two supercomputers, Traverse at PPPL and Perlmutter at NERSC, which utilize NVIDIA V100 SXM2 GPUs and NVIDIA A100 GPUs, respectively.

The results obtained demonstrate a notable speed improvement of approximately 4.5 times on Traverse, employing four GPUs and 32 MPI processes for the NUBEAM simulation with 100,000 Monte Carlo particles. Furthermore, Perlmutter showcases even better performance. Throughout this process, various challenges have been encountered and overcome, resulting in a compelling success story. These challenges and accomplishments will be highlighted and discussed in further detail.