Update: MPI-parallel working version

[fdrmrc]

The present version now support also distributed memory parallelism. It has been tested both in 2D and 3D on distributed grids with different number of processors and configurations. @andreacangiani

Here are some first working preliminary results:

• $\Omega=[0,1]^d$,
• $u=\Pi_{i=1}^{d}\sin(2 \pi x_i)$,
• Left: partition across processors. Each color represents a processor which owns that portion of triangulation (12 in these tests),
• Right: interpolation on the fine triangulation of the solution defined on the agglomerated mesh.

$d=2$

$d=3$

[andreacangiani]

Thanks Marco! Seems like great news. Looking forward to more detailed results. Not your fault but the way the mesh is subdivided among processors is weird...

[fdrmrc]

Yes the partitions are generated by the library p4est which guarantees (among other features) to have load balancing. Then, locally to each partition we are free to agglomerate elements in the way we prefer.

[fdrmrc]

@luca-heltai , @andreacangiani

Some more preliminary numbers for the 3D Laplace problem on $[0,1]^3$ (on my laptop):

• Number of processors: 12
• Fine mesh $\tau_h$: 2.097.152 hexas
• Agglomerated mesh: 180.000 polygons
• Finite Element Space: DGQ of degree $1$ => number of DoFs=180.000*8=1.440.000
• Assembly time: ~60[s]
• Solver: CG preconditioned by AMG
• Time to solution: ~6[s]
• $L^2$ error: 9e-4