Hybrid GPU Accelerated Mesoscopic Particle Simulation

During the past few years, the performance enhancement of existing algorithm implementations through massive parallelization on modern graphics processing units (GPUs) based on NVIDIA's Compute Unified Device Architecture (CUDA) has become very popular in a large variety of applications. Particle methods e.g. are dedicated for parallelization because of their basic algorithmic structure. The neighbourhood search, the computation of interaction forces and the final state update via time integration can be parallelized directly among the particles.

In order to evaluate the potential of such a GPU implementation compared to the single CPU execution, a parallel and hybrid implementation with OpenMP and CUDA has been developed for a special particle method, the so-called mesoscopic particle method. This method takes into account mechanical as well as thermodynamical degrees of freedom. On the basis of a hybrid implementation, the performances on different types of execution architectures (sequential on single CPU core, parallel on multi-core CPU, parallel on GPU) can be determined and compared. (© 2012 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim)