[petsc-users] positive definite MPISBAIJ-Matrix gets singular in the case of more than one MPI-process

Andreas Hauffe andreas.hauffe at tu-dresden.de
Tue Jun 8 02:33:47 CDT 2010


Dear all,

I want to solve a linear System with a symmetric positive definite matrix. I 
use the external package MUMPS as solver with the following commands 

 call KSPCreate(PETSC_COMM_WORLD,ksp,ierr)
 call KSPSetOperators(ksp,mat,mat,DIFFERENT_NONZERO_PATTERN,ierr)
 call KSPSetType(ksp,KSPPREONLY,ierr)
 call KSPGetPC(ksp,prec,ierr)
 call PCSetType(prec,PCCHOLESKY,ierr)
 call PCFactorSetMatSolverPackage(prec,MAT_SOLVER_MUMPS,ierr)
 call KSPSetFromOptions(ksp,ierr)
 call KSPSolve(ksp,vec1,vec2,ierr)

and give the option "-mat_mumps_sym 1" (symmetric positive definite).

If I run the program with one MPI-process everything works fine. For a low 
number of unknows it also works for more MPI-processes. If I increase the 
number of unknows (>50000) and use more then one MPI-process on one or more 
host, I get a MUMPS error that the matrix is singular. With one MPI-process 
everything is still fine. The critical value for the number of unknows 
depends also on the use of the debug-version or non-debug-version of PETSc.
If I do not use the option "-mat_mumps_sym 1" everything also works.

Right now the finite element assembling process (MatSetValue) is only done 
with rank 0. So the hole matrix gets all its values only from rank 0.

Could someone explain this behaviour? Where is my error?

Btw, I am currently using PETSc 3.0 since I work with SLEPc as well.

Thank you very much,
-- 
Andreas Hauffe

----------------------------------------------------------------------------------------------------
Technische Universität Dresden
Institut für Luft- und Raumfahrttechnik / Institute of Aerospace Engineering
Lehrstuhl für Luftfahrzeugtechnik / Chair of Aircraft Engineering

D-01062 Dresden
Germany

phone   : (++49)351 463 38496
fax     : (++49)351 463 37263
mail    : andreas.hauffe at tu-dresden.de
Website : http://tu-dresden.de/mw/ilr/lft
----------------------------------------------------------------------------------------------------


More information about the petsc-users mailing list