[petsc-users] MATSOLVERSUPERLU_DIST not giving the correct solution

Barry Smith bsmith at mcs.anl.gov
Wed Apr 30 13:19:28 CDT 2014


  Please send the same thing on one process.


On Apr 30, 2014, at 8:17 AM, Justin Dong <jsd1 at rice.edu> wrote:

> Here are the results of one example where the solution is incorrect: 
> 
>   0 KSP unpreconditioned resid norm 7.267616711036e-05 true resid norm 7.267616711036e-05 ||r(i)||/||b|| 1.000000000000e+00
> 
>   1 KSP unpreconditioned resid norm 4.081398605668e-16 true resid norm 4.017029301117e-16 ||r(i)||/||b|| 5.527299334618e-12
> 
> 
>   2 KSP unpreconditioned resid norm 4.378737248697e-21 true resid norm 4.545226736905e-16 ||r(i)||/||b|| 6.254081520291e-12
> 
> KSP Object: 4 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=10000, initial guess is zero
> 
>   tolerances:  relative=1e-13, absolute=1e-50, divergence=10000
> 
>   right preconditioning
> 
>   using UNPRECONDITIONED norm type for convergence test
> 
> PC Object: 4 MPI processes
> 
>   type: lu
> 
>     LU: out-of-place factorization
> 
>     tolerance for zero pivot 2.22045e-14
> 
>     matrix ordering: natural
> 
>     factor fill ratio given 0, needed 0
> 
>       Factored matrix follows:
> 
>         Matrix Object:         4 MPI processes
> 
>           type: mpiaij
> 
>           rows=1536, cols=1536
> 
>           package used to perform factorization: superlu_dist
> 
>           total: nonzeros=0, allocated nonzeros=0
> 
>           total number of mallocs used during MatSetValues calls =0
> 
>             SuperLU_DIST run parameters:
> 
>               Process grid nprow 2 x npcol 2 
> 
>               Equilibrate matrix TRUE 
> 
>               Matrix input mode 1 
> 
>               Replace tiny pivots TRUE 
> 
>               Use iterative refinement FALSE 
> 
>               Processors in row 2 col partition 2 
> 
>               Row permutation LargeDiag 
> 
>               Column permutation METIS_AT_PLUS_A
> 
>               Parallel symbolic factorization FALSE 
> 
>               Repeated factorization SamePattern_SameRowPerm
> 
>   linear system matrix = precond matrix:
> 
>   Matrix Object:   4 MPI processes
> 
>     type: mpiaij
> 
>     rows=1536, cols=1536
> 
>     total: nonzeros=17856, allocated nonzeros=64512
> 
>     total number of mallocs used during MatSetValues calls =0
> 
>       using I-node (on process 0) routines: found 128 nodes, limit used is 5
> 
> 
> 
> On Wed, Apr 30, 2014 at 7:57 AM, Barry Smith <bsmith at mcs.anl.gov> wrote:
> 
> On Apr 30, 2014, at 6:46 AM, Matthew Knepley <knepley at gmail.com> wrote:
> 
> > On Wed, Apr 30, 2014 at 6:19 AM, Justin Dong <jsd1 at rice.edu> wrote:
> > Thanks. If I turn on the Krylov solver, the issue still seems to persist though.
> >
> > mpiexec -n 4 -ksp_type gmres -ksp_rtol 1.0e-13 -pc_type lu -pc_factor_mat_solver_package superlu_dist
> >
> > I'm testing on a very small system now (24 ndofs), but if I go larger (around 20000 ndofs) then it gets worse.
> >
> > For the small system, I exported the matrices to matlab to make sure they were being assembled correct in parallel, and I'm certain that that they are.
> >
> > For convergence questions, always run using -ksp_monitor -ksp_view so that we can see exactly what you run.
> 
>   Also run with -ksp_pc_side right
> 
> 
> >
> >   Thanks,
> >
> >      Matt
> >
> >
> > On Wed, Apr 30, 2014 at 5:32 AM, Matthew Knepley <knepley at gmail.com> wrote:
> > On Wed, Apr 30, 2014 at 3:02 AM, Justin Dong <jsd1 at rice.edu> wrote:
> > I actually was able to solve my own problem...for some reason, I need to do
> >
> > PCSetType(pc, PCLU);
> > PCFactorSetMatSolverPackage(pc, MATSOLVERSUPERLU_DIST);
> > KSPSetTolerances(ksp, 1.e-15, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT);
> >
> > 1) Before you do SetType(PCLU) the preconditioner has no type, so FactorSetMatSolverPackage() has no effect
> >
> > 2) There is a larger issue here. Never ever ever ever code in this way. Hardcoding a solver is crazy. The solver you
> >      use should depend on the equation, discretization, flow regime, and architecture. Recompiling for all those is
> >      out of the question. You should just use
> >
> >     KSPCreate()
> >     KSPSetOperators()
> >     KSPSetFromOptions()
> >     KSPSolve()
> >
> > and then
> >
> >    -pc_type lu -pc_factor_mat_solver_package superlu_dist
> >
> >
> > instead of the ordering I initially had, though I'm not really clear on what the issue was. However, there seems to be some loss of accuracy as I increase the number of processes. Is this expected, or can I force a lower tolerance somehow? I am able to compare the solutions to a reference solution, and the error increases as I increase the processes. This is the solution in sequential:
> >
> > Yes, this is unavoidable. However, just turn on the Krylov solver
> >
> >   -ksp_type gmres -ksp_rtol 1.0e-10
> >
> > and you can get whatever residual tolerance you want. To get a specific error, you would need
> > a posteriori error estimation, which you could include in a custom convergence criterion.
> >
> >   Thanks,
> >
> >      Matt
> >
> > superlu_1process = [
> > -6.8035811950925553e-06
> > 1.6324030474375778e-04
> > 5.4145340579614926e-02
> > 1.6640521936281516e-04
> > -1.7669374392923965e-04
> > -2.8099208957838207e-04
> > 5.3958133511222223e-02
> > -5.4077899123806263e-02
> > -5.3972905090366369e-02
> > -1.9485020474821160e-04
> > 5.4239813043824400e-02
> > 4.4883984259948430e-04];
> >
> > superlu_2process = [
> > -6.8035811950509821e-06
> > 1.6324030474371623e-04
> > 5.4145340579605655e-02
> > 1.6640521936281687e-04
> > -1.7669374392923807e-04
> > -2.8099208957839834e-04
> > 5.3958133511212911e-02
> > -5.4077899123796964e-02
> > -5.3972905090357078e-02
> > -1.9485020474824480e-04
> > 5.4239813043815172e-02
> > 4.4883984259953320e-04];
> >
> > superlu_4process= [
> > -6.8035811952565206e-06
> > 1.6324030474386164e-04
> > 5.4145340579691455e-02
> > 1.6640521936278326e-04
> > -1.7669374392921441e-04
> > -2.8099208957829171e-04
> > 5.3958133511299078e-02
> > -5.4077899123883062e-02
> > -5.3972905090443085e-02
> > -1.9485020474806352e-04
> > 5.4239813043900860e-02
> > 4.4883984259921287e-04];
> >
> > This is some finite element solution and I can compute the error of the solution against an exact solution in the functional L2 norm:
> >
> > error with 1 process:    1.71340e-02 (accepted value)
> > error with 2 processes: 2.65018e-02
> > error with 3 processes: 3.00164e-02
> > error with 4 processes: 3.14544e-02
> >
> >
> > Is there a way to remedy this?
> >
> >
> > On Wed, Apr 30, 2014 at 2:37 AM, Justin Dong <jsd1 at rice.edu> wrote:
> > Hi,
> >
> > I'm trying to solve a linear system in parallel using SuperLU but for some reason, it's not giving me the correct solution. I'm testing on a small example so I can compare the sequential and parallel cases manually. I'm absolutely sure that my routine for generating the matrix and right-hand side in parallel is working correctly.
> >
> > Running with 1 process and PCLU gives the correct solution. Running with 2 processes and using SUPERLU_DIST does not give the correct solution (I tried with 1 process too but according to the superlu documentation, I would need SUPERLU for sequential?). This is the code for solving the system:
> >
> >         /* solve the system */
> >       KSPCreate(PETSC_COMM_WORLD, &ksp);
> >       KSPSetOperators(ksp, Aglobal, Aglobal, DIFFERENT_NONZERO_PATTERN);
> >       KSPSetType(ksp,KSPPREONLY);
> >
> >       KSPGetPC(ksp, &pc);
> >
> >       KSPSetTolerances(ksp, 1.e-13, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT);
> >       PCFactorSetMatSolverPackage(pc, MATSOLVERSUPERLU_DIST);
> >
> >       KSPSolve(ksp, bglobal, bglobal);
> >
> > Sincerely,
> > Justin
> >
> >
> >
> >
> >
> >
> > --
> > 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
> 
> 



More information about the petsc-users mailing list