[linear] Linear solve time (UMFPack) increased after reordering of free-flow matrix blocks
!1877 (merged) change the matrix block structure for free flow problems to be more consistent with the literature.
This change however increases the linear solve time for UMFPack for a larger system of one of my applications by a factor of 5. I implemented a hacky way to change back to the older reordering !2495 which restores the original performance.
My case is a coupled free-flow/pore-network problem (1p2cni-2p2cni) so the physics are rather complex.
Did anyone experience something similar? Would there be a way to heuristically change/optimize the block order?