I've added a flag: -pc_gamg_reuse_interpolation <true>
which can be set to false to redo the setup in GAMG every time. '-pc_ml_reuse_interpolation true' does seem to get ML to reuse some mesh setup. The setup time goes from .3 to .1 sec on one of my tests from the first to the second solve. Hong: this looks like a way to infer ML's RAP times. I think the second solves are just redoing the RAP with this flag, like what GAMG does by default. On Jan 7, 2013, at 11:36 AM, Mark F. Adams <mark.adams at columbia.edu> wrote: >> >> ML isn't using eigenvalue estimation (doesn't expose the algorithm). Sanjay >> is using the default smoother (Richardson + SOR) rather than >> chebyshev/pbjacobi. >> > > The only other problem that I can think of is that ML is not getting the null > space correctly. Unfortunately I do not see anything in MLs verbose output > that says what it is doing for the null space. > > It would be good to know if GAMG has the same problem.
