Fine-grained parallelization of a vlasov-poisson application on GPU

Abstract

Understanding turbulent transport in magnetised plasmas is a subject of major importance to optimise experiments in tokamak fusion reactors. Also, simulations of fusion plasma consume a great amount of CPU time on today's supercomputers. The Vlasov equation provides a useful framework to model such plasma. In this paper, we focus on the parallelization of a 2D semi-Lagrangian Vlasov solver on GPGPU. The originality of the approach lies in the needed overhaul of both numerical scheme and algorithms, in order to compute accurately and efficiently in the CUDA framework. First, we show how to deal with 32-bit floating point precision, and we look at accuracy issues. Second, we exhibit a very fine grain parallelization that fits well on a many-core architecture. A speed-up of almost 80 has been obtained by using a GPU instead of one CPU core. As far as we know, this work presents the first semi-Lagrangian Vlasov solver ported onto GPU. © 2011 Springer-Verlag Berlin Heidelberg.