An efficient and robust localization scheme on massively distributed systems

Abstract

A parallel localization scheme is presented to enable solution transfers between unstructured grids. The scheme relies on neighbor walks to reduce the number of candidate elements that are visited to find the enclosing element. An advancing front method efficiently allows a subset of nodes to efficiently sweep through the grid, progressively reducing search spaces. The algorithm is parallelized permitting solution transfers over arbitrary grid decompositions. A hierarchical localization process helps prevent the neighbor walk algorithm from failing when encountering the boundaries of a concave domain by localizing the boundaries before the interior of the domain is localized. Random selections of the next step interrupt cyclic loops that may occur during a neighbor walk. The complexity of the search algorithm is verified over parallel decompositions and is effectively independent of the number of partitions.