A PCISPH implementation using distributed multi-GPU acceleration for simulating industrial engineering applications

Predictive–corrective incompressible smoothed particle hydrodynamics (PCISPH) is a promising variant of the particle-based fluid modeling technique smoothed particle hydrodynamics (SPH). In PCISPH, a dedication prediction–correction scheme is employed which allows for using a larger time step and thereby outperforms other SPH variants by up to one order of magnitude. However, certain characteristics of the PCISPH lead to severe synchronization problems that, thus far, prevented PCISPH from being applied to industrial scenarios where high performance computing techniques need to leveraged in order to simulate in appropriate resolution. In this work, we are for the first time, presenting a highly accelerated PCISPH implementation which employs a distributed multi-GPU architecture. To that end, dedicated optimization techniques are presented that allow to overcome the drawbacks caused by the algorithmic characteristics of PCISPH. Experimental evaluations on a standard dam break test case and an industrial water splash scenario confirm that PCISPH can be efficiently employed to model real-world scenarios involving a large number of particles.