Accelerating the D3Q19 Lattice Boltzmann Model with OpenACC and MPI

Multi-GPU implementations of the Lattice Boltzmann method are of practical interest as they allow the study of turbulent flows on large-scale simulations at high Reynolds numbers. Although programming GPUs, and in general power-efficient accelerators, typically guarantees high performances, the lack of portability in their low-level programming models implies significant efforts for maintainability and porting of applications. Directive-based models such as OpenACC look promising in tackling these aspects. In this work we will evaluate the performances of a Multi-GPU implementation of the Lattice Boltzmann method accelerated with OpenACC. The implementation will allow for multi-node simulations of fluid flows in complex geometries, also supporting heterogeneous clusters for which the load balancing problem is investigated