- About some performance issues that occur when porting LES/DNS codes from vector machines to parallel platforms

In choosing the method for parallelization, a choice had to be made between the data parallel method (using High Performance Fortran, or native products such as CRAFT on Cray machines) on the one hand, or explicit message passing protocols such as MPI (Message Passing Interface and Parallel Virtual Machine (PVM) on the other hand. The data parallel method asks for less programming effort and takes many problems (such as dead-lock), which are typical for a parallel machine, out of the hands of the programmer. Thus, this method is the preferred one, since the step from a vector program to a parallel program is smaller. This chapter addresses the porting of flow simulation; in particular, a large-eddy simulation (LES) and direct numerical simulation (DNS) and codes from a vector machine to several parallel platforms. Specific issues which occur when porting these codes from vector architectures to parallel architectures are addressed. Because parallel machines are most probably architectures capable of handling the very large flow problems of the future, these issues will play an important role in the years to come. Instead of just showing how well a program scales, a comparison is made on the basis of wall-clock times between parallel models on parallel machines and an equivalent vector model on a vector machine.