I like the idea that Hypre (as a package) would get _one_ comm (for all the solvers/matrices created) that was duped from the one given to PETSc in Vec/MatCreate().
Seems like the tricky part would be figuring out _which_ comm that is based on the incoming comm. For instance - we would definitely have the case where we are doing a Hypre solve on effectively COMM_WORLD… and then many Hypre solves on sub-communicators (and even Hypre solves on sub-communicators of those sub-communicators). The system for getting “the” Hypre Comm would have to match up the incoming Comm in the Vec/MatCreate() call and find the correct Hypre comm to use. Derek On Tue, Apr 3, 2018 at 7:46 AM Satish Balay <[email protected]> wrote: > Fande claimed 49a781f5cee36db85e8d5b951eec29f10ac13593 made a difference. > [so assuming same hypre version was used before and after this commit - for > this bisection] > > So the extra MPI_Comm_dup() calls due to MATHYPRE must be pushing the > total communicators over the limit. > > And wrt debugging - perhaps we need to check MPI_Comm_free() aswell? > Presumably freed communicators can get reused so we have to look for > outstanding/unfreed communicators? > > Per message below - MPICH[?] provides a max of 2048 communicators. And > there is some discussion of this issue at: > https://lists.mpich.org/pipermail/discuss/2012-December/000148.html > > And wrt 'sharing' - I was thining in terms of: Can one use MPI_COMM_WORLD > with all hypre objects we create? If so - we could somehow attach one more > inner-comm - that could be obtained and reused with multiple hypre objects > [that got created off the same petsc_comm?] > > Satish > > On Tue, 3 Apr 2018, Smith, Barry F. wrote: > > > > > Each external package definitely needs its own duplicated > communicator; cannot share between packages. > > > > The only problem with the dups below is if they are in a loop and get > called many times. > > > > To debug the hypre/duplication issue in MOOSE I would run in the > debugger with a break point in MPI_Comm_dup() and see > > who keeps calling it an unreasonable amount of times. (My guess is this > is a new "feature" in hypre that they will need to fix but only debugging > will tell) > > > > Barry > > > > > > > On Apr 2, 2018, at 7:44 PM, Balay, Satish <[email protected]> wrote: > > > > > > We do a MPI_Comm_dup() for objects related to externalpackages. > > > > > > Looks like we added a new mat type MATHYPRE - in 3.8 that PCHYPRE is > > > using. Previously there was one MPI_Comm_dup() PCHYPRE - now I think > > > is one more for MATHYPRE - so more calls to MPI_Comm_dup in 3.8 vs 3.7 > > > > > > src/dm/impls/da/hypre/mhyp.c: ierr = > MPI_Comm_dup(PetscObjectComm((PetscObject)B),&(ex->hcomm));CHKERRQ(ierr); > > > src/dm/impls/da/hypre/mhyp.c: ierr = > MPI_Comm_dup(PetscObjectComm((PetscObject)B),&(ex->hcomm));CHKERRQ(ierr); > > > src/dm/impls/swarm/data_ex.c: ierr = > MPI_Comm_dup(comm,&d->comm);CHKERRQ(ierr); > > > src/ksp/pc/impls/hypre/hypre.c: ierr = > MPI_Comm_dup(PetscObjectComm((PetscObject)pc),&(jac->comm_hypre));CHKERRQ(ierr); > > > src/ksp/pc/impls/hypre/hypre.c: ierr = > MPI_Comm_dup(PetscObjectComm((PetscObject)pc),&(ex->hcomm));CHKERRQ(ierr); > > > src/ksp/pc/impls/hypre/hypre.c: ierr = > MPI_Comm_dup(PetscObjectComm((PetscObject)pc),&(ex->hcomm));CHKERRQ(ierr); > > > src/ksp/pc/impls/spai/ispai.c: ierr = > MPI_Comm_dup(PetscObjectComm((PetscObject)pc),&(ispai->comm_spai));CHKERRQ(ierr); > > > src/mat/examples/tests/ex152.c: ierr = MPI_Comm_dup(MPI_COMM_WORLD, > &comm);CHKERRQ(ierr); > > > src/mat/impls/aij/mpi/mkl_cpardiso/mkl_cpardiso.c: ierr = > MPI_Comm_dup(PetscObjectComm((PetscObject)A),&(mat_mkl_cpardiso->comm_mkl_cpardiso));CHKERRQ(ierr); > > > src/mat/impls/aij/mpi/mumps/mumps.c: ierr = > MPI_Comm_dup(PetscObjectComm((PetscObject)A),&(mumps->comm_mumps));CHKERRQ(ierr); > > > src/mat/impls/aij/mpi/pastix/pastix.c: ierr = > MPI_Comm_dup(PetscObjectComm((PetscObject)A),&(lu->pastix_comm));CHKERRQ(ierr); > > > src/mat/impls/aij/mpi/superlu_dist/superlu_dist.c: ierr = > MPI_Comm_dup(PetscObjectComm((PetscObject)A),&(lu->comm_superlu));CHKERRQ(ierr); > > > src/mat/impls/hypre/mhypre.c: ierr = > MPI_Comm_dup(PetscObjectComm((PetscObject)B),&hB->comm);CHKERRQ(ierr); > > > src/mat/partition/impls/pmetis/pmetis.c: ierr = > MPI_Comm_dup(pcomm,&comm);CHKERRQ(ierr); > > > src/sys/mpiuni/mpi.c: MPI_COMM_SELF, MPI_COMM_WORLD, and a > MPI_Comm_dup() of each of these (duplicates of duplicates return the same > communictor) > > > src/sys/mpiuni/mpi.c:int MPI_Comm_dup(MPI_Comm comm,MPI_Comm *out) > > > src/sys/objects/pinit.c: ierr = > MPI_Comm_dup(MPI_COMM_WORLD,&local_comm);CHKERRQ(ierr); > > > src/sys/objects/pinit.c: ierr = > MPI_Comm_dup(MPI_COMM_WORLD,&local_comm);CHKERRQ(ierr); > > > src/sys/objects/tagm.c: ierr = > MPI_Comm_dup(comm_in,comm_out);CHKERRQ(ierr); > > > src/sys/utils/mpiu.c: ierr = > MPI_Comm_dup(comm,&local_comm);CHKERRQ(ierr); > > > src/ts/impls/implicit/sundials/sundials.c: ierr = > MPI_Comm_dup(PetscObjectComm((PetscObject)ts),&(cvode->comm_sundials));CHKERRQ(ierr); > > > > > > Perhaps we need a PetscCommDuplicateExternalPkg() to somehow avoid > these MPI_Comm_dup() calls? > > > > > > Satish > > > > > > On Tue, 3 Apr 2018, Smith, Barry F. wrote: > > > > > >> > > >> Are we sure this is a PETSc comm issue and not a hypre comm > duplication issue > > >> > > >> frame #6: 0x00000001061345d9 > libpetsc.3.07.dylib`hypre_GenerateSubComm(comm=-1006627852, > participate=<unavailable>, new_comm_ptr=<unavailable>) + 409 at > gen_redcs_mat.c:531 [opt] > > >> > > >> Looks like hypre is needed to generate subcomms, perhaps it generates > too many? > > >> > > >> Barry > > >> > > >> > > >>> On Apr 2, 2018, at 7:07 PM, Derek Gaston <[email protected]> wrote: > > >>> > > >>> I’m working with Fande on this and I would like to add a bit more. > There are many circumstances where we aren’t working on COMM_WORLD at all > (e.g. working on a sub-communicator) but PETSc was initialized using > MPI_COMM_WORLD (think multi-level solves)… and we need to create > arbitrarily many PETSc vecs/mats/solvers/preconditioners and solve. We > definitely can’t rely on using PETSC_COMM_WORLD to avoid triggering > duplication. > > >>> > > >>> Can you explain why PETSc needs to duplicate the communicator so > much? > > >>> > > >>> Thanks for your help in tracking this down! > > >>> > > >>> Derek > > >>> > > >>> On Mon, Apr 2, 2018 at 5:44 PM Kong, Fande <[email protected]> > wrote: > > >>> Why we do not use user-level MPI communicators directly? What are > potential risks here? > > >>> > > >>> > > >>> Fande, > > >>> > > >>> On Mon, Apr 2, 2018 at 5:08 PM, Satish Balay <[email protected]> > wrote: > > >>> PETSC_COMM_WORLD [via PetscCommDuplicate()] attempts to minimize > calls to MPI_Comm_dup() - thus potentially avoiding such errors > > >>> > > >>> > https://urldefense.proofpoint.com/v2/url?u=http-3A__www.mcs.anl.gov_petsc_petsc-2Dcurrent_docs_manualpages_Sys_PetscCommDuplicate.html&d=DwIBAg&c=54IZrppPQZKX9mLzcGdPfFD1hxrcB__aEkJFOKJFd00&r=DUUt3SRGI0_JgtNaS3udV68GRkgV4ts7XKfj2opmiCY&m=jgv7gpZ3K52d_FWMgkK9yEScbLA7pkrWydFuJnYflsU&s=_zpWRcyk3kHuEHoq02NDqYExnXIohXpNnjyabYnnDjU&e= > > >>> > > >>> > > >>> Satish > > >>> > > >>> On Mon, 2 Apr 2018, Kong, Fande wrote: > > >>> > > >>>> On Mon, Apr 2, 2018 at 4:23 PM, Satish Balay <[email protected]> > wrote: > > >>>> > > >>>>> Does this 'standard test' use MPI_COMM_WORLD' to crate PETSc > objects? > > >>>>> > > >>>>> If so - you could try changing to PETSC_COMM_WORLD > > >>>>> > > >>>> > > >>>> > > >>>> I do not think we are using PETSC_COMM_WORLD when creating PETSc > objects. > > >>>> Why we can not use MPI_COMM_WORLD? > > >>>> > > >>>> > > >>>> Fande, > > >>>> > > >>>> > > >>>>> > > >>>>> Satish > > >>>>> > > >>>>> > > >>>>> On Mon, 2 Apr 2018, Kong, Fande wrote: > > >>>>> > > >>>>>> Hi All, > > >>>>>> > > >>>>>> I am trying to upgrade PETSc from 3.7.6 to 3.8.3 for MOOSE and its > > >>>>>> applications. I have a error message for a standard test: > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> *preconditioners/pbp.lots_of_variables: MPI had an > > >>>>>> errorpreconditioners/pbp.lots_of_variables: > > >>>>>> ------------------------------------------------ > > >>>>> preconditioners/pbp.lots_of_variables: > > >>>>>> Other MPI error, error > stack:preconditioners/pbp.lots_of_variables: > > >>>>>> PMPI_Comm_dup(177)..................: > MPI_Comm_dup(comm=0x84000001, > > >>>>>> new_comm=0x97d1068) failedpreconditioners/pbp.lots_of_variables: > > >>>>>> PMPI_Comm_dup(162)..................: > > >>>>>> preconditioners/pbp.lots_of_variables: > > >>>>>> MPIR_Comm_dup_impl(57)..............: > > >>>>>> preconditioners/pbp.lots_of_variables: > > >>>>>> MPIR_Comm_copy(739).................: > > >>>>>> preconditioners/pbp.lots_of_variables: > > >>>>>> MPIR_Get_contextid_sparse_group(614): Too many communicators > (0/2048 > > >>>>> free > > >>>>>> on this process; ignore_id=0)* > > >>>>>> > > >>>>>> > > >>>>>> I did "git bisect', and the following commit introduces this > issue: > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>>> *commit 49a781f5cee36db85e8d5b951eec29f10ac13593Author: Stefano > Zampini > > >>>>>> <[email protected] <[email protected]>>Date: > Sat Nov 5 > > >>>>>> 20:15:19 2016 +0300 PCHYPRE: use internal Mat of type MatHYPRE > > >>>>>> hpmat already stores two HYPRE vectors* > > >>>>>> > > >>>>>> Before I debug line-by-line, anyone has a clue on this? > > >>>>>> > > >>>>>> > > >>>>>> Fande, > > >>>>>> > > >>>>> > > >>>>> > > >>>> > > >>> > > >> > > > > >
