Roy Stogner <royst...@ices.utexas.edu> writes: > On Thu, 8 Jan 2015, Jed Brown wrote: > >>> 10 KSP preconditioned resid norm 1.580464771307e-13 true resid norm >>> 3.084021824157e-04 ||r(i)||/||b|| 2.711608086867e-03 >>> number of iterations to solve adjoint: 10 final residual of adjoint solve: >>> 1.58046e-13 >> >> This is usually caused by a singular preconditioner. > >> http://mid.mail-archive.com/87y6iyzl73.fsf@59A2.org > > Thanks! Is there a workaround that avoids false positives in such > cases? I've tried (with PETSc 3.5.2) "-ksp_pc_side right" and > "-ksp_type bcgs -ksp_norm_type unpreconditioned", but neither seems to > do much.
What do you mean? We use preconditioned norm by default mostly because some applications use penalty boundary conditions or poor scaling so that the unpreconditioned residual drops by many orders of magnitude without solving the problem. If you know your formulation is well-behaved and instead don't trust your preconditioner, then you should use unpreconditioned norm. Unfortunately, it costs a lot more in general to compute both norms. > It's odd that incomplete LU would be singular but incomplete Cholesky > would work fine, too, isn't it? Where do you see ILU being singular while ICC is not? What do you intend ICC to mean if the matrix is nonsymmetric? >>> It still seems like we've got to be doing something wrong in >>> PetscLinearSolver::adjoint_solve(). I haven't tried this myself, but >>> I'm told that setting up the transpose linearized problem here and >>> using a straight solve() works fine. >> >> What solver configuration is being used? Can you show >> -ksp_monitor_true_residual output for the original forward solve and the >> explicitly transposed solve? > > Okay, I take back the above. When I try this myself, doing a *linear* > forward solve from a zero initial guess, the behavior is nearly the > same as for the linear adjoint solve. (I am a little disturbed by > "nearly the same" for a symmetric matrix) Are you sure the matrix is exactly symmetric? What KSP are you using here? > Adjoint solve: > 0 KSP preconditioned resid norm 5.893300515271e-01 true resid norm > 1.137340546775e-01 ||r(i)||/||b|| 1.000000000000e+00 > 1 KSP preconditioned resid norm 1.323095860276e-01 true resid norm > 1.854948903571e-02 ||r(i)||/||b|| 1.630952935627e-01 > 2 KSP preconditioned resid norm 2.957976209084e-03 true resid norm > 9.076848423484e-04 ||r(i)||/||b|| 7.980765698734e-03 > 3 KSP preconditioned resid norm 4.630654713249e-04 true resid norm > 3.394455698647e-04 ||r(i)||/||b|| 2.984555248884e-03 > 4 KSP preconditioned resid norm 1.042927953583e-05 true resid norm > 3.082072165862e-04 ||r(i)||/||b|| 2.709893861255e-03 > 5 KSP preconditioned resid norm 7.161971654746e-07 true resid norm > 3.084038310662e-04 ||r(i)||/||b|| 2.711622582529e-03 > 6 KSP preconditioned resid norm 3.126650757674e-08 true resid norm > 3.084036219553e-04 ||r(i)||/||b|| 2.711620743934e-03 > 7 KSP preconditioned resid norm 1.102895578517e-09 true resid norm > 3.084021853448e-04 ||r(i)||/||b|| 2.711608112620e-03 > 8 KSP preconditioned resid norm 5.540930099957e-11 true resid norm > 3.084021840854e-04 ||r(i)||/||b|| 2.711608101548e-03 > 9 KSP preconditioned resid norm 2.937449992531e-12 true resid norm > 3.084021825944e-04 ||r(i)||/||b|| 2.711608088437e-03 > 10 KSP preconditioned resid norm 1.580464771307e-13 true resid norm > 3.084021824157e-04 ||r(i)||/||b|| 2.711608086867e-03 > > > Linear forward solve: > 0 KSP preconditioned resid norm 5.893175435351e-01 true resid norm > 1.137340546775e-01 ||r(i)||/||b|| 1.000000000000e+00 > 1 KSP preconditioned resid norm 1.323059758093e-01 true resid norm > 1.854739741057e-02 ||r(i)||/||b|| 1.630769030715e-01 > 2 KSP preconditioned resid norm 2.926513590091e-03 true resid norm > 9.072530459119e-04 ||r(i)||/||b|| 7.976969153912e-03 > 3 KSP preconditioned resid norm 4.631266360016e-04 true resid norm > 3.185176705725e-04 ||r(i)||/||b|| 2.800547922745e-03 > 4 KSP preconditioned resid norm 1.075300967442e-05 true resid norm > 2.824015720242e-04 ||r(i)||/||b|| 2.482999246135e-03 > 5 KSP preconditioned resid norm 5.982748177679e-07 true resid norm > 2.825507661757e-04 ||r(i)||/||b|| 2.484311026956e-03 > 6 KSP preconditioned resid norm 2.632996665496e-08 true resid norm > 2.825588496190e-04 ||r(i)||/||b|| 2.484382100156e-03 > 7 KSP preconditioned resid norm 2.222034839878e-09 true resid norm > 2.825586756678e-04 ||r(i)||/||b|| 2.484380570700e-03 > 8 KSP preconditioned resid norm 1.814426836084e-10 true resid norm > 2.825586967601e-04 ||r(i)||/||b|| 2.484380756153e-03 > 9 KSP preconditioned resid norm 2.844498948759e-12 true resid norm > 2.825586987199e-04 ||r(i)||/||b|| 2.484380773384e-03 > 10 KSP preconditioned resid norm 8.360209016600e-14 true resid norm > 2.825586987717e-04 ||r(i)||/||b|| 2.484380773840e-03 > > > *repeated* linear forward solves (using the solution of one as initial > iterate for the next) work, so perhaps NewtonSolver or PetscDiffSolver > was what was making up for the bad preconditioning. > --- > Roy
signature.asc
Description: PGP signature
------------------------------------------------------------------------------ Dive into the World of Parallel Programming! The Go Parallel Website, sponsored by Intel and developed in partnership with Slashdot Media, is your hub for all things parallel software development, from weekly thought leadership blogs to news, videos, case studies, tutorials and more. Take a look and join the conversation now. http://goparallel.sourceforge.net
_______________________________________________ Libmesh-devel mailing list Libmesh-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/libmesh-devel
