Parallel finite volume method-based fluid flow computations using OpenMP and CUDA applying different schemes

Parallelization of inhouse develpoed code for numerical computations on heterogeneous computing machine is becoming common. As the numerical solvers and problem complexity are evolving, the parallel computing facilities are also growing. This research study deals with the use of four different schemes to parallelly compute the numerical flow equations based on the finite volume method. The different schemes proposed are applied for parallelization using central processing units and graphical processing units. Open multiprocessing (OpenMP) and compute unified device architecture (CUDA) are the parallel computing tools used for parallelization of the code applying four schemes, viz. red and black successive over-relaxation (RBSOR), wavefront, combined RBSOR and wavefront, and alternate RBSOR and wavefront scheme. The flow analysis is carried out for internal and external flow at different Reynolds numbers on dissimilar machines having their individual computational capability. Speedup obtained and parallel efficiency achieved using the proposed unusual parallelization method are investigated separately. The grid size for both the flow conditions is fixed during the parallel computation performance analysis. The RBSOR scheme provided the maximum speedup in all cases of flow, scheme, and tool used. The wavefront scheme provides the lowest speedup and parallel efficiency. The alternate scheme is better than the wavefront scheme and combined scheme using OpenMP. The speedup achieved and parallel efficiency obtained for the CUDA parallelized code are in the range of 200 x and 70%, respectively, applying the RBSOR scheme.