Jed brings up good points about false negatives, there are also false positives to worry about. This is a bad idea, at least as a default solver.
That said, there is work on "adaptive" solvers like bootstrap AMG and using machine learning to formalize what it sounds like Barry is driving at. Perhaps make a '-ksp_type black_box' solver type where you put any heuristics methods like this. (this would require the KSP method to change the PC method which is ugly...) Mark On Aug 5, 2010, at 5:36 AM, Matthew Knepley wrote: > On Thu, Aug 5, 2010 at 2:10 AM, Jed Brown <jed at 59a2.org> wrote: > On Thu, 5 Aug 2010 00:21:13 -0500, Matthew Knepley > <knepley at gmail.com> wrote: > > If someone tells us, "I have a Stokes problem", we could search > for 0 > > diagonals and create the partition for FS. > > I suppose this would work for some common cases, but there are lot of > discretizations and pressure-dependent constitutive relations/boundary > conditions that have nonzeros in the pressure block. I think some > slip > conditions can also produce zero or negative values in the momentum > block. > > Do not disagree. > > I'm not convinced that it's so much to ask people to provide an index > set, considering that FieldSplit is a somewhat advanced thing anyway > (based on sheer number of choices available, and (typical) sensitivity > to those choices). > > I guess the point here is somewhat like the point of DAs. It is a > very limited > thing, but something many people do. So we provide a way to most > easily > do the very limited thing. > > Not sure if this is useful enough, but it might be. > > Matt > > > Jed > > > > -- > 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 -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20100805/7e275749/attachment.html>
