[petsc-users] Parallel Incomplete Choleski Factorization

[Michele Rosso]

On 07/17/2012 11:03 AM, Hong Zhang wrote:
> Michele :
>
>
>     I have some problems with the block jacobi preconditioner.
>     I am solving a 3D  Poisson equation with periodic BCs, discretized
>     by using finite differences (7-points stencil).
>     Thus the problem is singular and the nullspace has to be removed.
>
>
> For Poisson equations, multigrid precondition should be the method of
> choice.
Thank you for the suggestion. I do not have any experience with 
multigrid, but I will try.
>
>     If I solve with the PCG method + JACOBI preconditioner the results
>     are fine.
>     If I use PCG + Block Jacobi preconditioner + ICC on each block the
>     results are fine on the majority of the processors,
>     but on few of them the error is very large.
>
> How do you know " few of them"?
  Basically the solution is not correct on some grid points, say 6 grid 
nodes out of 64^3. The 6 grid nodes with problems belongs to 2 of the 
128 processors
I am using.
>
>     Do you have any idea/suggestions on how to fix this problem?
>     This is the fragment of code I am using ( petsc 3.1 and Fortran 90):
>
> Please update to petsc-3.3. petsc-3.1 is too old.
     I would do that but the version installed on the platform (Intrepid 
at ALCF) I am working on is 3.1-p2.

>
>         PetscErrorCode                petsc_err
>         Mat                                         A
>         PC                                           pc, subpc
>         KSP                                         ksp
>         KSP                                         subksp(1)
>         :
>         :
>         :
>         call KSPCreate(PETSC_COMM_WORLD,ksp,petsc_err)
>         call KSPSetOperators(ksp,A,A,DIFFERENT_NONZERO_PATTERN,petsc_err)
>
> call KSPSetType(ksp,KSPCG, ) !the default type is gmres. I guess you 
> want CG
>
>     call KSPGetPC(ksp,pc,petsc_err)
>         call PCSetType(pc,PCBJACOBI,petsc_err)
>     !    call KSPSetUp(ksp,petsc_err)    call this at the end
>
>         ! KSP context for each single block
>         call
>     PCBJacobiGetSubKSP(pc,PETSC_NULL_INTEGER,PETSC_NULL_INTEGER,subksp(1),petsc_err)
>
>         call KSPGetPC(subksp(1),subpc,petsc_err)
>         call PCSetType(subpc,PCICC,petsc_err)
>
>         call KSPSetType(subksp(1),KSPPREONLY petsc_err)
>
>         call KSPSetTolerances(subksp(1),tol
>     ,PETSC_DEFAULT_DOUBLE_PRECISION,&
>              &
>     PETSC_DEFAULT_DOUBLE_PRECISION,PETSC_DEFAULT_INTEGER,petsc_err)
>
>        ! Remove nullspace from the singular system (Check PETSC_NULL)
>         call
>     MatNullSpaceCreate(MPI_COMM_WORLD,PETSC_TRUE,0,PETSC_NULL,nullspace,petsc_err)
>         call KSPSetNullSpace(ksp, nullspace, petsc_err)
>         call MatNullSpaceRemove(nullspace, b, PETSC_NULL,petsc_err)
>
>         call KSPSolve(ksp,b,x,petsc_err)
>
>
> I modified your code slightly. All these options can be provided at 
> runtime:
> '-ksp_type cg -pc_type bjacobi -sub_pc_type icc'
>
> Hong
>
>
>
>
>
>
>
>
>
>
>     On 07/13/2012 12:14 PM, Hong Zhang wrote:
>>     Michele :
>>
>>
>>         I need to use the ICC factorization as preconditioner, but I
>>         noticed that no parallel version is supported.
>>         Is that correct?
>>
>>     Correct.
>>
>>         If so, is there a work around, like building  the
>>         preconditioner "by hand" by using PETSc functions?
>>
>>     You may try block jacobi with icc in the blocks  '-ksp_type cg
>>     -pc_type bjacobi -sub_pc_type icc'
>>
>>     Hong
>>
>>
>
>
>


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120717/077b6027/attachment.html>