[petsc-users] Choosing the best solvers for ISPH
Smith, Barry F.
bsmith at mcs.anl.gov
Sat Mar 17 11:46:59 CDT 2018
Algebraic multigrid has a variety of options that can effect convergence, you can find some of them at http://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/PC/PCGAMG.html these are the options you should try to improve convergence.
Changing the Krylov solver, as you discovered is generally not important to improve the speed of algebraic multigrid (just stick to the default KSPGMRES).
You can also try hypre's boomeramg with the option -pc_type hypre (you will need to ./configure PETSc with the option --download-hypre). hypre is a very good AMG implementation and may beat the one in PETSc.
But note that all the solvers in PETSc (and most on the web) are designed for finite element/finite difference discretizations, they are not designed for discretizations such as ISPH so they may not work particularly well for ISPH discretizations. You should do a web search for solvers customized for ISPH (I don't know if any exist).
> On Mar 17, 2018, at 7:41 AM, 我 <dayedut123 at 163.com> wrote:
> I use PETSC to solve the unsymmetrical matrix which is formed by the incompressible SPH method. Due to the scale of the problem is huge, I want to choose the best and the most efficient solver. I have checked the time-cost of the pre-conditioner PCGAMG together with the solvers like KSPCGS, KSPGMRES, KSPBCGS, KSPTFQMR and KSPGCR, but the difference between them is really really small. Except comparing the time-cost, are there any other properties that I can compare in order to find the most efficient solver? Are there any suggestions about choosing the best solver? Besides, I want to further speed the solving process up, are there any suggestions?
> Thank you very much!
More information about the petsc-users