ECP 5(b): Evaluate dropping Jacobian entries for sliding/overset

[nalucfd]

ECP 5(b): Evaluate dropping Jacobian entries for sliding/overset

[aprokop]

@spdomin Could you please elaborate on the ticket? I'd like to work on it but need more context.

[spdomin]

We should probably talk at the next stand-up then. The basic sliding mesh includes column entries of the opposing element for the DG contribution. If we remove those entries, the matrix structure does not change. However, this will come at the expense of more continuity iterations as noted above. I would go to the AssembleContinuityNonConformal algorithm and modify the connected nodes keeping only the current element connectivity and removing the opposing element connectivity from the stencil. You will also need to modify the Tpetra setup. Then, in the algorithm apply, you can remove the LHS += for the opposing nodes. Try one of the dg tests (possibly even removing the spinning of the block, i.e., omega = 0) and see what happens.

You can also avoid all of the above and see if there is an opportunity to "reuse" the preconditioner for Muelu.

[spdomin]

Attached are some preliminary overset and sliding mesh work when dropping Jacobian (LHS sensitivities)
slidingOversetMeshStencil.pdf
.

[spdomin]

Transition to Jira.