[petsc-users] Convergence problem in solving a symmetric positive definite matrix in a CFD code
Matthew Knepley
knepley at gmail.com
Thu Dec 12 20:12:30 CST 2013
On Thu, Dec 12, 2013 at 3:09 PM, Xiao, Jianjun (IKET)
<jianjun.xiao at kit.edu>wrote:
> Hi Matt,
>
> I am using PETSc-dev because only this version supports
> MatSetValuesBlockedStencil in Fortran. I tried many combinations of KSP
> and PC. None of them worked in my case. The same thing happened to
> PETSc-3.4.
>
> Then I changed back to 3.3-P7 and 3.2-P7 without using
> MatSetValuesBlockedStencil. I tried KSPCR+PCBJACOBI and many others, they
> worked fine.
>
> I read the changes of each version, and still could not find the reason.
> Did I miss something? Thank you.
>
1) "Did not work" is an insufficient description of the problem. If it did
not converge, give the output of -ksp_monitor_true_residual
-ksp_converged_reason
for each run
2) You changed two things at once. That makes it impossible to debug. Use
petsc-dev without MatSetValuesBlockedStencil(). Do you
reproduce the results from 3.3? If so, I believe your blocks may be
ordered incorrectly.
Matt
> ________________________________________
> From: Matthew Knepley [knepley at gmail.com]
> Sent: Wednesday, December 04, 2013 4:19 PM
> To: Xiao, Jianjun (IKET)
> Cc: petsc-users at mcs.anl.gov
> Subject: Re: [petsc-users] Convergence problem in solving a symmetric
> positive definite matrix in a CFD code
>
> On Wed, Dec 4, 2013 at 8:59 AM, Xiao, Jianjun (IKET) <jianjun.xiao at kit.edu
> <mailto:jianjun.xiao at kit.edu>> wrote:
> Hello,
>
> I am using Petsc to solve the linear equation in a CFD code. The matrix is
> symmetric positive definite.
>
> Please find my input and output below.
>
> 1. When I used KSPCR solver, ||r(i)||/||b|| is inf. What is the reason?
> Does it mean ||b|| is zero? When I used the KSPLGMRES solver, it seems
> ||r(i)||/||b|| is OK. However, it seems the calculated results are not
> right.
>
> It looks that way. You can check it with VecNorm() before calling
> KSPSolve().
>
> Matt
>
> 2. I am not sure if I set the solver and matrix properly. Did I miss
> something?
>
> Thank you.
>
>
> ****************************************
> INPUT:
>
> CALL KSPCreate(PETSC_COMM_WORLD,solver,ierr)
> CALL DMSetMatType(da_ksp,MATMPISBAIJ,ierr)
> CALL DMCreateMatrix(da_ksp,mat,ierr)
> CALL
> MatMPISBAIJSetPreallocation(mat,1,4,PETSC_NULL_INTEGER,1,PETSC_NULL_INTEGER,ierr)
> CALL MatZeroEntries(mat,ierr)
> CALL MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY,ierr)
> CALL MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY,ierr)
>
> CALL KSPSetOperators(solver,mat,mat,SAME_NONZERO_PATTERN,ierr)
>
> ! CALL KSPSetType(solver,KSPCR,ierr)
> CALL KSPSetType(solver,KSPLGMRES,ierr)
>
> epsi = 1.0-e5
> CALL KSPSetTolerances(solver,epsi,PETSC_DEFAULT_DOUBLE_PRECISION,&
> & PETSC_DEFAULT_DOUBLE_PRECISION,itmax,ierr)
>
> CALL KSPGetPC(solver,gfprec,ierr)
> CALL PCSetType(prec,PCBJACOBI,ierr)
>
> CALL KSPMonitorSet(solver,KSPMonitorTrueResidualNorm,&
> PETSC_NULL_OBJECT,PETSC_NULL_FUNCTION,ierr )
>
> CALL KSPSetInitialGuessNonzero(solver,PETSC_TRUE,ierr)
> CALL KSPSolve(solver,bvec,xsol,ierr)
>
> CALL KSPGetIterationNumber(solver,iter,ierr)
> CALL KSPGetResidualNorm(solver,dmax,ierr)
> CALL KSPView(solver,PETSC_VIEWER_STDOUT_WORLD,ierr)
>
> **************************************
> OUTPUT CALL KSPSetType(solver,KSPCR,ierr):
>
> 0 KSP preconditioned resid norm 2.226482634319e+05 true resid norm
> 1.204978940624e+07 ||r(i)||/||b|| inf
> 1 KSP preconditioned resid norm 1.684243557244e+05 true resid norm
> 6.742321430949e+06 ||r(i)||/||b|| inf
> 2 KSP preconditioned resid norm 1.039386033131e+05 true resid norm
> 5.094347016880e+06 ||r(i)||/||b|| inf
> 3 KSP preconditioned resid norm 3.767761162917e+04 true resid norm
> 2.085014289432e+06 ||r(i)||/||b|| inf
> 4 KSP preconditioned resid norm 2.220316358489e+04 true resid norm
> 1.039841616110e+06 ||r(i)||/||b|| inf
> 5 KSP preconditioned resid norm 1.009108756815e+04 true resid norm
> 6.764592620620e+05 ||r(i)||/||b|| inf
> 6 KSP preconditioned resid norm 7.266143334386e+03 true resid norm
> 4.713756053613e+05 ||r(i)||/||b|| inf
> 7 KSP preconditioned resid norm 4.925270026573e+03 true resid norm
> 3.276759177651e+05 ||r(i)||/||b|| inf
> 8 KSP preconditioned resid norm 2.595039666791e+03 true resid norm
> 1.774916597474e+05 ||r(i)||/||b|| inf
> 9 KSP preconditioned resid norm 1.970388137453e+03 true resid norm
> 1.449811653036e+05 ||r(i)||/||b|| inf
> 10 KSP preconditioned resid norm 1.455914234388e+03 true resid norm
> 7.916294162841e+04 ||r(i)||/||b|| inf
> 11 KSP preconditioned resid norm 8.335194818556e+02 true resid norm
> 4.530953608250e+04 ||r(i)||/||b|| inf
> 12 KSP preconditioned resid norm 3.022320555777e+02 true resid norm
> 1.728551635678e+04 ||r(i)||/||b|| inf
> 13 KSP preconditioned resid norm 7.190336024797e+01 true resid norm
> 4.186842086105e+03 ||r(i)||/||b|| inf
> 14 KSP preconditioned resid norm 1.718291655675e+01 true resid norm
> 1.089751055004e+03 ||r(i)||/||b|| inf
> 15 KSP preconditioned resid norm 1.150683059424e+01 true resid norm
> 8.672405273471e+02 ||r(i)||/||b|| inf
> 16 KSP preconditioned resid norm 8.663479440949e+00 true resid norm
> 5.776737380768e+02 ||r(i)||/||b|| inf
> 17 KSP preconditioned resid norm 5.282161990683e+00 true resid norm
> 2.977735906695e+02 ||r(i)||/||b|| inf
> 18 KSP preconditioned resid norm 3.802629315725e+00 true resid norm
> 2.789114564993e+02 ||r(i)||/||b|| inf
> 19 KSP preconditioned resid norm 1.722575171383e+00 true resid norm
> 1.051323829526e+02 ||r(i)||/||b|| inf
> KSP Object: 1 MPI processes
> type: cr
> maximum iterations=1000
> tolerances: relative=1e-05, absolute=1e-50, divergence=10000
> left preconditioning
> using nonzero initial guess
> using PRECONDITIONED norm type for convergence test
> PC Object: 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: (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: (sub_) 1 MPI processes
> type: icc
> 0 levels of fill
> tolerance for zero pivot 2.22045e-14
> using Manteuffel shift [POSITIVE_DEFINITE]
> matrix ordering: natural
> factor fill ratio given 1, needed 1
> Factored matrix follows:
> Mat Object: 1 MPI processes
> type: seqsbaij
> rows=27, cols=27
> package used to perform factorization: petsc
> total: nonzeros=81, allocated nonzeros=81
> total number of mallocs used during MatSetValues calls =0
> block size is 1
> linear system matrix = precond matrix:
> Mat Object: 1 MPI processes
> type: seqsbaij
> rows=27, cols=27
> total: nonzeros=81, allocated nonzeros=108
> total number of mallocs used during MatSetValues calls =0
> block size is 1
> linear system matrix = precond matrix:
> Mat Object: 1 MPI processes
> type: mpisbaij
> rows=27, cols=27
> total: nonzeros=81, allocated nonzeros=135
> total number of mallocs used during MatSetValues calls =0
> block size is 1
>
> OUTPUT: CALL KSPSetType(solver,KSPLGMRES,ierr)
>
> 0 KSP preconditioned resid norm 2.362537325084e+04 true resid norm
> 1.138584383312e+06 ||r(i)||/||b|| 1.000000000000e+00
> 1 KSP preconditioned resid norm 8.501213683423e+03 true resid norm
> 3.655528853686e+05 ||r(i)||/||b|| 3.210591070162e-01
> 2 KSP preconditioned resid norm 5.487567253725e+03 true resid norm
> 3.005741194777e+05 ||r(i)||/||b|| 2.639893220769e-01
> 3 KSP preconditioned resid norm 2.470452880657e+03 true resid norm
> 1.545469272201e+05 ||r(i)||/||b|| 1.357360328187e-01
> 4 KSP preconditioned resid norm 1.750803325456e+03 true resid norm
> 1.182312309352e+05 ||r(i)||/||b|| 1.038405520646e-01
> 5 KSP preconditioned resid norm 1.123492053552e+03 true resid norm
> 6.754319630701e+04 ||r(i)||/||b|| 5.932208213726e-02
> 6 KSP preconditioned resid norm 5.150241959277e+02 true resid norm
> 3.689413898730e+04 ||r(i)||/||b|| 3.240351749775e-02
> 7 KSP preconditioned resid norm 4.182894544871e+02 true resid norm
> 3.052196222024e+04 ||r(i)||/||b|| 2.680693909701e-02
> 8 KSP preconditioned resid norm 2.520093155629e+02 true resid norm
> 1.880976788356e+04 ||r(i)||/||b|| 1.652031079932e-02
> 9 KSP preconditioned resid norm 1.186491314806e+02 true resid norm
> 6.797080217853e+03 ||r(i)||/||b|| 5.969764136483e-03
> 10 KSP preconditioned resid norm 5.851092821372e+01 true resid norm
> 2.973659280245e+03 ||r(i)||/||b|| 2.611716201127e-03
> 11 KSP preconditioned resid norm 1.669909189055e+01 true resid norm
> 5.658829814125e+02 ||r(i)||/||b|| 4.970057465277e-04
> 12 KSP preconditioned resid norm 3.090594692756e+00 true resid norm
> 2.161527454147e+02 ||r(i)||/||b|| 1.898434130864e-04
> 13 KSP preconditioned resid norm 2.164618839184e+00 true resid norm
> 1.620745991834e+02 ||r(i)||/||b|| 1.423474637093e-04
> 14 KSP preconditioned resid norm 1.291593952428e+00 true resid norm
> 9.095542547366e+01 ||r(i)||/||b|| 7.988465923722e-05
> 15 KSP preconditioned resid norm 6.100583411632e-01 true resid norm
> 4.021646656091e+01 ||r(i)||/||b|| 3.532146334551e-05
> 16 KSP preconditioned resid norm 2.723496807925e-01 true resid norm
> 1.676660466866e+01 ||r(i)||/||b|| 1.472583403954e-05
> 17 KSP preconditioned resid norm 1.377718471538e-01 true resid norm
> 8.551854245272e+00 ||r(i)||/||b|| 7.510953400221e-06
> KSP Object: 1 MPI processes
> type: lgmres
> GMRES: restart=30, using Classical (unmodified) Gram-Schmidt
> Orthogonalization with no iterative refinement
> GMRES: happy breakdown tolerance 1e-30
> LGMRES: aug. dimension=2
> LGMRES: number of matvecs=170
> maximum iterations=1000
> tolerances: relative=1e-05, absolute=1e-50, divergence=10000
> left preconditioning
> using nonzero initial guess
> using PRECONDITIONED norm type for convergence test
> PC Object: 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: (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: (sub_) 1 MPI processes
> type: icc
> 0 levels of fill
> tolerance for zero pivot 2.22045e-14
> using Manteuffel shift [POSITIVE_DEFINITE]
> matrix ordering: natural
> factor fill ratio given 1, needed 1
> Factored matrix follows:
> Mat Object: 1 MPI processes
> type: seqsbaij
> rows=27, cols=27
> package used to perform factorization: petsc
> total: nonzeros=81, allocated nonzeros=81
> total number of mallocs used during MatSetValues calls =0
> block size is 1
> linear system matrix = precond matrix:
> Mat Object: 1 MPI processes
> type: seqsbaij
> rows=27, cols=27
> total: nonzeros=81, allocated nonzeros=108
> total number of mallocs used during MatSetValues calls =0
> block size is 1
> linear system matrix = precond matrix:
> Mat Object: 1 MPI processes
> type: mpisbaij
> rows=27, cols=27
> total: nonzeros=81, allocated nonzeros=135
> total number of mallocs used during MatSetValues calls =0
>
>
>
> --
> 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
>
--
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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20131212/81fc055c/attachment-0001.html>
More information about the petsc-users
mailing list