I see, I had a tiny time step that make the Jacobian big and the errors small.
On Thu, Aug 17, 2017 at 12:30 AM, Mark Adams <[email protected]> wrote: > > > On Thu, Aug 17, 2017 at 12:22 AM, Matthew Knepley <[email protected]> > wrote: > >> On Wed, Aug 16, 2017 at 10:14 AM, Mark Adams <[email protected]> wrote: >> >>> I just want to check the interpretation of snes_type test. Is this >>> pretty conclusively good? >>> >> >> It looks right, but the entires in the Jacobian are crazy. You should >> non-dimensionalize first. >> > > This is non-dimensionalize and it is 2D FE 2nd order equation so > everything should be O(1), that is, there is no 1/h like term. And the > domain is (1) and the grid is small so h is O(1) anyway. > > >> >> Matt >> >> >>> Testing hand-coded Jacobian, if the ratio is >>> O(1.e-8), the hand-coded Jacobian is probably correct. >>> Run with -snes_test_display to show difference >>> of hand-coded and finite difference Jacobian. >>> Norm of matrix ratio 1.38299e-16, difference 1.11738 (user-defined state) >>> Norm of matrix ratio 9.50529e-09, difference 7.67972e+07 (constant state >>> -1.0) >>> Norm of matrix ratio 1.07438e-08, difference 8.68036e+07 (constant state >>> 1.0) >>> >>> >> >> >> -- >> What most experimenters take for granted before they begin their >> experiments is infinitely more interesting than any results to which their >> experiments lead. >> -- Norbert Wiener >> >> http://www.caam.rice.edu/~mk51/ >> > >
