Please run with the -snes_fd and one level of grid sequencing (where Newton was having trouble converging on the second level) with -snes_monitor -ksp_monitor_true_residual and send the output. Too much speculation and not enough data.
Barry > On Sep 9, 2015, at 11:14 AM, Gideon Simpson <[email protected]> wrote: > > Checking the Jacobian, this is a pretty standard output: > > Testing hand-coded Jacobian, if the ratio is O(1.e-8), the hand-coded > Jacobian is probably correct. > Finite difference Jacobian > Hand-coded Jacobian > Hand-coded minus finite difference Jacobian > 6.08281e-10 = ||J - Jfd||//J|| 0.000151055 = ||J - Jfd|| > > > Things look favorable in terms of relative difference, but the absolute > difference is a bit more suspect. > > -gideon > >> On Sep 8, 2015, at 11:49 PM, Barry Smith <[email protected]> wrote: >> >>> >>> On Sep 8, 2015, at 10:28 PM, Gideon Simpson <[email protected]> >>> wrote: >>> >>> I should clarify, in that run, I ran with the analytic jacobian, but I did >>> not use -snes_mf_operator. If I use that flag, the performance is a bit >>> different. In particular, the true residual norms are not as good. How >>> should I interpret that? >> >> The "analytic Jacobians" are wrong. Of course, that is easy to say, the >> hard part is figuring out exactly what entries are wrong. >> >> You can try the following options. They are experimental so may be flaky >> run with the options below (but not -snes_fd or -snes_mf_operators) >> >> -snes_check_jacobian >> -snes_check_jacobian_view >> >
