Performance Analysis of Viscous Flow Computations on Various Parallel Architectures

A study on the performance of parallel viscous flow computations on various parallel architectures is presented. An implicit computational fluid dynamics code was utilized to solve the axisymmetric compressible Navier-Stokes equations, and parallelization of the code was obtained by the grid partitioning technique. Several hardware and software issues, such as, different communication possibilities and their effect on the parallel performance, the numerical and parallel efficiency for single grid and mesh-sequencing solutions, double and single precision calculations, and the effect of the number of processors on the computing time and the parallel performance, were investigated. Calculations for laminar and turbulent benchmark cases were performed on four parallel platforms and the results were compared with corresponding computations done on the Cray Y-MP utilizing only one processor. A theoretical model for the prediction of the parallel efficiency was also derived and it was verified for single grid and mesh-sequencing solutions by comparing the measured efficiencies with those of the predicted ones.