Accessing Petsc solver submethods in FEniCS python interface

Question

Hello. I am using Dolfin 1.2.0 installed through Dorsal.

We are looking to better configure our solver for parallel computing using MPI. Currently we have set:

problem = LinearVariationalProblem(a, L, u, bcs)
solver = LinearVariationalSolver(problem)
solver.parameters['linear_solver'] = 'gmres'
solver.parameters['preconditioner'] = 'additive_schwarz'

which returns the report:

PETSc Krylov solver (gmres, asm) converged in 2 iterations.
PETSc Krylov solver preconditioner (asm) submethods: (preonly, ilu)

This solver works great in serial, but the method fails to converge under the same parameters as we increase the number of processes. Is it possible to change the PETSc Krylov preconditioner parameter 'preonly'?

Thank you,
Brian

Comments

I should add that the solver converges quickly in serial (2 iterations in the posted example) but fails to converge in 3000 iterations when running in parallel.

Answer 1

Hi Brian, I can't say I've had experience with parallel computing but I managed to change some of the petsc parameters directly by adding the line:

from dolfin import *
parameters.parse()

in my file, and then launching the program by running:

python 2Component_snes.py --petsc.ksp_type preonly --petsc.pc_type lu --petsc.snes_ls_alpha 1e-2

With this I was able to change some of the petsc parameters directly although it doesn't work for all of them. Just add '--petsc.' to a petsc word given in the manual.

Comments

This is exactly what I needed. Thank you!

Answer 2

In addition to the answer from mwelland, the development version of DOLFIN now has a wrapper for PETSc options. You can now do, for example,

PETScOptions::set("pc_type", "ml");  

from C++, and

PETScOptions.set("pc_type", "ml")  

from Python.