Mark, I compared the eigen estimates and they were identical. I found an issue that was somehow miraculously covered up by -random_seed 1 for 10,000 matrix solves. I was initially sent on a wild chase by noting a strong behavior difference between DIFFERENT_NONZERO_PATTERN and SAME_PRECONDITIONER, where DIFFERENT_NONZERO_PATTERN gave the correct solution, but it was totally a fluke. Sorry about that.
John On Fri, Jul 12, 2013 at 12:31 PM, Mark F. Adams <[email protected]>wrote: > > > > > > if (set) { > > ierr = PetscRandomSetSeed(rnd,(unsigned long int)seed);CHKERRQ(ierr); > > ierr = PetscRandomSeed(rnd);CHKERRQ(ierr); > > } > > > > Otherwise the default seed is used, but that changes for repeat solves. > > OK, right I see. > > Anyway it is bad that the code is sensitive to the random RHS in the eigen > estimator. These operators are not symmetric apparently and I'm not sure > about the nature of this non-symetry and how that can effect eigen > estimators. This is a fundamental disadvantage of smoothed aggregation. > > So at this point it is not clear if 1) there is an simple code/parameter > change in the code that can fix this, 2) a code bug, or 3) a fundamental > problem with cheby on these (asymmetric operators). > > It would be nice to verify and get some number on the differences in eigen > estimates in these two (good and bad) solves. > > Also, hypre is a good robust alternative if you need to get stuff done now.
