[petsc-dev] METIS update -> MUMPS crash

[Patrick Lacasse]

Hi, I think I have this problem the same problem. It crash with this stask :

#004:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libmetis.so(libmetis__CreateCoarseGraphNoMask+0x2e1)
[0x7f349896db71]
#005:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libmetis.so(libmetis__Match_SHEM+0x5a3)
[0x7f349896f233]
#006:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libmetis.so(CoarsenGraphNlevels+0x1e5)
[0x7f349896f955]
#007:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libmetis.so(libmetis__MlevelNodeBisectionL2+0x50)
[0x7f34989669c0]
#008:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libmetis.so(libmetis__MlevelNestedDissection+0x1e)
[0x7f3498967d8e]
#009:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libmetis.so(METIS_NodeND+0x415)
[0x7f34989684c5]
#010:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so(dmumps_195_+0x3987)
[0x7f3499cfc1b9]
#011:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so(dmumps_26_+0x1488)
[0x7f3499c3de1f]
#012:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so(dmumps_+0x1032)
[0x7f3499cce3e3]
#013:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so(dmumps_f77_+0x1030)
[0x7f3499c1a08f]
#014:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so(dmumps_c+0xcc4)
[0x7f3499bfed38]
#015:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so(MatCholeskyFactorSymbolic_MUMPS+0x58b)
[0x7f34998a10cd]
#016:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so(MatCholeskyFactorSymbolic+0x2c3)
[0x7f349961ab70]
#017:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so(+0x645934)
[0x7f34999f5934]
#018:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so(PCSetUp+0x39a)
[0x7f3499a78232]
#019:
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so(KSPSetUp+0xd61)
[0x7f3499ad8e16]

alternatively, I can try a LU factorisation and then it hang into metis at :

#0  0x00007fea74ef5ada in libmetis__rpqUpdate () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libmetis.so
#1  0x00007fea74ee7c1f in libmetis__FM_2WayNodeRefine1Sided () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libmetis.so
#2  0x00007fea74ecdcae in libmetis__Refine2WayNode () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libmetis.so
#3  0x00007fea74ed3d8e in libmetis__MlevelNestedDissection () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libmetis.so
#4  0x00007fea74ed44c5 in METIS_NodeND () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libmetis.so
#5  0x00007fea762681b9 in dmumps_195_ () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so
#6  0x00007fea761a9e1f in dmumps_26_ () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so
#7  0x00007fea7623a3e3 in dmumps_ () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so
#8  0x00007fea7618608f in dmumps_f77_ () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so
#9  0x00007fea7616ad38 in dmumps_c () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so
#10 0x00007fea75e0c458 in MatLUFactorSymbolic_AIJMUMPS () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so
#11 0x00007fea75b86793 in MatLUFactorSymbolic () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so
#12 0x00007fea75f6678d in PCSetUp_LU () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so
#13 0x00007fea75fe4232 in PCSetUp () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so
#14 0x00007fea76044e16 in KSPSetUp () from
/home/mefpp_placasse/hpclibs/petsc-dev_arch-gcc64-opt_Patlac/lib/libpetsc.so

This only happen for big matrix and, for the same problem, not on every
computer.

I have tried scotch and it crash like this :
http://graal.ens-lyon.fr/MUMPS/index.php?page=faq#19
   Unfortunately, there is a an error in the SCOTCH 6.0 package which is
making it unusable with MUMPS. The SCOTCH developement team is aware of the
problem and is working on it. For the moment, you should use the previous
version of SCOTCH (5.1.12b) which works fine.

I have tried cholmod and it gave me very slow results (and it don't want do
to cholesky on indefinite symmetric matrix).

For now I'd like to know : if I change my version of metis, which version
should I choose?
Maybe I also need to choose appropriate parmetis.

Patrick Lacasse


2013/9/18 Sean Farley <sean.michael.farley at gmail.com>

>
> bsmith at mcs.anl.gov writes:
>
> > On Sep 1, 2013, at 1:27 PM, Matthew Knepley <knepley at gmail.com> wrote:
> >
> >> On Sun, Sep 1, 2013 at 1:07 PM, Satish Balay <balay at mcs.anl.gov> wrote:
> >> I see the errors with valgrind - and I don't know the reason.  Perhaps
> >> we should revert the metis/parmetis upgrade.. [unless someone can debug
> this..]
> >>
> >> valgrind doesn't give errors with the older metis-5.0.2
> >>
> >> This looks like a MUMPS bug. I guess we should report to them that they
> are not
> >> actually compatible with the latest METIS release.
> >
> >    They may not claim to be?   We should be careful about not casually
> upgrading our support for one package if other packages we use may not work
> with that upgrade. How to ensure this in practice I don't know.
>
> Sorry for the late chime-in; just now catching up on petsc-dev. MUMPS
> doesn't claim to use anything considered recent in terms of METIS. As
> for as I can see, they still haven't applied my patch to update MUMPS to
> work with METIS 5.0:
>
> https://listes.ens-lyon.fr/sympa/arc/mumps-users/2011-11/msg00027.html
>
> The same goes (as of this afternoon) about George not updating METIS
> with the patches we sent a while back.
>



2013/9/18 Sean Farley <sean.michael.farley at gmail.com>

>
> bsmith at mcs.anl.gov writes:
>
> > On Sep 1, 2013, at 1:27 PM, Matthew Knepley <knepley at gmail.com> wrote:
> >
> >> On Sun, Sep 1, 2013 at 1:07 PM, Satish Balay <balay at mcs.anl.gov> wrote:
> >> I see the errors with valgrind - and I don't know the reason.  Perhaps
> >> we should revert the metis/parmetis upgrade.. [unless someone can debug
> this..]
> >>
> >> valgrind doesn't give errors with the older metis-5.0.2
> >>
> >> This looks like a MUMPS bug. I guess we should report to them that they
> are not
> >> actually compatible with the latest METIS release.
> >
> >    They may not claim to be?   We should be careful about not casually
> upgrading our support for one package if other packages we use may not work
> with that upgrade. How to ensure this in practice I don't know.
>
> Sorry for the late chime-in; just now catching up on petsc-dev. MUMPS
> doesn't claim to use anything considered recent in terms of METIS. As
> for as I can see, they still haven't applied my patch to update MUMPS to
> work with METIS 5.0:
>
> https://listes.ens-lyon.fr/sympa/arc/mumps-users/2011-11/msg00027.html
>
> The same goes (as of this afternoon) about George not updating METIS
> with the patches we sent a while back.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20131002/52b1f8de/attachment.html>