Investigation of the parallel performance of the unstructured DLR-TAU-Code on distributed computing systems

High Performance Computing is widely used in the Computational Fluid Dynamics (CFD) community to satisfy the increasing needs for a faster prediction of simulated flows. Despite the advances in computer manufacturing and chip design over the last years, there is still a lack of computational resources required to solve CFD problems in real time. Moreover, for most research centers it is difficult to buy a supercomputer like the NEC-SX 6 based Earth Simulator out of their budgets. To encounter this gap, there is an increasing installation of supercomputers built out of commodity components like PC's or workstations connected via high speed networks. These cluster systems often lack tight integration between the processor, network and main memory. This often results in inefficiencies in the communication subsystem, such as high software overheads and/or message latencies. In this paper we investigate the parallel performance of the unstructured DLR TAU-Code on a set of up-to-date Linux-Clusters. The TAU-Code flow solver is parallelized by domain decomposition and uses MPI-1 for the inter domain communication. The influence of the different cluster configurations on the parallel performance is investigated with two real life applications, which are typical for the day-to-day work of a CFD-engineer or researcher. The parallel efficiency of the TAU-Code is investigated with two different versions of cache optimization techniques, which will have a significant impact on the scalability of the code. The potential of the installed high speed network to handle the enourmous amount of data transfer caused by a parallel CFD calculation will be evaluated to determine the most promising platform for this type of CFD-Code