Stefano, Checking PCRedundant, I see that current PCRedundant is hard-coded using PETSC_SUBCOMM_INTERLACED which 'avoid data scattering from subcomm back to original comm' See ~petsc/src/sys/objects/subcomm.c - PetscSubcommCreate_interlaced() Unless you write a new implementation, PETSC_SUBCOMM_CONTIGUOUS cannot be used for PCRedundant.
Hong On Mon, Oct 15, 2012 at 4:37 PM, Hong Zhang <hzhang at mcs.anl.gov> wrote: > Stefano : > >> Hong, INTERLACED is the default since >> http://petsc.cs.iit.edu/petsc/petsc-dev/rev/67df4fdba5bd >> >> See also http://petsc.cs.iit.edu/petsc/petsc-dev/rev/c04222d17a3d > > > Hmm, you are correct :-( > > >> Hong, in petsc-dev, default is interlaced. I ran some tests, I got malloc >>> errors when inserting mat values using subcomm of contiguous type. I think >>> that PCRedundant with PETSC_SUBCOMM_CONTIGUOUS is buggy. >> >> > PETSC_SUBCOMM_CONTIGUOUS was introduced for multiprocessor bjacobi > preconditioner, > which has not been used/tested in other situations, likely causes problems. > Can you send us a stand-alone code that reproduces the malloc errors? > > Hong > >> >>> >>> 2012/10/15 Hong Zhang <hzhang at mcs.anl.gov> >>> >>>> Stefano : >>>> >>>>> >>>>> is it possible to use PCRedundant with PETSC_SUBCOMM_CONTIGUOUS as >>>>> subcomm type? >>>>> >>>> I think "PETSC_SUBCOMM_CONTIGUOUS" is the default for PCRedundant. >>>> >>>> Hong >>>> >>> >>> >>> >>> -- >>> Stefano >>> >> >> >> >> -- >> Stefano >> > > -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20121016/db1bf8ee/attachment.html>
