I have been playing around with the MPI version. On installation, all tests pass, which is good news. However, I notice there is no test coverage (at least by default) for things like continuesoln and findsoln.
In fact, I find that if I converge a state I typically get a different answer if I build with MPI versus without MPI. I wonder if this is known about and has an explanation?
As a starting point, I’ve taken EQ1 at Re = 400 and run
findsoln -eqb -R 405 -T 20 -L eq1.nc
to converge to a new state at Re = 405.
If I compute the L2 distance with L2op between the states obtained with channelflow compiled as serial vs MPI, it is ~4.5e-8.
The converged solution in the case of the MPI build is always the same - regardless of how many MPI tasks there are (even if there’s just 1).