Efficiency of Monte Carlo collisional dynamics on GPUS

Summary form only given. Modeling collisional plasma systems with particle-in-cell (PIC) is a well-accepted method to model the statistical nature of the plasma. More accurate simulations with more particles and grid locations can easily become computationally prohibitive. Methods to implement the Monte Carlo portion of PIC with the Null Collision Method [1] reduces the number of collisional probabilities required to calculate particle dynamics. Graphical Processing Units (GPUs) have demonstrated performance speedup of over 140x on FDTD [2]. A two-dimensional GPU model for PIC has shown a speedup of nearly 30× [3] over serial CPU. Since PIC codes have two fundamental data-structures: mesh (FDTD cell) and particles, each requires a different method to parallelize. Timing [3] for particle management routines indicates MCC gained the least speed-up at 18× compared to 120× for the particle push, 300× for particle boundary conditions, and 78× for deposition of particle charge to the grid. With MCC still taking up 17%-34% of the processing time, improvements can be significant. The MCC method of [4] has shown that random sparse access to particles is the worst-case scenario for memory-bandwidth on GPUs. An alternative MCC method has optimal memory-bandwidth and is applied to this simulation, comparing timings for all particle algorithms.