[petsc-users] Automatically re-solving after MUMPS error

[Barry Smith]

  Matt,

   Yes, you must be right The MatDestroy() on the "partially factored" matrix should clean up everything properly but it sounds like it is not. I'll look at it right now but I only have a few minutes; if I can't resolve it really quickly it may take a day or two.


  Barry

> On Sep 30, 2015, at 4:10 PM, Matt Landreman <matt.landreman at gmail.com> wrote:
> 
> Hi Barry,
> I tried adding PetscMallocDump after SNESDestroy as you suggested. When mumps fails, PetscMallocDump shows a number of mallocs which are absent when mumps succeeds, the largest being MatConvertToTriples_mpiaij_mpiaij() (line 638 in petsc-3.6.0/src/mat/impls/aij/mpi/mumps/mumps.c).  The total memory reported by PetscMallocDump after SNESDestroy is substantially (>20x) larger when mumps fails than when mumps succeeds, and this amount increases uniformly with each mumps failure.  So I think some of the mumps-related structures are not being deallocated by SNESDestroy if mumps generates an error.
> Thanks,
> -Matt
> 
> On Wed, Sep 30, 2015 at 2:16 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
> 
> > On Sep 30, 2015, at 1:06 PM, Matt Landreman <matt.landreman at gmail.com> wrote:
> >
> > PETSc developers,
> >
> > I tried implementing a system for automatically increasing MUMPS ICNTL(14), along the lines described in this recent thread. If SNESSolve returns ierr .ne. 0 due to MUMPS error -9, I call SNESDestroy, re-initialize SNES, call MatMumpsSetIcntl with a larger value of ICNTL(14), call SNESSolve again, and repeat as needed. The procedure works, but the peak memory required (as measured by the HPC system) is 50%-100% higher if the MUMPS solve has to be repeated compared to when MUMPS works on the 1st try (by starting with a large ICNTL(14)), even though SNESDestroy is called in between the attempts. Are there some PETSc or MUMPS structures which would not be deallocated immediately by SNESDestroy?  If so, how do I deallocate them?
> 
>    They should be all destroyed automatically for you. You can use PetscMallocDump() after the SNES is destroyed to verify that all that memory is not properly freed.
> 
>    My guess is that your new malloc() with the bigger workspace cannot "reuse" the space that was previously freed; so to the OS it looks like you are using a lot more space but in terms of physical memory you are not using more.
> 
>   Barry
> 
> >
> > Thanks,
> > Matt Landreman
> >
> >
> > On Tue, Sep 15, 2015 at 7:47 AM, David Knezevic <david.knezevic at akselos.com> wrote:
> > On Tue, Sep 15, 2015 at 7:29 PM, Matthew Knepley <knepley at gmail.com> wrote:
> > On Tue, Sep 15, 2015 at 4:30 AM, David Knezevic <david.knezevic at akselos.com> wrote:
> > In some cases, I get MUMPS error -9, i.e.:
> > [2]PETSC ERROR: Error reported by MUMPS in numerical factorization phase: INFO(1)=-9, INFO(2)=98927
> >
> > This is easily fixed by re-running the executable with -mat_mumps_icntl_14 on the commandline.
> >
> > However, I would like to update my code in order to do this automatically, i.e. detect the -9 error and re-run with the appropriate option. Is there a recommended way to do this? It seems to me that I could do this with a PETSc error handler (e.g. PetscPushErrorHandler) in order to call a function that sets the appropriate option and solves again, is that right? Are there any examples that illustrate this type of thing?
> >
> > I would not use the error handler. I would just check the ierr return code from the solver. I think you need the
> > INFO output, for which you can use MatMumpsGetInfo().
> >
> >
> > OK, that sounds good (and much simpler than what I had in mind), thanks for the help!
> >
> > David
> >
> >
> 
>