Modeling the parallel execution profile of a CFD simulation on a cluster of workstations

This paper describes the derivation of a model which elucidates the asymptotic parallel execution profile of a real CFD application parallelized over a cluster of workstations connected via the Ethernet. The CFD simulation studies the formation of trailing vortices around a wing with no twist, no taper, no sweep and a rounded wing-tip during subsonic flight conditions. The simulation solves the conservation Navier-Stokes equations over a computational grid surrounding the wing of dimensions 61×71×41; 177,571 grid points. Modeling the execution of the parallelized version of such an application, before embarking on the actual parallel implementation, is important for two reasons. Firstly, it enables a better understanding of the intrinsic parallel structure and communication patterns of the chosen computational solution procedures. Secondly, it allows the scalability of the solution procedures, parallelized over a chosen parallel computer, to be analysed. We encourage parallel application developers to perform such modeling because it is our experience that better quality first time correct parallel programs can be derived more quickly when this is done