Advances in the sparse linear system solver Spliss: Mixed-Precision and Stripped Linear Multigrid Configurations

Kurzfassung

Within the German Aerospace Center (DLR), Spliss serves as the established sparse linear system solver for Computational Fluid Dynamics (CFD) applications, e.g. CODA or TRACE. In this talk, we present recent investigations into optimizing Spliss through mixed-precision arithmetic and varied linear multigrid configurations, analyzing their impact on both performance and solution accuracy. A key focus of this work is the presentation of a stripped linear multigrid configuration. In this approach, coarse-grid solvers operate exclusively on local matrix partitions, thereby avoiding global communication overhead. While this strategy yields inherently approximate coarse-grid corrections, we demonstrate that applying an outer solver to smooth the residual effectively compensates for the reduced accuracy. Crucially, this hybrid methodology converges to the exact solution of the original system, despite the communication-avoiding simplifications. Non-linear, steady-state computations, using the flow solver CODA, investigate the performance and scalability impacts of this technique.