Solver improvements #2716
Solver improvements #2716
Conversation
No mallocs now needed when setting matrix elements in serial or parallel for 1D-recycling example. Hypre preconditioner type can now be set with pc_hypre_type option. Defaults to pilut, as this seems to give the best performance in parallel.
cvode_nonlinear_convergence_coef
Sets parameter through call to CVodeSetNonlinConvCoef
https://sundials.readthedocs.io/en/latest/cvodes/Usage/SIM.html#c.CVodeSetNonlinConvCoef
cvode_linear_convergence_coef
Sets parameter through call to CVodeSetEpsLin
https://sundials.readthedocs.io/en/latest/cvodes/Usage/SIM.html#c.CVodeSetEpsLin
Add some documentation of new cvode and beuler solver options.
If PETSc is not configured with Hypre then PCHYPRESetType is not defined.
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
| @@ -458,6 +467,10 @@ int CvodeSolver::init() { | |||
| #endif | |||
| } | |||
|
|
|||
| // Set internal tolerance factors | |||
| CVodeSetNonlinConvCoef(cvode_mem, cvode_nonlinear_convergence_coef); | |||
| CVodeSetEpsLin(cvode_mem, cvode_linear_convergence_coef); | |||
There was a problem hiding this comment.
warning: use of undeclared identifier 'CVodeSetEpsLin' [clang-diagnostic-error]
;
^
There was a problem hiding this comment.
Ah, this is because we use the system version of SUNDIALS for the clang-tidy action, which turns out to be 3.1.2.
I think just updating clang-tidy-review will use a more recent version of Ubuntu and give us SUNDIALS 5.1.8, but I guess the question which versions of SUNDIALS do we want to support?
|
This is exciting. Can we add some flags for diagnostics? I remember that SNES had a separate tolerance setting for controlling when it's "steady state enough" to go fast and not worry about time accuracy. Could this be a factor that influences the tests? I don't remember whether we have it exposed in the input file or not. I seem to remember it was called |
If a step fails, re-calculate the Jacobian and preconditioner. This can be an expensive operation, so doing this on failure allows lag_jacobian to be increased.
Enables `petsc:log_view=true` option to set the PETSc `-log_view` switch correctly.
|
clang-tidy review says "All clean, LGTM! 👍" |
| @@ -234,10 +240,26 @@ int SNESSolver::init() { | |||
| std::vector<PetscInt> o_nnz(localN); | |||
|
|
|||
| // Set values for most points | |||
| const int ncells_x = (mesh->LocalNx > 1) ? 2 : 0; | |||
There was a problem hiding this comment.
| const int ncells_x = (mesh->LocalNx > 1) ? 2 : 0; | |
| const int ncells_x = (mesh->xstart > 0) ? 2 : 0; |
Maybe rather check whether we have guards in x and y?
| @@ -234,10 +240,26 @@ int SNESSolver::init() { | |||
| std::vector<PetscInt> o_nnz(localN); | |||
|
|
|||
| // Set values for most points | |||
| const int ncells_x = (mesh->LocalNx > 1) ? 2 : 0; | |||
| const int ncells_y = (mesh->LocalNy > 1) ? 2 : 0; | |||
There was a problem hiding this comment.
| const int ncells_y = (mesh->LocalNy > 1) ? 2 : 0; | |
| const int ncells_y = (mesh->ystart > 0) ? 2 : 0; |
Co-authored-by: David Bold <dschwoerer@users.noreply.github.com>
Co-authored-by: David Bold <dschwoerer@users.noreply.github.com>
Co-authored-by: David Bold <dschwoerer@users.noreply.github.com>
cvode
Added input options to control tolerances on the inner nonlinear and linear solvers. Tightening the linear solver tolerance, for example, might improve the robustness of the nonlinear iterations. The optimal choice will be problem dependent, so the defaults are unchanged.
beuler / snes
Improved / fixed the preallocation of Jacobian rows, and detection of boundary cells, so that no allocations are needed while setting the stencil pattern. Tested in 1D and 2D (tokamak X-point).
Added an option
pc_hypre_typeto specify the type of hypre solver (choices areboomeramg,euclidandpilut). The default ispilutwhich seems to work best for a 1D test problem;euclidseems to be better in 2D.boomeramgdoesn't seem to work very well for the transport problems tested.Moved the PETSc options setting to before PETSc is initialized. This fixes the log_view option. In BOUT.inp:
or on the command line
petsc:log_view