[petsc-users] fieldsplit_0_ monitor in combination with selfp

Klaij, Christiaan C.Klaij at marin.nl
Thu Sep 4 07:26:32 CDT 2014


Sorry, here's the ksp_view. I'm expecting 

-fieldsplit_1_inner_ksp_type preonly

to set the ksp(A00) in the Schur complement only, but it seems to set it in the inv(A00) of the diagonal as well.

Chris

  0 KSP Residual norm 1.229687498638e+00 
    Residual norms for fieldsplit_1_ solve.
    0 KSP Residual norm 7.185799114488e+01 
    1 KSP Residual norm 3.873274154012e+01 
  1 KSP Residual norm 1.107969383366e+00 
KSP Object: 1 MPI processes
  type: fgmres
    GMRES: restart=30, using Classical (unmodified) Gram-Schmidt Orthogonalization with no iterative refinement
    GMRES: happy breakdown tolerance 1e-30
  maximum iterations=1, initial guess is zero
  tolerances:  relative=1e-05, absolute=1e-50, divergence=10000
  right preconditioning
  using UNPRECONDITIONED norm type for convergence test
PC Object: 1 MPI processes
  type: fieldsplit
    FieldSplit with Schur preconditioner, factorization LOWER
    Preconditioner for the Schur complement formed from Sp, an assembled approximation to S, which uses (the lumped) A00's diagonal's inverse
    Split info:
    Split number 0 Defined by IS
    Split number 1 Defined by IS
    KSP solver for A00 block
      KSP Object:      (fieldsplit_0_)       1 MPI processes
        type: preonly
        maximum iterations=1, initial guess is zero
        tolerances:  relative=1e-05, absolute=1e-50, divergence=10000
        left preconditioning
        using NONE norm type for convergence test
      PC Object:      (fieldsplit_0_)       1 MPI processes
        type: bjacobi
          block Jacobi: number of blocks = 1
          Local solve is same for all blocks, in the following KSP and PC objects:
          KSP Object:          (fieldsplit_0_sub_)           1 MPI processes
            type: preonly
            maximum iterations=10000, initial guess is zero
            tolerances:  relative=1e-05, absolute=1e-50, divergence=10000
            left preconditioning
            using NONE norm type for convergence test
          PC Object:          (fieldsplit_0_sub_)           1 MPI processes
            type: ilu
              ILU: out-of-place factorization
              0 levels of fill
              tolerance for zero pivot 2.22045e-14
              using diagonal shift on blocks to prevent zero pivot [INBLOCKS]
              matrix ordering: natural
              factor fill ratio given 1, needed 1
                Factored matrix follows:
                  Mat Object:                   1 MPI processes
                    type: seqaij
                    rows=48, cols=48
                    package used to perform factorization: petsc
                    total: nonzeros=200, allocated nonzeros=200
                    total number of mallocs used during MatSetValues calls =0
                      not using I-node routines
            linear system matrix = precond matrix:
            Mat Object:            (fieldsplit_0_)             1 MPI processes
              type: seqaij
              rows=48, cols=48
              total: nonzeros=200, allocated nonzeros=240
              total number of mallocs used during MatSetValues calls =0
                not using I-node routines
        linear system matrix = precond matrix:
        Mat Object:        (fieldsplit_0_)         1 MPI processes
          type: mpiaij
          rows=48, cols=48
          total: nonzeros=200, allocated nonzeros=480
          total number of mallocs used during MatSetValues calls =0
            not using I-node (on process 0) routines
    KSP solver for S = A11 - A10 inv(A00) A01 
      KSP Object:      (fieldsplit_1_)       1 MPI processes
        type: gmres
          GMRES: restart=30, using Classical (unmodified) Gram-Schmidt Orthogonalization with no iterative refinement
          GMRES: happy breakdown tolerance 1e-30
        maximum iterations=1, initial guess is zero
        tolerances:  relative=1e-05, absolute=1e-50, divergence=10000
        left preconditioning
        using PRECONDITIONED norm type for convergence test
      PC Object:      (fieldsplit_1_)       1 MPI processes
        type: bjacobi
          block Jacobi: number of blocks = 1
          Local solve is same for all blocks, in the following KSP and PC objects:
          KSP Object:          (fieldsplit_1_sub_)           1 MPI processes
            type: preonly
            maximum iterations=10000, initial guess is zero
            tolerances:  relative=1e-05, absolute=1e-50, divergence=10000
            left preconditioning
            using NONE norm type for convergence test
          PC Object:          (fieldsplit_1_sub_)           1 MPI processes
            type: bjacobi
              block Jacobi: number of blocks = 1
              Local solve is same for all blocks, in the following KSP and PC objects:
              KSP Object:              (fieldsplit_1_sub_sub_)               1 MPI processes
                type: preonly
                maximum iterations=10000, initial guess is zero
                tolerances:  relative=1e-05, absolute=1e-50, divergence=10000
                left preconditioning
                using NONE norm type for convergence test
              PC Object:              (fieldsplit_1_sub_sub_)               1 MPI processes
                type: ilu
                  ILU: out-of-place factorization
                  0 levels of fill
                  tolerance for zero pivot 2.22045e-14
                  using diagonal shift on blocks to prevent zero pivot [INBLOCKS]
                  matrix ordering: natural
                  factor fill ratio given 1, needed 1
                    Factored matrix follows:
                      Mat Object:                       1 MPI processes
                        type: seqaij
                        rows=24, cols=24
                        package used to perform factorization: petsc
                        total: nonzeros=120, allocated nonzeros=120
                        total number of mallocs used during MatSetValues calls =0
                          not using I-node routines
                linear system matrix = precond matrix:
                Mat Object:                 1 MPI processes
                  type: seqaij
                  rows=24, cols=24
                  total: nonzeros=120, allocated nonzeros=120
                  total number of mallocs used during MatSetValues calls =0
                    not using I-node routines
            linear system matrix = precond matrix:
            Mat Object:             1 MPI processes
              type: mpiaij
              rows=24, cols=24
              total: nonzeros=120, allocated nonzeros=120
              total number of mallocs used during MatSetValues calls =0
                not using I-node (on process 0) routines
        linear system matrix followed by preconditioner matrix:
        Mat Object:        (fieldsplit_1_)         1 MPI processes
          type: schurcomplement
          rows=24, cols=24
            Schur complement A11 - A10 inv(A00) A01
            A11
              Mat Object:              (fieldsplit_1_)               1 MPI processes
                type: mpiaij
                rows=24, cols=24
                total: nonzeros=0, allocated nonzeros=0
                total number of mallocs used during MatSetValues calls =0
                  using I-node (on process 0) routines: found 5 nodes, limit used is 5
            A10
              Mat Object:              (a10_)               1 MPI processes
                type: mpiaij
                rows=24, cols=48
                total: nonzeros=96, allocated nonzeros=96
                total number of mallocs used during MatSetValues calls =0
                  not using I-node (on process 0) routines
            KSP of A00
              KSP Object:              (fieldsplit_1_inner_)               1 MPI processes
                type: preonly
                maximum iterations=1, initial guess is zero
                tolerances:  relative=1e-05, absolute=1e-50, divergence=10000
                left preconditioning
                using NONE norm type for convergence test
              PC Object:              (fieldsplit_1_inner_)               1 MPI processes
                type: jacobi
                linear system matrix = precond matrix:
                Mat Object:                (fieldsplit_0_)                 1 MPI processes
                  type: mpiaij
                  rows=48, cols=48
                  total: nonzeros=200, allocated nonzeros=480
                  total number of mallocs used during MatSetValues calls =0
                    not using I-node (on process 0) routines
            A01
              Mat Object:              (a01_)               1 MPI processes
                type: mpiaij
                rows=48, cols=24
                total: nonzeros=96, allocated nonzeros=480
                total number of mallocs used during MatSetValues calls =0
                  not using I-node (on process 0) routines
        Mat Object:         1 MPI processes
          type: mpiaij
          rows=24, cols=24
          total: nonzeros=120, allocated nonzeros=120
          total number of mallocs used during MatSetValues calls =0
            not using I-node (on process 0) routines
  linear system matrix = precond matrix:
  Mat Object:   1 MPI processes
    type: nest
    rows=72, cols=72
      Matrix object: 
        type=nest, rows=2, cols=2 
        MatNest structure: 
        (0,0) : prefix="fieldsplit_0_", type=mpiaij, rows=48, cols=48 
        (0,1) : prefix="a01_", type=mpiaij, rows=48, cols=24 
        (1,0) : prefix="a10_", type=mpiaij, rows=24, cols=48 
        (1,1) : prefix="fieldsplit_1_", type=mpiaij, rows=24, cols=24


From: Matthew Knepley <knepley at gmail.com>
Sent: Thursday, September 04, 2014 2:20 PM
To: Klaij, Christiaan
Cc: petsc-users at mcs.anl.gov
Subject: Re: [petsc-users] fieldsplit_0_ monitor in combination with selfp
  



On Thu, Sep 4, 2014 at 7:06 AM, Klaij, Christiaan  <C.Klaij at marin.nl> wrote:
 I'm playing with the selfp option in fieldsplit using
snes/examples/tutorials/ex70.c. For example:

mpiexec -n 2 ./ex70 -nx 4 -ny 6 \
-ksp_type fgmres \
-pc_type fieldsplit \
-pc_fieldsplit_type schur \
-pc_fieldsplit_schur_fact_type lower \
-pc_fieldsplit_schur_precondition selfp \
-fieldsplit_1_inner_ksp_type preonly \
-fieldsplit_1_inner_pc_type jacobi \
-fieldsplit_0_ksp_monitor -fieldsplit_0_ksp_max_it 1 \
-fieldsplit_1_ksp_monitor -fieldsplit_1_ksp_max_it 1 \
-ksp_monitor -ksp_max_it 1

gives the following output

  0 KSP Residual norm 1.229687498638e+00
    Residual norms for fieldsplit_1_ solve.
    0 KSP Residual norm 2.330138480101e+01
    1 KSP Residual norm 1.609000846751e+01
  1 KSP Residual norm 1.180287268335e+00

To my suprise I don't see anything for the fieldsplit_0_ solve,
why?



Always run with -ksp_view for any solver question.


  Thanks,


    Matt
   Furthermore, if I understand correctly the above should be
exactly equivalent with

mpiexec -n 2 ./ex70 -nx 4 -ny 6 \
-ksp_type fgmres \
-pc_type fieldsplit \
-pc_fieldsplit_type schur \
-pc_fieldsplit_schur_fact_type lower \
-user_ksp \
-fieldsplit_0_ksp_monitor -fieldsplit_0_ksp_max_it 1 \
-fieldsplit_1_ksp_monitor -fieldsplit_1_ksp_max_it 1 \
-ksp_monitor -ksp_max_it 1

  0 KSP Residual norm 1.229687498638e+00
    Residual norms for fieldsplit_0_ solve.
    0 KSP Residual norm 5.486639587672e-01
    1 KSP Residual norm 6.348354253703e-02
    Residual norms for fieldsplit_1_ solve.
    0 KSP Residual norm 2.321938107977e+01
    1 KSP Residual norm 1.605484031258e+01
  1 KSP Residual norm 1.183225251166e+00

because -user_ksp replaces the Schur complement by the simple
approximation A11 - A10 inv(diag(A00)) A01. Beside the missing
fielsplit_0_ part, the numbers are pretty close but not exactly
the same. Any explanation?

Chris


dr. ir. Christiaan Klaij
CFD Researcher
Research & Development
E mailto:C.Klaij at marin.nl
T +31 317 49 33 44


MARIN
2, Haagsteeg, P.O. Box 28, 6700 AA Wageningen, The Netherlands
T +31 317 49 39 11, F +31 317 49 32 45, I www.marin.nl

 



 -- 
What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
-- Norbert Wiener      


More information about the petsc-users mailing list