Parallel unstructured mesh CFD codes: A role for recursive clustering techniques in mesh decomposition

Abstract

In principle, unstructured mesh computational fluid dynamics (CFD) codes can be parallelized using a mesh decomposition approach similar to structured mesh codes. However, for unstructured codes the mesh structure is problem dependent and algorithms for automatically decomposing the mesh onto the processors are required. An algorithm based upon a recursive clustering technique for decomposing meshes into an arbitrary number of subdomains with a nearest neighbor connectivity for a specified processor topology is used as a preprocessor to a parallel unstructured mesh CFD code. The key result indicated in this chapter is that the control volume based unstructured mesh CFD codes can be effectively mapped onto distributed-memory parallel architectures with efficiencies that are equivalent to structured mesh codes. Undoubtedly, an important factor in this success is the mesh decomposition algorithm that maps onto to processor topology and so minimizes the distance over which communications have to travel and minimizes the number of interface nodes in the halo layers among neighboring submeshes, and therefore the length of messages exchanged among neighboring processors in the chain.