Laplace Equation preconditioner
Good morning PETSC users, I have a current solver that requires the solution of a Laplace equation to be reused for all future time steps. The configuration is axisymmetric with Dirichlet BCs at the top and bottom boundaries, and Zero Neumman conditions at the axis and far field. The grid is curvilinear and structured. So far I have been using PCHYPRE boomeramg as the preconditioner, which often works well, but I have also experienced DIVERGED_BREAKDOWN on many occasions. When I use direct solver PCMUMPS it always produces a satisfactory answer, which gives me confidence that the solution exists for the given grid in all these configurations where boomeramg fails. I am looking for recommendations on other preconditioners to try in this problem that can produce a solution faster than PCMUMPS, or recommendations on which parameters to adjust on boomeramg. Thank you, -- Alfredo Duarte Graduate Research Assistant The University of Texas at Austin
On Wed, Oct 12, 2022 at 1:38 PM Alfredo J Duarte Gomez <[email protected]> wrote:
Good morning PETSC users,
I have a current solver that requires the solution of a Laplace equation to be reused for all future time steps.
The configuration is axisymmetric with Dirichlet BCs at the top and bottom boundaries, and Zero Neumman conditions at the axis and far field. The grid is curvilinear and structured.
So far I have been using PCHYPRE boomeramg as the preconditioner, which often works well, but I have also experienced DIVERGED_BREAKDOWN on many occasions. When I use direct solver PCMUMPS it always produces a satisfactory answer, which gives me confidence that the solution exists for the given grid in all these configurations where boomeramg fails.
I do not know why Hypre is breaking down. Did you try ML or GAMG? They are easier to diagnose I think. Thanks, Matt
I am looking for recommendations on other preconditioners to try in this problem that can produce a solution faster than PCMUMPS, or recommendations on which parameters to adjust on boomeramg.
Thank you,
-- Alfredo Duarte Graduate Research Assistant The University of Texas at Austin
-- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
What KSP are you using? DIVERGED_BREAKDOWN is very rare for KSPGMRES. If you are using one of its lesser cousins like bcgs you might consider switching to bcgsl or gmres. I assume because of boundary conditions or the discretization you do not have symmetric positive definite and thus cannot use CG. Barry
On Oct 12, 2022, at 2:00 PM, Matthew Knepley <[email protected]> wrote:
On Wed, Oct 12, 2022 at 1:38 PM Alfredo J Duarte Gomez <[email protected] <mailto:[email protected]>> wrote: Good morning PETSC users,
I have a current solver that requires the solution of a Laplace equation to be reused for all future time steps.
The configuration is axisymmetric with Dirichlet BCs at the top and bottom boundaries, and Zero Neumman conditions at the axis and far field. The grid is curvilinear and structured.
So far I have been using PCHYPRE boomeramg as the preconditioner, which often works well, but I have also experienced DIVERGED_BREAKDOWN on many occasions. When I use direct solver PCMUMPS it always produces a satisfactory answer, which gives me confidence that the solution exists for the given grid in all these configurations where boomeramg fails.
I do not know why Hypre is breaking down. Did you try ML or GAMG? They are easier to diagnose I think.
Thanks,
Matt
I am looking for recommendations on other preconditioners to try in this problem that can produce a solution faster than PCMUMPS, or recommendations on which parameters to adjust on boomeramg.
Thank you,
-- Alfredo Duarte Graduate Research Assistant The University of Texas at Austin
-- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener
https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
Hello, I am using KSPGMRES. I was surprised to see DIVERGED_BREAKDOWN as well, which is why I thought there could be some grid issues. However, other preconditioners were able to retrieve a satisfactory solution (PCGAMG, PCGASM, PCLU MUMPS). Is there maybe a problem size for which boomerAMG has decreased performance? I am dealing with problem sizes of 2 million to 20 million points for reference. I tried out PCGAMG and it seems to be working much better than boomerAMG, so thank you so much for your suggestion. Sincerely, -Alfredo On Wed, Oct 12, 2022 at 1:32 PM Barry Smith <[email protected]> wrote:
What KSP are you using? DIVERGED_BREAKDOWN is very rare for KSPGMRES. If you are using one of its lesser cousins like bcgs you might consider switching to bcgsl or gmres.
I assume because of boundary conditions or the discretization you do not have symmetric positive definite and thus cannot use CG.
Barry
On Oct 12, 2022, at 2:00 PM, Matthew Knepley <[email protected]> wrote:
On Wed, Oct 12, 2022 at 1:38 PM Alfredo J Duarte Gomez < [email protected]> wrote:
Good morning PETSC users,
I have a current solver that requires the solution of a Laplace equation to be reused for all future time steps.
The configuration is axisymmetric with Dirichlet BCs at the top and bottom boundaries, and Zero Neumman conditions at the axis and far field. The grid is curvilinear and structured.
So far I have been using PCHYPRE boomeramg as the preconditioner, which often works well, but I have also experienced DIVERGED_BREAKDOWN on many occasions. When I use direct solver PCMUMPS it always produces a satisfactory answer, which gives me confidence that the solution exists for the given grid in all these configurations where boomeramg fails.
I do not know why Hypre is breaking down. Did you try ML or GAMG? They are easier to diagnose I think.
Thanks,
Matt
I am looking for recommendations on other preconditioners to try in this problem that can produce a solution faster than PCMUMPS, or recommendations on which parameters to adjust on boomeramg.
Thank you,
-- Alfredo Duarte Graduate Research Assistant The University of Texas at Austin
-- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener
https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
-- Alfredo Duarte Graduate Research Assistant The University of Texas at Austin
participants (3)
-
Alfredo J Duarte Gomez -
Barry Smith -
Matthew Knepley