simplify access to PETSc/backend parameters

[Nico Schlömer]

(with reference to https://answers.launchpad.net/dolfin/+question/225891)

The mechanism of handing parameters to the linear backends currently uses a FEniCS-style dictionary that is parsed and translated into a PETSc/Epetra set of parameters. This parsing methods have grown more and more complex over the years to satisfy the ever evolving backends. There are still many parameters that are not accessible from Dolfin since the parameter interface wasn't updated accordingly (e.g., https://bugs.launchpad.net/dolfin/+bug/1159384).

A more maintainable approach to the problem would be useful. The obvious suggestion is to allow for any input parameters in Dolfin, and have the backend figure out whether or not they make sense. This would completely remove the burden of keeping the lists of possible parameters (and with it the documentation) in sync. The code base of Dolfin would get smaller. One downside of this is that most options would not be portable between the backends, e.g., using a Jacobian solver on the coarsest level of AMG would require a different setting in PETSc than in Epetra. One could argue though that switching the backend and thus using an entirely different implementation naturally requires some user interaction anyways.

[Martin Sandve Alnæs]

"Core" parameters should be backend-independent (make easy things easy), but I agree a mechanism for injecting more esoteric backend-specific parameters would be a good thing (make hard things possible).

[Prof Garth Wells]

Interface now exists to set PETSc option, see https://bitbucket.org/fenics-project/dolfin/commits/36860144ca08b0379d24eb5d83049f3265f19193

[Nico Schlömer]

Hm, not working for me in Python. The symbols are there,

$ nm --dynamic ~/Work/FEniCS/lib/libdolfin.so | grep PETScOptions
0000000000b948c0 T _ZN6dolfin12PETScOptions3setESs
0000000000b94bf0 T _ZN6dolfin12PETScOptions3setESsb
0000000000b95210 T _ZN6dolfin12PETScOptions3setESsd
0000000000b956c0 T _ZN6dolfin12PETScOptions3setESsi
0000000000b94ef0 T _ZN6dolfin12PETScOptions3setESsSs
0000000000b94710 T _ZN6dolfin12PETScOptions5clearESs

so maybe something is missing in the SWIG interface.

[Johan Hake]

@nico : could you please be more specific of what is not working, maybe provide a minimal example that does not work?

Also it might help triggering regeneration of SWIG interface by running:

make generate_swig_files

from your build directory.

[Prof Garth Wells]

See

demo/undocumented/mixed-poisson-sphere/python/demo_mixed-poisson-sphere.py

for a usage example.

[Nico Schlömer]

This very example fails for me with

Traceback (most recent call last):
  File "demo_mixed-poisson-sphere.py", line 56, in <module>
    PETScOptions.set("mat_mumps_icntl_14", 40.0)
NameError: name 'PETScOptions' is not defined

If it works for everyone else, I suppose we could consider this a configuration error that will likely be gone with a fresh install(tm).

[Johan Hake]

Works fine here. Have you tried regenerating the swig interface?

You can also clean your source directory from old generated files by:

git clean -fdx

triggering the automatic generation of the SWIG interface.

[Nico Schlömer]

Still not working. Meh, let's just let it sit there for a while; sooner or later I'll wipe everything and with it the problematic cache.