On Jan 7, 2013, at 8:15 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
> On Mon, Jan 7, 2013 at 8:12 PM, Barry Smith <bsmith at mcs.anl.gov> wrote: > > On Jan 7, 2013, at 8:05 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote: > > > It's sort of awkward to deal with PCREDUNDANT for coarse grids, and I have > > yet to see any evidence that it's better than gathering to rank 0, solving > > sequentially, and scattering the result. What about implementing LU for > > MPIAIJ to do the latter algorithm? Then we could trivially make it smart > > enough to skip the gather when Mark uses it on the coarse grid of GAMG > > (where the entire system is already on rank 0). In additon to being simpler > > for the user to interact with, this would fix a lot of fighting between the > > PCMG and its children. > > I thought PCREDUNDANT had this ability? What about adding it? > > That wouldn't be redundant at all, so if we put it there, we'd have to at > least rename it. I can teach you how to use etags or you can teach me how to use PERL :-) > > Have an option that cause PCREDUNDANT to only collect a single copy of the > matrix from rank 0 to r-1 and in your case r is 1. I hate the idea of having > another separate place with similar stuff to PCREDUNDANT. You could then > implement LU for MPIAIJ using this variant of PCREDUNDANT, that is a call to > PCREDUNDANT stuff with the right options. > > Barry > > > >
