[petsc-users] Preconditioned JFNK using SuperLU_dist for time dependent problems

[Barry Smith]

On Jul 29, 2011, at 11:27 AM, Travis C. Fisher wrote:

> Hong,
> 
> In petsc-dev, superlu_dist now fails at the first attempt to apply the preconditioner in the first linear solve. In the release version, it worked fine for the first time step (that's five nonlinear solves using the same preconditioner), but then failed on the second time step (which I suppose is more disconcerting). I have tried on 1 and 4 processors and get the same problem in the development version.

   Are you absolutely sure the non-zero pattern is fixed and doesn't change? After the call to MatAssemblyEnd() make a call MatSetOption(mat, MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);  Then rerun.

   Also it is much more useful if you include the entire error message (cut and paste) instead just a few words, there is lots of valuable information in the error message that can help other people see what the solution might be.

   Barry

> 
> To clarify, when I mentioned ASM/ILU before, the ILU routine I was using was the default PETSc routine. That still works with my implementation using petsc-dev.
> 
> Mumps worked fine with petsc-dev, but I don't think it's as fast as superlu_dist for the test problem I am using.
> 
> Thanks,
> 
> Travis