[petsc-users] Doubt in MPIPreallocation

Sun Feb 21 08:11:55 CST 2016

Thank you very much *Barry* and *Matt*, for your help. Its working now. The
way I did is I created a dummy matrix of the same layout with the name *'A'.
*Then I set it up using MatSetUp(A); . Once this is done, I got the local
sizes of the matrix, then created the o_nnz[] and d_nnz[]. After that I
*destroyed* *A*, and made the actual working matrix *B, *with the same
layout and set the preallocation with the already calculated o_nnz[] and
d_nnz[]. And checked it via MatView. :)

Many thanks.

----
John Albequerque.

On Sat, Feb 20, 2016 at 3:07 AM, Barry Smith <bsmith at mcs.anl.gov> wrote:

>
> > On Feb 19, 2016, at 3:31 PM, John Albequerque <johncfdcfd at gmail.com>
> wrote:
> >
> > Barry,
> > I have let PETSc decide the local sizes I have only specified the global
> values,
>
>    This cannot work. The preallocation information is provided on a
> process by process information. Each process has to provide information for
> its rows. This means you need to know and set the number of local rows of
> the matrix you cannot use PETSC_DECIDE.
>
> > I will also check with the debugger and tell you the status.
>
>
> >
> > And just another query,is it a problem if I have already used
> MatSetUp(A); once and after that assembled the matrix and after that
> applying Preallocation. I guess that should not be a problem as the code
> ran fine with the one processor case. Just confirming though.
>
>   No you only do one preallocation ever. If you need another matrix with a
> different preallocation you create a new matrix and set its preallocation
> (and delete the old one if you don't need it anymore).
>
>   Barry
>
>
> >
> > Many thanks.
> >
> > ----
> > John Albequerque.
> >
> > On Sat, Feb 20, 2016 at 2:53 AM, Barry Smith <bsmith at mcs.anl.gov> wrote:
> >
> >   Are you sure you have divided up the matrix so each process gets some
> of the rows?
> >
> >   You are just going to have to break down and run with
> -start_in_debugger to see exactly why it is crashing.
> >
> >   Barry
> >
> > > On Feb 19, 2016, at 3:16 PM, John Albequerque <johncfdcfd at gmail.com>
> wrote:
> > >
> > > Barry,
> > > I forgot to mention that the code is working fine with a single
> processor, but when I increase the number of processors, as in this case 3
> processors have been used when the error message was generated. And yes, it
> has something to do with the off-diagonal terms as I am increasing the
> number of processes, the same error triggers for each row for which
> o_nnz[i] is not zero.
> > >
> > >
> > > Many thanks.
> > >
> > > ----
> > > John Albequerque.
> > >
> > > On Sat, Feb 20, 2016 at 2:38 AM, Barry Smith <bsmith at mcs.anl.gov>
> wrote:
> > >
> > >   You are setting nonzero preallocations for the "off diagonal"
> portion of the matrix, but one one process there is no off diagonal portion
> so you should be setting those values to all zero.
> > >
> > >   Barry
> > >
> > > > On Feb 19, 2016, at 2:29 PM, John Albequerque <johncfdcfd at gmail.com>
> wrote:
> > > >
> > > >
> > > > Dear Barry,
> > > > I am sorry I could not get you. I have also posted the entire error
> message so that you could get a deeper insight into it.
> > > >
> > > >
> > > > Many thanks.
> > > >
> > > > ----
> > > > John Albequerque.
> > > >
> ========================================================================================
> > > > Argument out of range
> > > > [0]PETSC ERROR: nnz cannot be greater than row length: local row 0
> value 1 rowlength 0
> > > > [0]PETSC ERROR: See
> http://www.mcs.anl.gov/petsc/documentation/faq.html for trouble shooting.
> > > > [0]PETSC ERROR: Petsc Release Version 3.6.3, unknown
> > > > [0]PETSC ERROR: ./MatSparse on a linux-gnu-c-debug named John by
> johncfd Sat Feb 20 01:36:01 2016
> > > > [0]PETSC ERROR: Configure options --with-cc=gcc --with-cxx=g++
> --with-fc=gfortran --download-fblaslapack --download-mpich
> > > > [0]PETSC ERROR: #1 MatSeqAIJSetPreallocation_SeqAIJ() line 3567 in
> /home//MyStuff/ClonedRepos/petsc/src/mat/impls/aij/seq/aij.c
> > > > [0]PETSC ERROR: #2 MatSeqAIJSetPreallocation() line 3539 in
> /home/MyStuff/ClonedRepos/petsc/src/mat/impls/aij/seq/aij.c
> > > > [0]PETSC ERROR: #3 MatMPIAIJSetPreallocation_MPIAIJ() line 2835 in
> /home/MyStuff/ClonedRepos/petsc/src/mat/impls/aij/mpi/mpiaij.c
> > > > [0]PETSC ERROR: #4 MatMPIAIJSetPreallocation() line 3532 in
> /home/MyStuff/ClonedRepos/petsc/src/mat/impls/aij/mpi/mpiaij.c
> > > >
> =========================================================================================
> > > >
> > > >
> > > > On Sat, Feb 20, 2016 at 12:35 AM, Barry Smith <bsmith at mcs.anl.gov>
> wrote:
> > > >
> > > > > On Feb 19, 2016, at 12:54 PM, John Albequerque <
> johncfdcfd at gmail.com> wrote:
> > > > >
> > > > > Jed, one more
> > > > >
> > > > > nnz cannot be greater than row length: local row 0 value 1
> rowlength 0
> > > >
> > > >   Always send the entire error message, this provides the context to
> know what the hey is going on.
> > > >
> > > >   It looks like you set a matrix block (if you are using an MPI
> matrix on one process this could the the "off-diagonal" block)  which has
> no columns (hence row length is zero) but you claim you need to preallocate
> an entry (the 1).
> > > >
> > > >   Barry
> > > >
> > > > >
> > > > > How do I deal with this error?
> > > > >
> > > > > Thanks.
> > > > >
> > > > > ----
> > > > > John Albequerque.
> > > > >
> > > > >
> > > > >
> > > > > On Fri, Feb 19, 2016 at 8:17 PM, John Albequerque <
> johncfdcfd at gmail.com> wrote:
> > > > > Thank you very much, I will try it.
> > > > >
> > > > >
> > > > > Thanks
> > > > >
> > > > > ----
> > > > > John Albequerque.
> > > > >
> > > > >
> > > > >
> > > > > On Fri, Feb 19, 2016 at 8:16 PM, Jed Brown <jed at jedbrown.org>
> wrote:
> > > > > John Albequerque <johncfdcfd at gmail.com> writes:
> > > > >
> > > > > > So Jed, what you are suggesting is that I should set only the
> non-zero
> > > > > > elements while using
> > > > > > *  MatSetValues(A,(high-low),**idxm,nc,idxn,values,INSERT_*
> > > > > >
> > > > > > *VALUES);*
> > > > > > And not mention the zero elements and for that I should loop
> over all local
> > > > > > rows and then set each of the value.
> > > > >
> > > > > Yes.  The whole point of a sparse matrix is to spend no time or
> storage
> > > > > on 0.0 entries.  If you allocate a dense matrix and store all the
> zeros,
> > > > > you're just being inefficient relative to a dense format.
> > > > >
> > > > >
> > > >
> > > >
> > >
> > >
> >
> >
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20160221/47f61f29/attachment.html>