Beowulf performance in CFD multigrid applications

Abstract

Computational fluid dynamics is probably one of the most computationally demanding disciplines, a driving force behind the development of new computer architectures. In fact, the design and evaluation of high-performance parallel systems is commonly based on CFD workloads. One of the most remarkable examples of such workloads is the NAS parallel benchmark, which aims to mimic the computation and data-movement characteristics of large scale CFD applications. We have paid specific attention to the NAS-MG (multigrid) kernel, since these methods represent one of the most promising solvers in the field of CFD. Nevertheless, practical flow computations demand robust multigrid algorithms which differ from the NAS-MG kernel. The paper presents a performance evaluation of a Beowulf system using both a state-of-the-art multigrid solver and the NAS-MG benchmark. These two codes have been used to compare several of its design choices, namely, the interconnection network (GigaNet versus Fast-Ethernet) as well as the node configuration (dual nodes versus single nodes). The results highlight that the optimal combination strongly depends on the target application.