Scalable Two-Level Preconditioners for CFD Computations on Many-Core Systems

Abstract

At the German Aerospace Center (DLR), the parallel simulation systems TAU and TRACE have been developed for the aerodynamic design of aircrafts or turbines for jet engines. Large-scale computing resources allow more detailed numerical investigations with bigger numerical configurations. Up to 50 million grid points in a single simulation are becoming a standard configuration in the industrial design. Both CFD solvers, TAU and TRACE, require the parallel solution of large, sparse real or complex systems of linear equations. For the parallel iterative solution of these equation systems, various block-local preconditioners are compared with a global two-level incomplete factorization preconditioning method for real or complex matrix problems. The two-level preconditioner is based on an improved distributed Schur complement ansatz. Compared with the original Schur complement method, it significantly saves incomplete factorization operations. Numerical, performance and scalability results of preconditioned FGMRes algorithms are presented for typical TAU and TRACE problems from external or internal flow simulations on many-core systems.