Given the discussion in Section , it is most
convenient to describe our prototype implementation of a parallel sweeping
preconditioner based upon its deviations from our proposed approach. The
primary difference is that there is not yet support for simultaneously
factoring the subdomain auxiliary problems and then redistributing each frontal
tree to the entire set of processes. This will certainly lead to large
improvements in the scalability of the setup phase, but it is left for future
work.