I've experienced the impact of dof-ordering on Krylov solver performance for specific preconditioners. Switching between boost and SCOTCH as dof-ordering library resulted in roughly 4 times difference in Krylov Solver iterations (in serial).
My problem is the (linearized) Navier-Stokes problem on complex geometries, and therefore a good dof ordering would be one that follows the flow.
Since I know the flow direction, could it be possible to order the dofs either a priori (with e.g. a meshfunction with weights)? Alternatively, could it somehow be done "on-the-fly" (e.g. with the previous solution)?
