[petsc-dev] Bug introduced in MatGetSubmatrices()

[Zhang, Hong]

We set C->submat_singleis = false as default. 
User turns it on when he knows a single IS will be used for MatGetSubMatrices(), e.g., when asm is used.
User should turn the flag off after it is being used.

This flag is for optimization -- avoid MPI_Allreduce, which is expensive for np>10k.
MatGetSubMatrices() works without this flag turns on. 
I just merged the improved MatGetSubMatrices_MPIAIJ() which only calls MPI_Allreduce once
for repeated use. Previously, it calls MPI_Allreduce twice for each call of MatGetSubMatrices_MPIAIJ().

Hong

________________________________________
From: Barry Smith [bsmith at mcs.anl.gov]
Sent: Friday, January 20, 2017 1:11 PM
To: Matthew Knepley
Cc: Zhang, Hong; PETSc
Subject: Re: [petsc-dev] Bug introduced in MatGetSubmatrices()

> On Jan 20, 2017, at 12:45 PM, Matthew Knepley <knepley at gmail.com> wrote:
>
> On Fri, Jan 20, 2017 at 11:55 AM, Barry Smith <bsmith at mcs.anl.gov> wrote:
>
> > On Jan 20, 2017, at 11:49 AM, Matthew Knepley <knepley at gmail.com> wrote:
> >
> > On Fri, Jan 20, 2017 at 10:49 AM, Hong <hzhang at mcs.anl.gov> wrote:
> > Matt,
> > By default, the flag C->submat_singleis = false.
> > In PCSetUp_ASM(), we set it as 'true' to use MatGetSubMatrices_MPIAIJ_SingleIS().
> >
> > Can you check the value of this flag in your case?
> >
> > The problem is the following:
> >
> >   1) We use MatGetSubMatrices() to extract small matrices in order to form a preconditioner
> >
> >   2) We do this at each Newton iteration
> >
> >   3) We use ASM as a preconditioner for the eventual Newton solve
> >
> >   4) The second time we call MatGetSubMatrices(), it has this flag set, even though we are using multiple ISes
> >
> > Solution: ALSO check that the user is in fact passing a single IS.
> >
>     This requires a communication. I am confused, is the number of IS changing each time? If not why is the flag set?
>
> ASM sets this flag because it knows that IT is going to call MatGetSubMatrices() later, but it unsafe if any user calls
> MatGetSubMatrices() as well. I think overall its a fragile design and should be scrapped.

   You mean the user calls the SAME MatGetSubMatrices with the same matrices, right? Not a completely different unrelated MatGetSubMatrices() which should not be affect by the previous unrelated call.


   Barry

>
>   Matt
>
>
>    Barry
>
> >   Matt
> >
> > Hong
> >
> > It comes from here:
> >
> >   https://bitbucket.org/petsc/petsc/commits/c10200c1442b553b7ad65c70101560db4fa22e78
> >
> > If we ask for more than 1 matrix, it dispatches to
> >
> >   MatGetSubMatrices_MPIAIJ_SingleIS()
> >
> > but then fails here
> >
> >   https://bitbucket.org/petsc/petsc/annotate/2e559809f9aee9c95ee79eb0939630cfe5502c8d/src/mat/impls/aij/mpi/mpiov.c?at=master&fileviewer=file-view-default#mpiov.c-1306
> >
> > because ismax > 1. I think the ismax check needs to move up to here
> >
> >   https://bitbucket.org/petsc/petsc/annotate/2e559809f9aee9c95ee79eb0939630cfe5502c8d/src/mat/impls/aij/mpi/mpiov.c?at=master&fileviewer=file-view-default#mpiov.c-2012
> >
> > but I don't know for sure. Please fix this since it is breaking PyLith.
> >
> >   Matt
> >
> > --
> > 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
> >
> >
> >
> >
> >
> > --
> > 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
>
>
>
>
> --
> 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