Implementing particle and ray-based algorithms in heterogeneous environments for plasma simulations

The proliferation of CPU-GPU heterogeneous comput-

ing presents an opportunity for the acceleration of high-fidelity plasma

physics simulations. The TriForce 1 code’s ray-tracing package is be-

ing developed to leverage this architecture and uses a parallel imple-

mentation of a recently developed fixed-point iteration algorithm for

including the effect of crossed-beam energy transfer (CBET) 2 . A hy-

brid object-oriented and grid-based dereferencing scheme allows for si-

multaneous tracking of ray and grid parameters while preserving data

locality and allowing low-synchronization SIMD iteration kernels. Intro-

ducing OpenMP vectorization resulted in a 10× speedup on a 12-thread

processor and the addition of GPU computation increased this further.

Furthermore, implementations of particle-based algorithms are in de-

velopment for multi-node distributed memory GPU clusters to enable

large-scale plasma PIC simulations on high-performance computing sys-

tems. Previous efforts have been limited to either shared memory GPUs

or distributed memory CPU-only packages. We present progress on us-

ing both approaches in concert to create a scalable and heterogeneous

platform for particle- and ray-based scientific computing. References:

1 [A. B. Sefkow, et al., Bulletin of the APS 64, JP10.125 (2019)], 2 [R. K.

Follett, et al, Phys. Rev. E 98, 043202 (2018)].