On 04.17.2013 10:26, Jozsef Bakosi wrote: > On 04.05.2013 08:34, Jozsef Bakosi wrote: > > Hi folks, > > > > In switching from 3.1-p8 to 3.3-p6, keeping the same ML ml-6.2.tar.gz, I get > > indefinite preconditioner with the newer PETSc version. Has there been > > anything > > substantial changed around how PCs are handled, e.g. in the defaults? > > > > I know this request is pretty general, I would just like to know where to > > start > > looking, where changes in PETSc might be clobbering the (supposedly same) > > behavior of ML. > > > > Alright, here is a little more information about what we see. Running the same > setup/solve using ML (using the same ML and application source code) and > switching from PETSc 3.1-p8 to 3.3-p6 appears to work differently, in some > cases, resulting in divergence compared to the old version. > > I attach the output from KSPView() called after KSPSetup() for the 3.1-p8 > (old.out) and for the 3.3-p6 (new.out), both running on 4 MPI ranks. > > A diff reveals some notable differences: > > * using (PRECONDITIONED -> NONE) norm type for convergence test > > * (using -> not using) I-node routines > > * tolerance for zero pivot (1e-12 -> 2.22045e-14) for > PPE_mg_levels_[12]_sub_ > (stayed the same for PPE_mg_coarse_redundant_) > > So we are wondering what might have changed in the PETSc defaults around how > PCs, in particular ML, is used. > > Thanks, and please let me know if I can give you more information, > > Thanks, > Jozsef
Another piece of info: The error we get is indefinite preconditioner. Interestingly the same setup produces an OK preconditioner using 1,2,3, and 5 ranks -- indefinite only on 4. ps: Please CC me as I'm not on the email list. Thanks, J
