Non-overlapping coupling with >2 subdomains does not work

[ikalash]

I checked in a simple test case involving 3 subdomains, for which the non-overlapping Schwarz method does not converge: https://github.com/sandialabs/Norma.jl/tree/main/examples/debug-cases/3-cuboid-symmetry-bc-bottom-fixed/nonoverlap . In the test, each subdomain is discretized by just 1 element. Schwarz diverges in the first time-step regardless of whether one end is pulled or fixed. Since the problem is so small, it is a good one to troubleshoot the issue.

[ikalash]

So, I am finding something fascinating. If I switch the order of the BCs to do Neumann-Dirichlet (ND) instead of Dirichlet-Neumann (DN), both versions of the 3 subdomain problem converge with Schwarz (though a lot of iters are required). Swapping the BCs with @lxmota 's fix to the swapping doesn't change the base ND or DN behavior at all.

[lxmota]

Somewhat consistent with the behavior that I observed on the 2 cubes problem, where the order of the BCs has a significant effect on convergence. I suppose this is due that at the interface, as one cube is pulled away from the other, which BC is applied first will lead to a better prediction of the displacement. Food for thought.

[ikalash]

Somewhat consistent with the behavior that I observed on the 2 cubes problem, where the order of the BCs has a significant effect on convergence.

Yes, I noticed also that there is an impact of the order on convergence. ND yields more Schwarz iters for cases where DN converges, but I am making more runs to gather more information.  

I suppose this is due that at the interface, as one cube is pulled away from the other, which BC is applied first will lead to a better prediction of the displacement. Food for thought.

Yes, something like this must be happening. I'll have more data on how the ordering impacts convergence soon.