Dave Lee <davelee2...@gmail.com> writes: > Thanks Jed, > > I will reconfigure my PETSc with MUMPS or SuperLU and see if that helps. > (my code is configured to only run in parallel on 6*n^2 processors (n^2 > procs on each face of a cubed sphere, which is a little annoying for > situations like these where a serial LU solver would be handy for testing). > > I've tried setting a tight tolerance on the slab-wise block solve, and > preconditioning based on the pressure, but these haven't helped. > > Unlike other atmosphere codes I'm using slab (layer)-minor indexing, as my > horizontal discretisation is arbitrary order, while I'm only piecewise > constant/linear in the vertical, so I've already got the slab-wise dofs > close together in memory.
Being high order does not necessarily mean more tightly coupled. A useful test is to compute a column of the Jacobian inverse (e.g., by solving with a right-hand side that is a column of the identity) and plot it to see how it extends in each spatial dimension. > I feel like maybe the construction of the Hessenberg during the Arnoldi > iteration and/or the least squares minimisation for the GMRES solve is > leading to additional coupling between the slabs, which is maybe degrading > the convergence, just a thought... GMRES does not care how you order degrees of freedom. My guess based on what I've seen in this thread is that there is a bug in your discretization.
