Parallelization in RF3p takes advantage of the first and third approaches listed in “Parallel Processing”. Domain decomposition is not used. During the “AMR” process, different AMR frequencies are solved on separate processes and independent port solutions are generated concurrently via threading. During the final frequency sweep work is divided differently for the discrete sweep and the fast frequency sweep; see “Frequencies” for details on frequency sweeps. During the discrete sweep, contiguous portions of the frequency spectrum are distributed among the available processes subject to available memory with threading used locally to take advantage of any further resources. During the fast frequency sweep, individual expansion points are distributed among available processes subject to available memory, and again, threading is used as possible.
The RF3p linear solver is parallelized as described in “Linear Solver”.