This report summarizes the weak scaling performance of the diffusion-synthetic acceleration (DSA) of transport sweeps in PDT (a massively parallel deterministic transport solver developed at Texas A\&M University). It provides an assessment of the cost of the DSA based on the modified symmetric interior-penalty discontinuous Galerkin discretization and solved by the AMG-preconditioned conjugate gradient method provided by the HYPRE/BoomerAMG library, relative to the transport sweeps. Different parallel partitionings (``hybrid-KBA'', volumetric), per-core workload (512/4k cells per core) and number of discrete directions (128 to 2k) are compared and a set of optimized HYPRE/BoomerAMG parameters is selected that leads to best overall performance of the transport+DSA solution at scale.
@article{arxiv.1911.04624,
title = {Weak Scaling of DSA Preconditioning of Transport Sweeps using HYPRE},
author = {Milan Hanus},
journal= {arXiv preprint arXiv:1911.04624},
year = {2019}
}