ECP 14: in situ matrix modifications about nalu

[mhoemmen]

@spdomin wrote:

Establish local row- and column-maps for dofs only at moving mesh interfaces.

Would it be correct to translate this into Tpetra terms as follows? "Let Tpetra's sparse graph and matrices change their local structure, yet not change their communication pattern for sparse matrix-vector multiplies, if changing the communication pattern is unnecessary."

[alanw0]

Some combination of steps 1 and 2 probably do have the potential to improve performance on our side, without initially requiring Tpetra changes. That would then allow us plenty of lead time to collaborate with the Tpetra folks to stage in some enhancements in priority order, driven by good performance data.

[spdomin]

@mhoemmen; this is a long term ECP requirement on Tpetra that we may need elevated to the Trilinos backlog if it gets too deep for the team. Hopefully, this is a synergistic activity that other application teams require. You can see that we have some time to get this done...

[alanw0]

@mhoemmen, I'm not in deep enough yet to be sure if that's a correct description. But if a primary source of change is due to changing ghost entities, then a division between local structure vs boundary/communication patterns may indeed be a good way to describe it.

[spdomin]

Transition to Jira.