Massively Parallel Computing of Turbulent Transport in Tokamaks

Abstract

With the advent of the gyrokinetic formalism, recent developments in low-noise nonlinear $\delta f$ methods, and enormous gains in computing power, large-scale gyrokinetic simulations have become an important tool for improved understanding of anomalous transport in tokamaks. Simulating the non-linear behaviour requires solving for the perturbations of distrbution function in five dimensions. We use a particle-in-cell approach to solve the equations via the non-linear characteristic method. The code has been parallelized for a variety of architecures (C90, CM-5, T3D) using a 1-D domain decomposition along the torroidal axis, for which the number of particles in each cell remains approximately constant. The quasi uniform distribution of particles, which minimizes load imbalance, coupled with the relatively small movement of particles across cells, which minimizes communications, makes this problem ideally suited to massively parallel architectures. We present the performance of the program as a function of number of processors and problem size, which demonstrates the near perfect scalability of the code. In addition, we discuss the scientific results obtained from the code and the types of problems that will be addressable as the next generation of super-computers become available.