Screenshot instructions:
Windows
Mac
Red Hat Linux
Ubuntu
Click URL instructions:
Rightclick on ad, choose "Copy Link", then paste here →
(This may not be possible with some types of ads)
From: Lorenzo Zanon <zanon@ai...>  20131125 18:19:41

> Are you sure the Jacobian is correct? Is this a buckling problem, or > something nonsmooth like plasticity or contact? If not, Newton should > converge fairly fast. It should be, running on a 10x2x1 mesh I get convergence and a roughly correct output (wrt FEAP). The problem is just a bar with an applied load, modelled with nonlinear hysotropic elasticity (no plasticity, no contact). > Did you set a near null space (rigid body modes) using > MatSetNearNullSpace and perhaps MatNullSpaceCreateRigidBody? Are you > using a nodal basis? How do you enforce boundary conditions? I'm using Lagrange first order on HEX8 elements. Boundary conditions are set as in the example: std::set<boundary_id_type> boundary_ids; boundary_ids.insert(4); // 3D ... // Create a ZeroFunction to initialize dirichlet_bc ZeroFunction<> zf; DirichletBoundary dirichlet_bc(boundary_ids, variables, &zf); system.get_dof_map().add_dirichlet_boundary(dirichlet_bc); > Do you need the line search? with an NL decay rate that slow one of two things is happening I'd bet  either the Jacobian is incorrect, so your computed dU is not a descent direction(see the increase from steps 24) or the line search is behaving nonoptimally. Testing on the small 10x2x1 problem, without any option it fails: > exampleopt > NL step 0, residual_2 = 3.464102e05 > StVen system solved at nonlinear iteration 0 , final nonlinear residual norm: 3.464102e05 With linesearch and JFNK it works: > exampleopt snes_type ls snes_linesearch_type basic snes_mf_operator > NL step 0, residual_2 = 3.464102e05 > NL step 1, residual_2 = 2.417673e04 > NL step 2, residual_2 = 6.139470e08 > NL step 3, residual_2 = 5.468950e13 > NL step 4, residual_2 = 7.179521e18 > StVen system solved at nonlinear iteration 4 , final nonlinear residual norm: 7.179521e18 Using either linesearch or JFNK it doesn't converge, it looks like I need both. Apparently I can do without LU, but still it's very slow. Thanks Lorenzo On Nov 25, 2013, at 4:01 PM, Kirk, Benjamin (JSCEG311) wrote: > > On Nov 25, 2013, at 6:23 AM, Lorenzo Zanon <zanon@...> > wrote: > >> Hello, >> >> Thanks for the detailed answer! >> >> I've just rerun the executable taking off the snes_mf_operator and LU options, but it doesn't look any better... : >> >> Running ./exampleopt snes_type ls snes_linesearch_type basic ksp_rtol 1e4 >> NL step 0, residual_2 = 1.936492e05 >> NL step 1, residual_2 = 1.938856e05 >> NL step 2, residual_2 = 1.941222e05 >> NL step 3, residual_2 = 1.943592e05 >> NL step 4, residual_2 = 1.945964e05 >> >> Then I stopped it because it didn't look like converging, after one hour or so. >> > > Do you need the line search? with an NL decay rate that slow one of two things is happening I'd bet  either the Jacobian is incorrect, so your computed dU is not a descent direction(see the increase from steps 24) or the line search is behaving nonoptimally. > > I think snes_info (?) will present more verbose information that may be useful to you. > > Ben > 