<div dir="ltr">I think I found it. I thought initially that the problem was in the shift from KSPSetNullSpace to MatSetNullSpace, but in retrospect that doesn't make much sense as they ostensibly are offering the exact same functionality. I discovered that our code returns the same KSP_DIVERGED_INDEFINITE_PC error in commit e8f7834, which is the last commit before KSPSetNullSpace was removed. <div><br></div><div>I've attached the bisection log between e8f7834 and 9fbf19a (v3.5.4) and the problem seems to have been introduced in 25a145a ("fixed gamg coarse grid to be general").</div><div><br></div><div>Thanks,</div><div>Gil Forsyth</div></div><div class="gmail_extra"><br><div class="gmail_quote">On Wed, Sep 30, 2015 at 4:36 PM, Gil Forsyth <span dir="ltr"><<a href="mailto:gforsyth@gwu.edu" target="_blank">gforsyth@gwu.edu</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr">The exact same thing ran in serial without the indefinite problem, but it does crop up in all parallel runs. <div><br></div><div>I've attached the failure run log and I'll start bisecting against 3.5.4 to try to track down the change.</div><div><br></div><div>Thanks!</div><span class="HOEnZb"><font color="#888888"><div>Gil Forsyth</div></font></span></div><div class="HOEnZb"><div class="h5"><div class="gmail_extra"><br><div class="gmail_quote">On Wed, Sep 30, 2015 at 4:24 PM, Barry Smith <span dir="ltr"><<a href="mailto:bsmith@mcs.anl.gov" target="_blank">bsmith@mcs.anl.gov</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><br>
Did the exact same thing run in parallel without the indefinite problem?<br>
<br>
Run that failure with -info and send all the output<br>
<br>
You could use bisection to find out exactly what change in the library breaks your example.<br>
<br>
Barry<br>
<div><div><br>
> On Sep 30, 2015, at 3:11 PM, Gil Forsyth <<a href="mailto:gforsyth@gwu.edu" target="_blank">gforsyth@gwu.edu</a>> wrote:<br>
><br>
> Using PETSc master branch solved the problem in serial, but I'm still seeing the same KSP_DIVERGED_INDEFINITE_PC error when I run with MPI. This runs to completion when I don't use GAMG. Log is attached for the following run.<br>
><br>
> $PETSC_DIR/$PETSC_ARCH/bin/mpirun -n 2 $PETIBM_DIR/petibm-git/bin/petibm2d -directory . -poisson_pc_type gamg -poisson_pc_gamg_type agg -poisson_pc_gamg_agg_nsmooths 1 -poisson_ksp_view -poisson_ksp_monitor_true_residual -poisson_ksp_converged_reason<br>
><br>
><br>
> Thanks again,<br>
> Gil Forsyth<br>
><br>
><br>
> On Tue, Sep 29, 2015 at 1:12 PM, Gil Forsyth <<a href="mailto:gforsyth@gwu.edu" target="_blank">gforsyth@gwu.edu</a>> wrote:<br>
> Ah, got it. I'll checkout the master branch and see if the behavior persists.<br>
><br>
> Many thanks,<br>
> Gil<br>
><br>
> On Tue, Sep 29, 2015 at 1:10 PM, Matthew Knepley <<a href="mailto:knepley@gmail.com" target="_blank">knepley@gmail.com</a>> wrote:<br>
> On Tue, Sep 29, 2015 at 12:08 PM, Gil Forsyth <<a href="mailto:gforsyth@gwu.edu" target="_blank">gforsyth@gwu.edu</a>> wrote:<br>
> PETSc is version 3.6.1 -- I just included a log from 3.5.4 to show that the behavior seems to have changed between versions. The only difference in our code between 3.5.4 and 3.6.1 is the change from KSPSetNullSpace to MatSetNullSpace.<br>
><br>
> Mark made some GAMG changes which were later reversed because they had unintended consequences like this.<br>
> I think what Barry means is, "you should get the behavior you expect using the master branch from PETSc development"<br>
><br>
> Thanks,<br>
><br>
> Matt<br>
><br>
> On Tue, Sep 29, 2015 at 1:04 PM, Barry Smith <<a href="mailto:bsmith@mcs.anl.gov" target="_blank">bsmith@mcs.anl.gov</a>> wrote:<br>
><br>
> Update your PETSc<br>
><br>
><br>
> > On Sep 29, 2015, at 12:00 PM, Gil Forsyth <<a href="mailto:gforsyth@gwu.edu" target="_blank">gforsyth@gwu.edu</a>> wrote:<br>
> ><br>
> > Hi Barry,<br>
> ><br>
> > We aren't explicitly setting GMRES anywhere in the code and I'm not sure why it's being used. Running our 3.5.4 code using KSPSetNullSpace works with:<br>
> ><br>
> > $PETIBM_DIR/petibm3.5/bin/petibm2d -directory . -poisson_pc_type gamg -poisson_pc_gamg_type agg -poisson_pc_gamg_agg_nsmooths 1 -poisson_ksp_view -poisson_ksp_monitor_true_residual -poisson_ksp_converged_reason > kspview3.5.4<br>
> ><br>
> > and shows that the coarse grid solver is of type:preonly<br>
> ><br>
> > running the newer version that uses MatSetNullSpace in its stead and adding in -poisson_mg_coarse_ksp_type preonly<br>
> ><br>
> > $PETIBM_DIR/petibm3.6/bin/petibm2d -directory . -poisson_pc_type gamg -poisson_pc_gamg_type agg -poisson_pc_gamg_agg_nsmooths 1 -poisson_mg_coarse_ksp_type preonly -poisson_ksp_view -poisson_ksp_monitor_true_residual -poisson_ksp_converged_reason > kspview3.6.1<br>
> ><br>
> > still shows<br>
> ><br>
> > KSP Object:(poisson_) 1 MPI processes<br>
> > type: cg<br>
> > maximum iterations=10000<br>
> > tolerances: relative=1e-05, absolute=1e-50, divergence=10000<br>
> > left preconditioning<br>
> > using nonzero initial guess<br>
> > using PRECONDITIONED norm type for convergence test<br>
> > PC Object:(poisson_) 1 MPI processes<br>
> > type: gamg<br>
> > MG: type is MULTIPLICATIVE, levels=3 cycles=v<br>
> > Cycles per PCApply=1<br>
> > Using Galerkin computed coarse grid matrices<br>
> > GAMG specific options<br>
> > Threshold for dropping small values from graph 0<br>
> > AGG specific options<br>
> > Symmetric graph false<br>
> > Coarse grid solver -- level -------------------------------<br>
> > KSP Object: (poisson_mg_coarse_) 1 MPI processes<br>
> > type: gmres<br>
> > GMRES: restart=30, using Classical (unmodified) Gram-Schmidt Orthogonalization with no iterative refinement<br>
> > GMRES: happy breakdown tolerance 1e-30<br>
> > maximum iterations=1, initial guess is zero<br>
> > tolerances: relative=1e-05, absolute=1e-50, divergence=10000<br>
> > left preconditioning<br>
> > using NONE norm type for convergence test<br>
> ><br>
> ><br>
> > both logs are attached.<br>
> ><br>
> ><br>
> > On Tue, Sep 29, 2015 at 12:37 PM, Barry Smith <<a href="mailto:bsmith@mcs.anl.gov" target="_blank">bsmith@mcs.anl.gov</a>> wrote:<br>
> ><br>
> > This can't work. You can't use a GMRES inside a CG. Try changing to -poisson_mg_coarse_ksp_type preonly<br>
> ><br>
> > KSP Object:(poisson_) 1 MPI processes<br>
> > type: cg<br>
> ><br>
> > KSP Object: (poisson_mg_coarse_) 1 MPI processes<br>
> > type: gmres<br>
> > GMRES: restart=30, using Classical (unmodified) Gram-Schmidt Orthogonalization with no iterative refinement<br>
> > GMRES: happy breakdown tolerance 1e-30<br>
> > maximum iterations=1, initial guess is zero<br>
> ><br>
> ><br>
> > > On Sep 29, 2015, at 10:53 AM, Gil Forsyth <<a href="mailto:gforsyth@gwu.edu" target="_blank">gforsyth@gwu.edu</a>> wrote:<br>
> > ><br>
> > ><br>
> > > On Tue, Sep 29, 2015 at 11:42 AM, Matthew Knepley <<a href="mailto:knepley@gmail.com" target="_blank">knepley@gmail.com</a>> wrote:<br>
> > > On Tue, Sep 29, 2015 at 10:28 AM, Gil Forsyth <<a href="mailto:gforsyth@gwu.edu" target="_blank">gforsyth@gwu.edu</a>> wrote:<br>
> > > Hi all,<br>
> > ><br>
> > > I've been having some trouble with what should be a relatively simple update to an immersed boundary CFD solver from PETSc 3.5.4 to 3.6.1<br>
> > ><br>
> > > I'm getting indefinite PC errors for a simple lid-driven cavity test problem, 32x32 at Re 100<br>
> > ><br>
> > > Under PETSc 3.5.4 using KSPSetNullSpace we used the following to set the null space. This is for a 2D Poisson system with no immersed boundary and so the null space is the constant vector.<br>
> > ><br>
> > > MatNullSpace nsp;<br>
> > > ierr = MatNullSpaceCreate(PETSC_COMM_WORLD, PETSC_TRUE, 0, NULL, &nsp); CHKERRQ(ierr);<br>
> > > ierr = KSPSetNullSpace(ksp2, nsp); CHKERRQ(ierr);<br>
> > > ierr = MatNullSpaceDestroy(&nsp); CHKERRQ(ierr);<br>
> > ><br>
> > > Clearly this has to happen in the reverse order, since ksp2 would not be created yet.<br>
> > ><br>
> > > For questions about solvers, we HAVE to see the complete output of -ksp_view so we<br>
> > > know what we are dealing with. Its also nice to have -ksp_monitor_true_residual -ksp_converged_reason<br>
> > ><br>
> > > Matt<br>
> > ><br>
> > > Yes -- sorry, those are both in inline files and are called in the reverse order that I wrote them out.<br>
> > ><br>
> > > I've attached the output of<br>
> > ><br>
> > > $PETIBM_DIR/petibm3.6/bin/petibm2d -directory . -poisson_pc_type gamg -poisson_pc_gamg_type agg -poisson_gamg_agg_nsmooths 1 -poisson_ksp_view -poisson_ksp_monitor_true_residual -poisson_ksp_converged_reason > kspview.log<br>
> > ><br>
> > ><br>
> > ><br>
> > > And then setup the KSP with<br>
> > > ierr = KSPCreate(PETSC_COMM_WORLD, &ksp2); CHKERRQ(ierr);<br>
> > > ierr = KSPSetOptionsPrefix(ksp2, "poisson_"); CHKERRQ(ierr);<br>
> > > ierr = KSPSetOperators(ksp2, QTBNQ, QTBNQ); CHKERRQ(ierr);<br>
> > > ierr = KSPSetInitialGuessNonzero(ksp2, PETSC_TRUE); CHKERRQ(ierr);<br>
> > > ierr = KSPSetType(ksp2, KSPCG); CHKERRQ(ierr);<br>
> > > ierr = KSPSetReusePreconditioner(ksp2, PETSC_TRUE); CHKERRQ(ierr);<br>
> > > ierr = KSPSetFromOptions(ksp2); CHKERRQ(ierr);<br>
> > ><br>
> > > The matrix QTBNQ does not change, only the rhs of the system is updated.<br>
> > ><br>
> > > We run this with `-pc_type gamg -pc_gamg_type agg -pc_gamg_agg_nsmooths 1` and everything seems to work as expected.<br>
> > ><br>
> > > Under PETSc 3.6.1, we change only the KSPSetNullSpace line, to<br>
> > ><br>
> > > ierr = MatSetNullSpace(QTBNQ, nsp); CHKERRQ(ierr);<br>
> > ><br>
> > > and the same code diverges after 1 timestep and returns a -8 KSP_DIVERGED_INDEFINITE_PC<br>
> > ><br>
> > > This is weird, especially because if we change nsmooths to 2, it runs for 264 timesteps and the returns the same error. But we have explicitly set KSPSetReusePreconditioner so it should be using the same PC, right?<br>
> > ><br>
> > > Change nsmooths to 3 and it again diverges after 1 timestep.<br>
> > ><br>
> > > Change nsmooths to 4 and it runs to completion.<br>
> > ><br>
> > > It seems like either gamg's behavior has changed, or that KSPSetNullSpace was doing something implicitly that we now need to do explicitly in addition to MatSetNullSpace?<br>
> > ><br>
> > > Thanks,<br>
> > > Gil Forsyth<br>
> > ><br>
> > ><br>
> > ><br>
> > > --<br>
> > > What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.<br>
> > > -- Norbert Wiener<br>
> > ><br>
> > > <kspview.log><br>
> ><br>
> ><br>
> > <kspview3.5.4><kspview3.6.1><br>
><br>
><br>
><br>
><br>
><br>
> --<br>
> What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.<br>
> -- Norbert Wiener<br>
><br>
><br>
</div></div>> <mpi_n2_indefinite_pc.log><br>
<br>
</blockquote></div><br></div>
</div></div></blockquote></div><br></div>