[dune-pdelab] Dune::PDELab::ISTLBackend_NOVLP_CG_AMG_SSOR with ALUGrid

[Eike Mueller]

Hi Rebecca,

I presume you always meant P0ParallelGhostConstraints, not 
P0GhostConstraints?

I just had another go, and I find the following:

* If I use CG, preconditioned by AMG with ILU0 smoother on ALUGrid 
(cubed sphere), it does not matter as to whether I use 
P0GhostConstraints or the (wrong) constraints that I used before, the 
results agree in so far as the solution looks identical in both cases 
(in particular, no imprinted processor boundaries) and the number of 
iterations and convergence rate is exactly the same in both cases.
* If I use CG, preconditioned with ILU0 or SSOR on ALUGrid and 
P0Constraints, it does not work, I get convergence rates that are larger 
than 1 and the solution just looks wrong (I see the processor edges 
imprinted on the solution).
* If I use CG, preconditioned with ILU0 or SSOR on YaspGrid, it all 
works fine.

Is it possible that I'm doing something wrong when setting up the 
P0ParallelGhostConstraints?

I should also add that all the runs reported in this email were done 
with the trunk, not with version 2.1. Also, I'm now solving the Poisson 
equation on a spherical shell.

Cheers,

Eike

Rebecca Neumann wrote:
> Sorry,
> of course the P0GhostConstraints are only for ALU,
> for ovlp YaspGrid the P0Constraints have to be used.
> 
> Best Regards,
> Rebecca
>