A novel multi-GPU parallelization paradigm for SPH applied to solid mechanics in complex industrial applications

Abstract

A novel parallelization paradigm has been developed for multi-GPU architectures. Classical multi-GPU parallelization for SPH rely on domain decomposition. In our approach each particle can be assigned to a GPU independently of its position in space. This ensures a kind of natural load balancing because the number of particles per GPU remains constant. The data exchange domain is no more a surface as in the classical approach or in mesh-based method, but it is a volume which is growing as mixing occurs in time between the particles assigned to different GPUs. This growth must be prevented because the efficiency in terms of computation time and memory consumption is rapidly dropping with the mixing. A simple heuristic is suggested to periodically detect particles to swap between GPUs in order to keep the exchange volume close to a surface. The final decomposition is much alike a domain decomposition and can be applied to any geometry. We developed this approach within the framework of the industrial code IMPETUS Solver using the Gamma-SPH-ALE method. Industrial applications in the space and military fields are presented. They demonstrate the relevance of the approach developed in an industrial context and for complex applications.