Danyang : I tested your data. Your matrices encountered zero pivots, e.g. petsc/src/ksp/ksp/examples/tutorials (master) $ mpiexec -n 24 ./ex10 -f0 a_react_in_2.bin -rhs b_react_in_2.bin -ksp_monitor -ksp_error_if_not_converged
[15]PETSC ERROR: Zero pivot in LU factorization: http://www.mcs.anl.gov/petsc/documentation/faq.html#zeropivot [15]PETSC ERROR: Zero pivot row 1249 value 2.05808e-14 tolerance 2.22045e-14 ... Adding option '-sub_pc_factor_shift_type nonzero', I got mpiexec -n 24 ./ex10 -f0 a_react_in_2.bin -rhs b_react_in_2.bin -ksp_monitor -ksp_error_if_not_converged -sub_pc_factor_shift_type nonzero -mat_view ascii::ascii_info Mat Object: 24 MPI processes type: mpiaij rows=450000, cols=450000 total: nonzeros=6991400, allocated nonzeros=6991400 total number of mallocs used during MatSetValues calls =0 not using I-node (on process 0) routines 0 KSP Residual norm 5.849777711755e+01 1 KSP Residual norm 6.824179430230e-01 2 KSP Residual norm 3.994483555787e-02 3 KSP Residual norm 6.085841461433e-03 4 KSP Residual norm 8.876162583511e-04 5 KSP Residual norm 9.407780665278e-05 Number of iterations = 5 Residual norm 0.00542891 Hong > Hi Matt, > > Yes. The matrix is 450000x450000 sparse. The hypre takes hundreds of > iterates, not for all but in most of the timesteps. The matrix is not well > conditioned, with nonzero entries range from 1.0e-29 to 1.0e2. I also made > double check if there is anything wrong in the parallel version, however, > the matrix is the same with sequential version except some round error > which is relatively very small. Usually for those not well conditioned > matrix, direct solver should be faster than iterative solver, right? But > when I use the sequential iterative solver with ILU prec developed almost > 20 years go by others, the solver converge fast with appropriate > factorization level. In other words, when I use 24 processor using hypre, > the speed is almost the same as as the old sequential iterative solver > using 1 processor. > > I use most of the default configuration for the general case with pretty > good speedup. And I am not sure if I miss something for this problem. > > Thanks, > > Danyang > > On 17-05-24 11:12 AM, Matthew Knepley wrote: > > On Wed, May 24, 2017 at 12:50 PM, Danyang Su <danyang...@gmail.com> wrote: > >> Hi Matthew and Barry, >> >> Thanks for the quick response. >> >> I also tried superlu and mumps, both work but it is about four times >> slower than ILU(dt) prec through hypre, with 24 processors I have tested. >> > You mean the total time is 4x? And you are taking hundreds of iterates? > That seems hard to believe, unless you are dropping > a huge number of elements. > >> When I look into the convergence information, the method using ILU(dt) >> still takes 200 to 3000 linear iterations for each newton iteration. One >> reason is this equation is hard to solve. As for the general cases, the >> same method works awesome and get very good speedup. >> > I do not understand what you mean here. > >> I also doubt if I use hypre correctly for this case. Is there anyway to >> check this problem, or is it possible to increase the factorization level >> through hypre? >> > I don't know. > > Matt > >> Thanks, >> >> Danyang >> >> On 17-05-24 04:59 AM, Matthew Knepley wrote: >> >> On Wed, May 24, 2017 at 2:21 AM, Danyang Su <danyang...@gmail.com> wrote: >> >>> Dear All, >>> >>> I use PCFactorSetLevels for ILU and PCFactorSetFill for other >>> preconditioning in my code to help solve the problems that the default >>> option is hard to solve. However, I found the latter one, PCFactorSetFill >>> does not take effect for my problem. The matrices and rhs as well as the >>> solutions are attached from the link below. I obtain the solution using >>> hypre preconditioner and it takes 7 and 38 iterations for matrix 1 and >>> matrix 2. However, if I use other preconditioner, the solver just failed at >>> the first matrix. I have tested this matrix using the native sequential >>> solver (not PETSc) with ILU preconditioning. If I set the incomplete >>> factorization level to 0, this sequential solver will take more than 100 >>> iterations. If I increase the factorization level to 1 or more, it just >>> takes several iterations. This remind me that the PC factor for this >>> matrices should be increased. However, when I tried it in PETSc, it just >>> does not work. >>> >>> Matrix and rhs can be obtained from the link below. >>> >>> https://eilinator.eos.ubc.ca:8443/index.php/s/CalUcq9CMeblk4R >>> >>> Would anyone help to check if you can make this work by increasing the >>> PC factor level or fill? >>> >> >> We have ILU(k) supported in serial. However ILU(dt) which takes a >> tolerance only works through Hypre >> >> http://www.mcs.anl.gov/petsc/documentation/linearsolvertable.html >> >> I recommend you try SuperLU or MUMPS, which can both be downloaded >> automatically by configure, and >> do a full sparse LU. >> >> Thanks, >> >> Matt >> >> >>> Thanks and regards, >>> >>> Danyang >>> >>> >>> >> >> >> -- >> 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 >> >> http://www.caam.rice.edu/~mk51/ >> >> >> > > > -- > 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 > > http://www.caam.rice.edu/~mk51/ > > >