A highly scalable approach to solving linear systems using two-stage multisplitting

Abstract: Iterative methods for solving large sparse systems of linear equations are widely used in many HPC applications. Extreme scaling of these methods can be difficult, however, since global communication to form dot products is typically required at every iteration. To try to overcome this limitation we propose a hybrid approach, where the matrix is partitioned into blocks. Within each block, we use a highly optimised (parallel) conventional solver, but we then couple the blocks together using block Jacobi or some other multisplitting technique that can be implemented in either a synchronous or an asynchronous fashion. This allows us to limit the block size to the point where the conventional iterative methods no longer scale, and to avoid global communication (and possibly synchronisation) across all processes. Our block framework has been built to use PETSc, a popular scientific suite for solving sparse linear systems, as the synchronous intra-block solver, and we demonstrate results on up to 32768 cores of a Cray XE6 system. At this scale, the conventional solvers are still more efficient, though trends suggest that the hybrid approach may be beneficial at higher core counts.