The default preconditioner with ILU(0) on each process is not appropriate for your problem and is producing overflow. Try -sub_pc_type lu and see if that produces a different result.
Is this a Stokes-like problem? Barry > On Mar 26, 2015, at 10:10 AM, Manav Bhatia <[email protected]> wrote: > > Thanks, Matt. > > Following is the output with: -ksp_monitor_lg_residualnorm -ksp_log -ksp_view > -ksp_monitor_true_residual -ksp_converged_reason > > 0 KSP preconditioned resid norm inf true resid norm > 2.709083260443e+06 ||r(i)||/||b|| 1.000000000000e+00 > Linear solve did not converge due to DIVERGED_NANORINF iterations 0 > KSP Object: 12 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=1000 > tolerances: relative=1e-10, absolute=1e-50, divergence=10000 > left preconditioning > using nonzero initial guess > using PRECONDITIONED norm type for convergence test > PC Object: 12 MPI processes > type: bjacobi > block Jacobi: number of blocks = 12 > 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: 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=667070, cols=667070 > package used to perform factorization: petsc > total: nonzeros=4.6765e+07, allocated nonzeros=4.6765e+07 > total number of mallocs used during MatSetValues calls =0 > using I-node routines: found 133414 nodes, limit used is 5 > linear system matrix = precond matrix: > Mat Object: () 1 MPI processes > type: seqaij > rows=667070, cols=667070 > total: nonzeros=4.6765e+07, allocated nonzeros=5.473e+07 > total number of mallocs used during MatSetValues calls =0 > using I-node routines: found 133414 nodes, limit used is 5 > linear system matrix = precond matrix: > Mat Object: () 12 MPI processes > type: mpiaij > rows=6723030, cols=6723030 > total: nonzeros=4.98852e+08, allocated nonzeros=5.38983e+08 > total number of mallocs used during MatSetValues calls =0 > using I-node (on process 0) routines: found 133414 nodes, limit used is > 5 > > > Anything jumps out at you as odd? > > -Manav > > > >> On Mar 26, 2015, at 9:34 AM, Matthew Knepley <[email protected]> wrote: >> >> On Thu, Mar 26, 2015 at 9:21 AM, Manav Bhatia <[email protected]> wrote: >> Hi, >> >> I am using the KSP linear solver for my system of equations, without any >> command line options at this point. I have checked that the L1 norms of my >> system matrix and the force vector are finite values, but the KSP solver is >> returning with an “inf” residual in the very first iteration. >> >> The problem has 6.7M dofs and I have tried this on multiple machines with >> different number of nodes with the same result. >> >> Is there a reason why the solver would return after the first iteration >> with an inf? >> >> I am not sure on where to start debugging this case, so I would >> appreciate any pointers. >> >> For all solver questions, we want to see the output of >> >> -ksp_view -ksp_monitor_true_residual -ksp_converged_reason >> >> The problem here would be that there is an error, so we would never see the >> output >> of -ksp_view and know what solver you are using. If you are using something >> complex, >> can you try using >> >> -pc_type jacobi >> >> and send the output from the options above? Then we can figure out why the >> other solver >> gets an inf. >> >> Thanks, >> >> Matt >> >> Thanks, >> Manav >> >> >> >> >> >> -- >> 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 >
