Optimization of Mesh Partitions and Impact on Parallel CFD

Most of the recently proposed parallel CFD algorithms stem from the "divide and conquer" paradigm and involve some form of domain decomposition. When the discretization is highly regular and when the number of desired subdomains is such that a regular mesh decomposition is possible (i.e. box or strip decompositions), finding an appropriate mesh partition is a trivial task. However, the problem becomes more challenging when dealing with unstructured meshes. Here, we explore a two-step procedure for partitioning a given unstructured mesh and attaining a specific objective. First, we construct an initial partition using a fast deterministic algorithm. Next, we refine the initial decomposition with a non-deterministic optimization algorithm in order to reach a specific goal. This method has proven to produce decompositions with close-to-optimal interface sizes and with a relatively low CPU overhead.