Is direct solving method supported in palace?

[yanfeit]

No description provided.

[hughcars]

@yanfeit I am not sure what you mean by "direct" in this context, if you mean using a dense/direct linear solver, then no Palace does not support this as it is fundamentally built around iterative krylov methods. If you can clarify your question a bit I might be able to provide more of an answer.

[yanfeit]

@yanfeit I am not sure what you mean by "direct" in this context, if you mean using a dense/direct linear solver, then no Palace does not support this as it is fundamentally built around iterative krylov methods. If you can clarify your question a bit I might be able to provide more of an answer.

Thank you for the kind reply. I was meant to ask whether Palace supports sparse matrix solver based on LU decomposition. I think you have already answered my question.

[hughcars]

@yanfeit I am not sure what you mean by "direct" in this context, if you mean using a dense/direct linear solver, then no Palace does not support this as it is fundamentally built around iterative krylov methods. If you can clarify your question a bit I might be able to provide more of an answer.

Thank you for the kind reply. I was meant to ask whether Palace supports sparse matrix solver based on LU decomposition. I think you have already answered my question.

Ah, we do make use of SuperLU and STRUMPACK (as seen here in the docs) as the coarse preconditioner in a multigrid hierarchy when doing frequency and time domain solves. However they are solving a real valued approximation of the complex system, thus are not directly solving the system. There is a project we have planned to solve the complex system matrix itself, but that hasn't been attempted yet. Historically we have found the sparse direct solvers are very helpful for the coarsest solve in a multigrid hierarchy, as the system can be quite indeterminate and thus difficult for traditional preconditioning techniques.

[sebastiangrimberg]

Hi @yanfeit, sorry to chime in here but to add to Hugh's answer there is a keyword "MGMaxLevels" which you can set to 1 in your configuration file to disable multigrid and use a sparse direct solver on the assembled (potentially high-order) system. As Hugh says this is real-valued, but the GMRES solve should converge in very few iterations as it is typically a (provably) good preconditioner for the actual complex-valued system.

[yanfeit]

Hi @sebastiangrimberg , thanks for more clarifications and instructions. Since the answer has some technical details that I am not familiar with, I will discuss your answer with my colleagues.