A parallel version of Polyflow

摘要

A domain decomposition technique is used to introduce coarse grain parallelism in the Polyflow solver. Two methods are used: either on a full direct approach (implementing a nested dissection algorithm on the top of a frontal solver), or an hybrid direct/iterative approach. Results on non-elliptic flow problems are presented. Parallel efficiencies which can even be super-linear on a relatively small number of processors are reported. An original non-deterministic heuristic helps the user to find a suitable elimination tree.