How to Maintain Efficiency on Vector and Smp Platforms for Large Aerodynamic Calculations

Large industrial aerodynamic calculations are nowadays performed indifferently on parallel vector computers or on clusters of SMPs. From a software design point of view it is crucial to ensure the best possible sustained megaflops rate on both platforms while for sake of minimal labour effort it is important to maintain only one source code. In this paper we describe software techniques that have been implemented to comply with these constraints. We first show how we succeeded to obtain a sustained megaflops rate that is independent of the mesh size. We then investigate the possibilities of using OpenMP in some of the main time consuming routines to replace the vectorization by fine grain parallelism to better exploit the shared memory available within each node of the SMP computers. Numerical experiments are reported on HP-Compaq, IBM SP based on Power 3 and Power 4 where tremendous savings can be obtained at a cost of a very little code change.