Upgrading a neknek case to v26: stability & variable timestep

[joneuhauser]

In nekRS v23, variable timestep was supported in neknek:

nekRS/src/lib/nekrs.cpp

Lines 270 to 274 in 04492a7

	// during a neknek simulation, sync dt across all ranks
	if (nrs->neknek) {
	MPI_Allreduce(MPI_IN_PLACE, &nrs->dt[0], 1, MPI_DFLOAT, MPI_MIN, platform->comm.mpiCommParent);
	}

In v26, this was removed:

nekRS/src/lib/nekrs.cpp

Lines 418 to 423 in 96b3cf9

	if (platform->app->neknek) {
	// call only once as neknek doesn't support variable dt
	if (tstep == 1) {
	dt_ = platform->app->neknek->adjustDt(dt_);
	}
	}

What is the recommended replacement?

I have a case where I need to reduce the timestep considerably to an (initial) CFL of about 0.1 to get it to complete, while it works just fine with a variable timestep of targetcfl=3. I tried stabilizing it with the regularisation options without success (but it's laminar anyway).

Here's both cases for v23 and v26, as tagged on github:

overset_grid-v23.zip

overset_grid-v26.zip

(In reality the geometry of the "shell" is much more complex, hence the need for neknek, but the problem can be demonstrated in this simple case as well).

[stgeke]

In general, the time step size can vary, but it must be identical across all sessions. You’re correct that the current implementation does not fully support this use case. This will be addressed in the next upcoming release.