[petsc-users] Convergence problem in solving a symmetric positive definite matrix in a CFD code

[Matthew Knepley]

On Fri, Dec 13, 2013 at 4:16 AM, Xiao, Jianjun (IKET)
<jianjun.xiao at kit.edu>wrote:

> Hello Jed and Matt,
>
> Thanks for your reply.
>
> I design a small 3D case with 10 cells in each direction. Please find the
> binary files generated with PETSc-dev in the attachment. The matrix should
> be symmetric and positive definite.
>
> The calculation did not converge at the first time step with PCR+JACOBI.
> With PCG+JACOBI, it seems it converges in this case. I have very limited
> experience in using PETSc. In the past, I felt more flexible to change the
> combination of linear solvers and preconditioners. I could get divergence
> anyhow, and the main difference was the speed. But now, it seems only
> PCG+JACOBI works fine for my case. Many other combinations, such as
> PCR+JACOBI, PCR+BJACOBI,  PCG+BJACOBI, divergent at the first time step or
> take much longer time. I think there must be something wrong.
>

Can you see that the above paragraph gives NO IDEA what your problem is?

1) You completely failed to do the one thing I asked you to do. That is not
the way to get help.

     *** Run the same solver with 3.3 and dev. Does it give the same answer
***

    Until you do this, why would I believe anything has changed.

2) Iterative solvers fail to converge all the time. That is routine, and
the reason for preconditioning. Preconditioners
     are always tied to the specific problem. Thus telling us that
something did not converge without outlining the
     problem, or providing the output I asked for
(-ksp_monitor_true_residual -ksp_converged_reason) tells us
     nothing.

   Matt


> Below is how I set up the matrix and the solver.
>
>        CALL DMSetMatType(da_ksp,MATMPISBAIJ,ierr)
>        CALL DMCreateMatrix(da_ksp,gmat,ierr)
>        CALL DMSetMatrixPreallocateOnly(da_ksp,PETSC_TRUE,ierr)
>        CALL
> MatMPISBAIJSetPreallocation(gmat,1,4,PETSC_NULL_INTEGER,3,PETSC_NULL_INTEGER,ierr)
>        CALL MatZeroEntries(gmat,ierr)
>
>        CALL KSPCreate(PETSC_COMM_WORLD,gsolver,ierr)
> !       CALL KSPSetType(gsolver,KSPCR,ierr)
>        CALL KSPSetType(gsolver,KSPCG,ierr)
>
>        CALL KSPGetPC(gsolver,gprec,ierr)
>        CALL PCSetType(gprec,PCBJACOBI,ierr)
> !       CALL PCSetType(gprec,PCBJACOBI,ierr)
>
> ********************************** matsetvalues
>        ...
>            IF (k < kbar-1 .AND. j < jbar-1 .AND. i < ibar-1) THEN ! d-x-y-z
>
>              valpm(1)     = value_ad_gfpp(m)
>              colpm(MatStencil_i,1) = i
>              colpm(MatStencil_j,1) = j
>              colpm(MatStencil_k,1) = k
>
>              valpm(2)     = value_ax_gfpp(m)
>              colpm(MatStencil_i,2) = i+1
>              colpm(MatStencil_j,2) = j
>              colpm(MatStencil_k,2) = k
>
>              valpm(3)     = value_ay_gfpp(m)
>              colpm(MatStencil_i,3) = i
>              colpm(MatStencil_j,3) = j+1
>              colpm(MatStencil_k,3) = k
>
>              valpm(4)     = value_az_gfpp(m)
>              colpm(MatStencil_i,4) = i
>              colpm(MatStencil_j,4) = j
>              colpm(MatStencil_k,4) = k+1
>
>              CALL
> MatSetValuesBlockedStencil(gmat,1,rowpm,4,colpm,valpm,INSERT_VALUES,ierr)
>        ...
>
>        ...
> **********************************
>        CALL MatAssemblyBegin(gmat,MAT_FINAL_ASSEMBLY,ierr)
>        CALL MatAssemblyEnd(gmat,MAT_FINAL_ASSEMBLY,ierr)
>
>        CALL KSPSetOperators(gsolver,gmat,gmat,SAME_NONZERO_PATTERN,ierr)
>
>        epsi_rtol = 1.0e-50_dble
>        epsi_atol = 1.0e-5_dble
>
>        CALL KSPSetTolerances(gsolver,epsi_rtol, epsi_atol,&
>            & PETSC_DEFAULT_DOUBLE_PRECISION,itmax,ierr)
>
>        CALL KSPSetNormType(gsolver,KSP_NORM_PRECONDITIONED,ierr)
>
>        CALL KSPSolve(gsolver,bvec0_gfpp,xsol0_gfpp,ierr)
>
> Thank you for your help.
>
> JJ
>
>
> ________________________________________
> From: Jed Brown [jedbrown at mcs.anl.gov]
> Sent: Friday, December 13, 2013 8:34 AM
> To: Xiao, Jianjun (IKET); Matthew Knepley
> Cc: petsc-users at mcs.anl.gov
> Subject: Re: [petsc-users] Convergence problem in solving a symmetric
> positive definite matrix in a CFD code
>
> "Xiao, Jianjun (IKET)" <jianjun.xiao at kit.edu> writes:
>
> > 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.
>
> We're going to need exact options and output for the cases you tried.
> You should be able to use the same algorithm with both versions (with
> suitable options).  And if you have a modest-sized problem, you can run
> with -ksp_view_mat and send binaryoutput.  We'll use
> src/ksp/ksp/examples/tutorials/ex10.c to reproduce the convergence you
> see.
>



-- 
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/20131213/d597722b/attachment.html>