Demo multimesh-poisson has spikes

[Anders Logg (Chalmers)]

The solution computed by the demo multimesh-poisson has spikes in the solution that shouldn't be there. Missing or wrong stabilization?

Concerns both the Python and C++ version of the demos (which should be identical).

[Anders Logg (Chalmers)]

Solution as seen from below at step 6.

[August Johansson]

I've done the following:

• I checked the formulation and it's the same as the one I've used previously.
• I ran the demo with geometry debugging and didn't encounter any errors.
• I increased \beta and I got a larger spike!

I need to think about this one a bit more.

[Magne Nordaas]

The dof causing the spike is fully covered by the upper mesh. The current implementation is using GenericMatrix.ident_zeros to eliminate such dofs, making it sensitive to round-off errors from the inclusion/exclusion calculations. This is what happened here - the roundoff errors were marginally larger than the threshold value which is DOLFIN_EPS.

[Jan Blechta]

Yes, there's missing correct scaling by some sort of a row norm.

[Anders Logg (Chalmers)]

Resolved by Magne (improved stabilization parameters/terms).