On Thu, 6 Oct 2016, David Knezevic wrote:
> I'm using GhostingFunctor for a contact solve, in which I consider a 1/4 > domain with partial Dirichlet boundary conditions that impose a symmetry > condition (i.e. displacement normal to the symmetry boundary is clamped to > zero, and tangential displacement is unconstrained). This means that I > have Dirichlet constraints that affect the dofs on the contact surface. > What I find is that the solve works fine in serial, but in parallel the > nonlinear convergence fails, presumably because of an incorrect Jacobian. > I have actually run into this exact issue before (when I was augmenting the > sparsity pattern "manually", prior to GhostingFunctor) and I found > that the issue was that the dof constraints on the contact surface were not > being communicated in parallel, which caused the incorrect Jacobian > in parallel. I fixed it by adding artificial Edge2 elements into the mesh > (like in systems_of_equations_ex8) to ensure that the dof constraints > are communicated properly in parallel. > > So, my question is, is there a way to achieve the necessary dof constraint > communication with the new GhostingFunctor API? I had hoped that using > "add_algebraic_ghosting_functor" would do the job, but it apparently doesn't. Hmm... If you only needed algebraic ghosting, then add_algebraic_ghosting_functor should have been sufficient - processors won't know about all their ghosted dofs' constraints, but the ghosted dofs will get properly added to the send_list, and enforce_constraints_exactly() will ensure that the dofs, once constrained on the processors which own them, will have their values updated on all the processors which ghost them. But you need to augment the sparsity pattern too, so you should be using add_coupling_functor instead... and in *that* case, we're broken, aren't we? You build a Jacobian connecting the remotely coupled dofs, and you try to hit it with constrain_element_foo() or heterogeneously_constrain_element_foo(), but the processor isn't aware of all the ghosted dofs' constraints, so the element constraint matrix is wrong and so is your final answer. I think the proper fix is to call the coupling functors in scatter_constraints(), then query the processors who own the elements which are to be coupled for any constraints on those elements' dofs. I can take a crack at that tomorrow or Monday. Any chance you could set up a unit test (or even just stuff an assertion into the misc ex9 example?) that checks for the problem? --- Roy ------------------------------------------------------------------------------ Check out the vibrant tech community on one of the world's most engaging tech sites, SlashDot.org! http://sdm.link/slashdot _______________________________________________ Libmesh-devel mailing list Libmesh-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/libmesh-devel
