Improving the Performance of Large-Scale Unstructured PDE Applications

Abstract. This paper investigates two types of overhead due to duplicated local computations, which are frequently encountered in the parallel software of overlapping domain decomposition methods. To remove the duplication-induced overhead, we propose a parallel scheme that disjointly re-distributes the overlapping mesh points among irregularly shaped subdomains. The essence is to replace the duplicated local computations by an increased volume of the inter-processor communication. Since the number of inter-processor messages remains the same, the bandwidth consumption by an increased number of data values can often be justified by the removal of a considerably larger number of floating-point operations and irregular memory accesses in unstructured applications. Obtainable gain in the resulting parallel performance is demonstrated by numerical experiments.

• Type: Conference
• Year: 2004
• Where: PARA
• Authors: Xing Cai
• Keywords: Applied Computing, Domain Decomposition Methods, Duplicated Local Computations, Overlapping Mesh Points, PARA 2004