Data-parallel algorithms for Monte Carlo simulation of neoclassical transport in magnetically confined plasmas

Two data-parallel algorithms for simulating test particle transport in magnetically confined plasmas are compared. The first approach, a direct parallelization of conventional algorithms, is found to suffer from a low processor utilization factor unless parallel time stepping and particle splitting combined with an appropriate termination criterion are used. The second approach utilizes weak methods for solving stochastic differential equations to calculate relevant moments of the test particle distribution function. This shift in focus from the individual test particles to the test particle distribution avoids the load balancing problems found for the conventional algorithm. The approach based on stochastic differential equations is found to outperform the conventional algorithm when implemented on a Connection Machine 2 computer.