Parallel CPU/GPU computing for smoothed particle hydrodynamics models

Abstract

Smoothed Particle Hydrodynamics (SPH) is a numerical method particularly suitable to describe a variety of complex free-surface flows with large discontinuities. However, SPH simulations are computationally expensive and typical runtimes are too high to study real problems with high resolution. The proposed solution is the parallel computation to accelerate the SPH executions. This work introduces several high performance techniques applied to SPH to allow simulation of real problems at reasonable time. In this way, OpenMP was used to exploit all cores in the classical CPUs. On the other hand, CUDA language was used to take advantage of the high parallel computing power of GPUs (Graphics Processing Units). Finally, Message Passing Interface (MPI) was implemented to combine the power of several machines connected by network. These parallelization techniques are implemented in the code DualSPHysics and results are shown in terms of performance, efficiency and scalability using different CPU and GPU models.