Simulating and benchmarking the shallow-water fluid dynamical equations on multiple graphical processing units
Daniel Peter Playne, Ken A. Hawick, Mitchell G.B. Johnson · 2014
The shallow-water model equations provide a simple yet realistic benchmark problem in computational fluid dynamics (CFD) that can be implemented on a variety of computational platforms. Graphical Processing Units can be used to accelerate such problems either singly using a data parallel decompositional scheme or with multiple devices using a domain decompositional approach. We implement the SW equations on a range of modern GPUs with both parallel schemes and report on the typical performance. We compare integer optimised GPUs and very modern floating-point intensive GPU devices such as NVidia's Kepler K20X, and also investigate different m-GPU communication methods for geometric decompositions. We give detailed performance results and a summary of the main parallelisation issues.