Thank you for your suggestions. Could you please suggest me what function
can work well for using a Krylov solver? I can no see examples.
My actual code is implemented using PETS (for sparsematrix, solver, etc). I
can see that SLEPcWrappers::SolverKrylovSchur allows PETS matrices.
Thank you again
El jueves, 10 de marzo de 2022 a las 15:12:19 UTC+1, bruno.t...@gmail.com
escribió:
> Hermes,
>
> I think Cuthill-McKee only works on symmetric matrices, is your matrix
> symmetric? Also, the goal of Cuthill-McKee is to help with the fill in
> of the matrix.There is no guarantee that it helps with the
> performance. If you don't know which preconditioner to use, you can
> use ILU (Incomplete LU decomposition). Basically, you use a direct
> solver but you drop all the "small" entries in the matrix. It's not
> the best preconditioner but you can control how much time you spend in
> the "direct solver". The problem with direct solvers is that there is
> not much you can do to speed them up. In practice, everybody uses
> Krylov solvers because of the problems you are encountering now.
>
> Best,
>
> Bruno
>
> Le jeu. 10 mars 2022 à 09:00, Hermes Sampedro
> <hermes...@gmail.com> a écrit :
> >
> > Hi Bruno,
> >
> > Yes, for now, I have to use a direct solver due to the preconditioner.
> > I am experiencing long computational times with the solver function. I
> am trying to use DoFRenumbering::Cuthill_McKee(dof_handler),
> DoFRenumbering::boost::Cuthill_McKee(dof_handler,false,false)
> > but I get even higher computational times. Am I doing something wrong?
> >
> > In the setup_system() function I do:
> > dof_handler.distribute_dofs(fe);
> > DoFRenumbering::Cuthill_McKee(dof_handler);
> >
> > Then thee solver is
> > void LaplaceProblem<dim>::solve()
> > {
> > PETScWrappers::MPI::Vector
> completely_distributed_solution(locally_owned_dofs,mpi_communicator);
> > SolverControl cn;
> > PETScWrappers::SparseDirectMUMPS solver(cn, mpi_communicator);
> > solver.solve(system_matrix, completely_distributed_solution, system_rhs);
> > constraints.distribute(completely_distributed_solution);
> > locally_relevant_solution = completely_distributed_solution;
> > }
> >
> > Thank you
> > Regards,
> > H
> >
> > El jueves, 10 de marzo de 2022 a las 14:54:13 UTC+1,
> bruno.t...@gmail.com escribió:
> >>
> >> Hermes,
> >>
> >> For large systems, Krylov solvers are faster and require less memory
> >> than direct solvers. Direct solvers scale poorly, in terms of memory
> >> and performance, with the number of unknowns. The only problem with
> >> Krylov solvers is that you need to use a good preconditioner. The
> >> choice of the preconditioner depends on the system that you want to
> >> solve.
> >>
> >> Best,
> >>
> >> Bruno
> >>
> >> Le jeu. 10 mars 2022 à 02:51, Hermes Sampedro
> >> <hermes...@gmail.com> a écrit :
> >> >
> >> > Dear Bruno,
> >> >
> >> > Thank you again for your answer.
> >> >
> >> > I managed to solve now a system of 3.5 million DOF using the same
> solver as I posted above, SparseDirectMUMPS. Now, in release mode, the
> assembling takes a few minutes instead of hours, however, the solver
> function takes approximately 1.5h (per frequency iteration) using 40
> processes in parallel (similar to step-40).
> >> >
> >> > I was expecting to get faster performance when running in parallel
> with 40 processes, especially because I need to run for several
> frequencies. I would like to ask if you also would expect faster
> performance. Would that be solved using the solver that you suggested
> (Krylov)?
> >> >
> >> >
> >> > Thank you
> >> >
> >> > Regards,
> >> >
> >> > H
> >> >
> >> >
> >> > El lunes, 7 de marzo de 2022 a las 15:04:19 UTC+1,
> bruno.t...@gmail.com escribió:
> >> >>
> >> >> Hermes,
> >> >>
> >> >> The problem is that you are using a direct solver. Direct solvers
> >> >> require a lot of memory because the inverse of a sparse matrix is
> >> >> generally not sparse. If you use a LU decomposition, which I think
> >> >> MUMPS does, you need a dense matrix to store the LU decomposition.
> >> >> That's a lot of memory! You will need to use a Krylov to solve a
> >> >> problem of this size.
> >> >>
> >> >> Best,
> >> >>
> >> >> Bruno
> >> >>
> >> >> Le dim. 6 mars 2022 à 07:19, Hermes Sampedro <hermes...@gmail.com>
> a écrit :
> >> >> >
> >> >> > Dear Bruno,
> >> >> >
> >> >> > Thank you very much for the comments. The problem was that I was
> running in Debug mode without knowing. Now, after changing to Release the
> assembling time is considerably reduced.
> >> >> >
> >> >> > Moreover, I am experiencing another issue that I would like to
> ask. My mesh is done with hyper_cube() in 3D and 5 refinements. The dof is
> around 3 million. When running, I always get a memory issue and the program
> stops. I realized that the problem is in the line that executes
> solver.solve(system_matrix, completely_distributed_solution, system_rhs);
> >> >> > I am using SparseMatrix and I do not fully understand where the
> problem could come from. The matrices are initialized beforehand, what
> reason do you think It could produce a memory issue in the solver?
> >> >> >
> >> >> > Below is the full solver function:
> >> >> >
> >> >> > template <int dim>
> >> >> > void LaplaceProblem<dim>::solve()
> >> >> > {
> >> >> > PETScWrappers::MPI::Vector
> completely_distributed_solution(locally_owned_dofs,mpi_communicator);
> >> >> > SolverControl cn;
> >> >> > PETScWrappers::SparseDirectMUMPS solver(cn, mpi_communicator);
> >> >> > solver.solve(system_matrix, completely_distributed_solution,
> system_rhs);
> >> >> > constraints.distribute(completely_distributed_solution);
> >> >> > locally_relevant_solution = completely_distributed_solution;
> >> >> > }
> >> >> >
> >> >> >
> >> >> > Thank you again for your help
> >> >> > Regards
> >> >> > H.
> >> >> >
> >> >> > El jueves, 3 de marzo de 2022 a las 15:13:30 UTC+1,
> bruno.t...@gmail.com escribió:
> >> >> >>
> >> >> >> Hermes,
> >> >> >>
> >> >> >> There is a couple of things that you could do but it probably
> won't give you a significant speed up. Are you sure that you are running in
> Release mode and not in Debug? Do you evaluate complicated functions in the
> assembly?
> >> >> >> A couple changes that could help:
> >> >> >> - don't use fe.system_to_component_index(i).first and
> fe.system_to_component_index(j).first everywhere. Just define const k = ...
> and const m = ... and use k and m. That might help the compiler with some
> optimizations
> >> >> >> - move the two if for the cell assembly outside the for loop on
> the quadrature point, similar to what you did for the boundaries. This
> could potentially help quite a bit if the cpu often gets the branch
> prediction wrong
> >> >> >>
> >> >> >> Best,
> >> >> >>
> >> >> >> Bruno
> >> >> >>
> >> >> >> On Thursday, March 3, 2022 at 4:31:04 AM UTC-5
> hermes...@gmail.com wrote:
> >> >> >>>
> >> >> >>> Dear all,
> >> >> >>>
> >> >> >>> I am experiencing long times when computing the assembling and I
> would like to ask if this is common or there is something wrong with my
> implementation.
> >> >> >>>
> >> >> >>> My model is built in a similar way as step-29 and step-40 (using
> complex values ad solving with a direct solver using distributed parallel
> implementation).
> >> >> >>> Now I am running larger systems with 3.5million dof and the
> assembling took 16h, while the solver function took much less.
> >> >> >>>
> >> >> >>> I can show the structure of my assembly_system() function to ask
> if there is something that can be done in order to speed up the process:
> >> >> >>>
> >> >> >>> void Problem<dim>::assemble_system()
> >> >> >>> {
> >> >> >>> for (unsigned int i = 0; i < dofs_per_cell; ++i) {
> >> >> >>> for (unsigned int j = 0; j < dofs_per_cell; ++j)
> >> >> >>> {
> >> >> >>> for (unsigned int q_point = 0; q_point < n_q_points; ++q_point)
> >> >> >>> {
> >> >> >>> if (fe.system_to_component_index(i).first ==
> fe.system_to_component_index(j).first)
> >> >> >>> {
> >> >> >>> cell_matrix(i, j) += ....
> >> >> >>> }
> >> >> >>> if (fe.system_to_component_index(i).first !=
> fe.system_to_component_index(j).first)
> >> >> >>> {
> >> >> >>> cell_matrix(i, j) += ....
> >> >> >>> }
> >> >> >>> }
> >> >> >>>
> >> >> >>> // Boundaries
> >> >> >>> if (fe.system_to_component_index(i).first ==
> fe.system_to_component_index(j).first)
> >> >> >>> {
> >> >> >>> for (unsigned int face_no : GeometryInfo<dim>::face_indices())
> >> >> >>> if (cell->face(face_no)->at_boundary() &&
> (cell->face(face_no)->boundary_id() == 0))
> >> >> >>> {
> >> >> >>> fe_face_values.reinit(cell, face_no);
> >> >> >>> for (unsigned int q_point = 0; q_point < n_face_q_points;
> ++q_point)
> >> >> >>> cell_matrix(i, j) += ....
> >> >> >>> }
> >> >> >>> }
> >> >> >>> if (fe.system_to_component_index(i).first !=
> fe.system_to_component_index(j).first)
> >> >> >>> {
> >> >> >>> for (unsigned int face_no : GeometryInfo<dim>::face_indices())
> >> >> >>> {
> >> >> >>> if (cell->face(face_no)->at_boundary() &&
> (cell->face(face_no)->boundary_id() == 0))
> >> >> >>> {
> >> >> >>> fe_face_values.reinit(cell, face_no);
> >> >> >>> for (unsigned int q_point = 0; q_point <
> n_face_q_points;++q_point)
> >> >> >>> cell_matrix(i, j) += ....
> >> >> >>> }
> >> >> >>> }
> >> >> >>> }
> >> >> >>> }
> >> >> >>> }
> >> >> >>>
> >> >> >>>
> >> >> >>> Thank you very much.
> >> >> >>> Regards,
> >> >> >>> Hermes
> >> >> >
> >> >> > --
> >> >> > The deal.II project is located at http://www.dealii.org/
> >> >> > For mailing list/forum options, see
> https://groups.google.com/d/forum/dealii?hl=en
> >> >> > ---
> >> >> > You received this message because you are subscribed to a topic in
> the Google Groups "deal.II User Group" group.
> >> >> > To unsubscribe from this topic, visit
> https://groups.google.com/d/topic/dealii/KQGIxkJZL6w/unsubscribe.
> >> >> > To unsubscribe from this group and all its topics, send an email
> to dealii+un...@googlegroups.com.
> >> >> > To view this discussion on the web visit
> https://groups.google.com/d/msgid/dealii/d80cbd1c-7d1f-4042-a9e8-e8382a721790n%40googlegroups.com
> .
> >> >
> >> > --
> >> > The deal.II project is located at http://www.dealii.org/
> >> > For mailing list/forum options, see
> https://groups.google.com/d/forum/dealii?hl=en
> >> > ---
> >> > You received this message because you are subscribed to a topic in
> the Google Groups "deal.II User Group" group.
> >> > To unsubscribe from this topic, visit
> https://groups.google.com/d/topic/dealii/KQGIxkJZL6w/unsubscribe.
> >> > To unsubscribe from this group and all its topics, send an email to
> dealii+un...@googlegroups.com.
> >> > To view this discussion on the web visit
> https://groups.google.com/d/msgid/dealii/0438f349-0ecd-4503-944c-6a5068599ccfn%40googlegroups.com
> .
> >
> > --
> > The deal.II project is located at http://www.dealii.org/
> > For mailing list/forum options, see
> https://groups.google.com/d/forum/dealii?hl=en
> > ---
> > You received this message because you are subscribed to a topic in the
> Google Groups "deal.II User Group" group.
> > To unsubscribe from this topic, visit
> https://groups.google.com/d/topic/dealii/KQGIxkJZL6w/unsubscribe.
> > To unsubscribe from this group and all its topics, send an email to
> dealii+un...@googlegroups.com.
> > To view this discussion on the web visit
> https://groups.google.com/d/msgid/dealii/cb8066c4-d160-4ebb-977b-d96ca07a93e5n%40googlegroups.com
> .
>
--
The deal.II project is located at http://www.dealii.org/
For mailing list/forum options, see
https://groups.google.com/d/forum/dealii?hl=en
---
You received this message because you are subscribed to the Google Groups
"deal.II User Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to dealii+unsubscr...@googlegroups.com.
To view this discussion on the web visit
https://groups.google.com/d/msgid/dealii/bea0a976-3413-4e85-aade-a05934440648n%40googlegroups.com.
