Luca Bertagna

> > The easiest way to do this could be to: 1) disable push to master (i.e., force to go through PR, even for small fixes), 2) have PR reviewers...

Isn't Kokkos already a requirement for Albany to build? I think we should remove the host implementation whenever the kokkos one is available, and do away with the kokkos macro....

Ok, I can see how it can be beneficial to keep the non-kokkos version. @jewatkins your concern about kokkos kernels optimality makes me wonder whether we should run a performance...

Is this an absolute difference or a relative one?

Ok. It might be pointing to a wrongly assembled jacobian. Meanwhile, I double checked a basic pb (SteadyHeat2D), running it with 1 and 32 ranks: the NOX residuals start to...

However, J_serial-J_parallel ~ 1e-16. So this is interesting: the jacobian is basically the same, and yet we get very different convergence histories... I'm not an expert of Ifpack, but: when...

Using MueLu (instead of Ifpack2) gives _the same_ linear iterations (serial vs 32-ranks). Like serial: ``` ******************************************************* ***** Belos Iterative Solver: Block Gmres ***** Maximum Iterations: 100 ***** Block Size:...

1e-13 is not quite machine precision, but it might be due to some carries, so it might still be a roundoff diff. This is good, since it suggests that we're...

Thanks! No need to add the exact input file.

So, I tried that input file, with the meshes in `/home/projects/albany/ali-perf-tests-meshes`. `Use Serial Mesh` causes little change in the convergence history (serial mesh gives 1 less linear solver iteration at...