Parallel memory-efficient all-at-once algorithms for the sparse matrix triple products in multigrid methods

Abstract

Multilevel/multigrid methods is one of the most popular approaches for solving a large sparse linear system of equations, typically, arising from the discretization of partial differential equations. One critical step in the multilevel/multigrid methods is to form coarse matrices through a sequence of sparse matrix triple products. A commonly used approach for the triple products explicitly involves two steps, and during each step a sparse matrix-matrix multiplication is employed. This approach works well for many applications with a good computational efficiency, but it has a high memory overhead since some auxiliary matrices need to be temporarily stored for accomplishing the calculations. In this work, we propose two new algorithms that construct a coarse matrix with taking one pass through the input matrices without involving any auxiliary matrices for saving memory. The new approaches are referred to as "all-at-once" and "merged all-at-once", and the traditional method is denoted as "two-step". The all-at-once and the merged all-at-once algorithms are implemented based on hash tables in PETSc as part of this work with a careful consideration on the performance in terms of the compute time and the memory usage. We numerically show that the proposed algorithms and their implementations are perfectly scalable in both the compute time and the memory usage with up to 32,768 processor cores for a model problem with 27 billions of unknowns. The scalability is also demonstrated for a realistic neutron transport problem with over 2 billion unknowns on a supercomputer with 10,000 processor cores. Compared with the traditional two-step method, the all-at-once and the merged all-at-once algorithms consume much less memory for both the model problem and the realistic neutron transport problem meanwhile they are able to maintain the computational efficiency.