Core Concepts
Flexible inner-outer Krylov subspace methods combined with spectral deflation and advanced preconditioning techniques can efficiently solve large, sparse and ill-conditioned adjoint linear systems arising from high-fidelity computational fluid dynamics models.
Abstract
The content presents a comparative study of inner-outer Krylov solvers for the efficient solution of large, sparse and ill-conditioned linear systems that arise in adjoint-based aerodynamic shape optimization problems.
Key highlights:
The authors investigate the use of flexible inner-outer GMRES (FGMRES) solvers, which allow for variable preconditioning, in conjunction with spectral deflation techniques to address the challenges posed by the stiff adjoint systems.
Two different discretization schemes are considered - a finite volume (FV) method and a high-order discontinuous Galerkin (DG) method. This affects the arithmetic intensity and memory-bandwidth requirements of the linear algebra operations.
For the FV case, the authors compare the performance of FGMRES preconditioned by LU-SGS and BILU(0) applied to either an approximate or an exact Jacobian matrix. Deflation is shown to be crucial to overcome the stagnation of the standard GMRES solver.
For the DG case, the BILU(0) preconditioner combined with a Restricted Additive Schwarz (RAS) domain decomposition method is used. Deflation provides a slight improvement in convergence.
Strong scalability analysis demonstrates satisfactory parallel efficiency for both the FV and DG cases, with the DG case exhibiting some memory-bandwidth limitations.
The authors discuss the recommended numerical practices based on the representative test cases.
Stats
The number of nonzero entries (NNZ) of the preconditioner matrices are:
For the FV case:
JEX_O1: 28 million NNZ
J_APP_O1: 21 million NNZ
For the DG case:
J_EX_O3: 288 million NNZ
Quotes
"Typically in our implementation the efficiency of the preconditioner is enhanced with a domain decomposition method with overlapping."
"Maintaining the performance of the preconditioner may be challenging since scalability and efficiency of a preconditioning technique are properties often antagonistic to each other."
"We demonstrate how flexible inner-outer Krylov methods are able to overcome this critical issue."