FYI, ML has some different defaults (eg, its "threshold" is 0 and GAMG is finite as I recall). I have "verified" GAMG against ML a year or so ago, but I had to set: -pc_gamg_threshold 0. Mark
On Fri, Jun 12, 2015 at 7:38 AM, Matthew Knepley <knep...@gmail.com> wrote: > On Fri, Jun 12, 2015 at 6:24 AM, Garth N. Wells <gn...@cam.ac.uk> wrote: > >> I've noticed very different performance with ML when I (a) link PETSc >> to a Trilinos installation; or (b) use --download-ml. I observe >> >> - Case (a): solver behaves exactly as I would expect. Iteration counts >> very close to GAMG iteration counts, minimal dependency on problem >> size. >> >> - Case (b): iteration counts higher and substantial increase in >> iteration count with problem size. >> >> I haven't debugged since I usually link to Trilinos, in which case >> performance is as I expect. > > > ML stopped being distributed independent of Trilinos a while ago, so we > might not > have upstream changes. This sounds like a difference of settings. The > first step I think > is to look at -ksp_view for both runs. > > Thanks, > > Matt > > >> Garth >> > > > > -- > 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 >