Optimizing Performance of Three-Dimensional Astrophysical Plasma Simulations

Abstract

Particle-in-cell simulations of collisionless shocks have been shown to produce self-consistent shock acceleration of non-thermal particles, which are an important area of astrophysical research. A challenge that remains, however, is acquiring results from these simulations without a prohibitively long runtime. In this thesis we introduce the TRISTAN-MP simulation and examine methods of optimizing its performance. These methods include replacing multiple arrays of single precision values with one array of memory-aligned structures, precomputing coefficients that may be computed multiple times, and minimizing slow double precision calculations. We then discuss the local implementations of each optimization and present the resulting change in performance for each implementation. Finally, we discuss the optimizations that were selected for use in the full simulation and the full implementation of those optimizations. We show that the output of the optimized simulation maintains the properties of the original output and we find that running time is significantly reduced.