Mark Adams <mfad...@lbl.gov> writes: > That is odd. Are these problems symmetric positive definite? > > Eigen estimates are a pain in practice but I've never seen this. Hypre has > (better) smoothers that don't need this and the AMG algorithm does not need > them either. I think ML does pretty much the same thing as me.
We could certainly offer an "l1 Jacobi" for operator smoothing; it's effectively using the Gershgorin bound, though it can be very pessimistic. At a minimum, it would be useful diagnostic information and we could stop the Krylov estimate early if its maxbound (which includes safety and is non-increasing with iteration) exceeds the Gershgorin bound. > If SPD then you definitely want '-pc_gamg_esteig_ksp_type cg'. CG converges > faster and is more robust. Mark, this comes up a lot. Should we select it by default if MAT_SPD is set? > Even if not strictly SPD like with large > deformation or plasticity CG is probably better. You can run with -info and > grep on GAMG to see what eigen (high) estimates it came up with for each > level. They should be >1 and < 4, or so. > > I wonder if your LAPACK is funky. You might try a fresh build (delete the > "arch" directory) and download LAPACK, but this problem will go away with > CG. Sanjay, you could use the debugger to print out the small dense matrix that LAPACK is choking on. It isn't supposed to divide by zero ever.
