Improving multigrid performance for unstructured mesh drift–diffusion simulations on 147,000 cores

SUMMARY This study considers the scaling of three algebraic multigrid aggregation schemes for a finite element discretization of a drift–diffusion system, specifically the drift–diffusion model for semiconductor devices. The approach is more general and can be applied to other systems of partial differential equations. After discretization on unstructured meshes, a fully coupled multigrid preconditioned Newton–Krylov solution method is employed. The choice of aggregation scheme for generating coarser levels has a significant impact on the performance and scalability of the multigrid preconditioner. For the test cases considered, the uncoupled aggregation scheme, which aggregates/combines the immediate neighbors, followed by repartitioning and data redistribution for the coarser level matrices on a subset of the Message Passing Interface (MPI) processes, outperformed the two other approaches, including the baseline aggressive coarsening scheme. Scaling results are presented up to 147,456 cores on an IBM Blue Gene/P platform. A comparison of the scaling of a multigrid V-cycle and W-cycle is provided. Results for 65,536 cores demonstrate that a factor of 3.5 × reduction in time between the uncoupled aggregation and baseline aggressive coarsening scheme can be obtained by significantly reducing the iteration count due to the increased number of multigrid levels and the generation of better quality aggregates. Copyright © 2012 John Wiley & Sons, Ltd.