Hi, this problem occurs because each Function.sub(deepcopy=True) call creates a new function space (which in turn uses 1 MPI communicator). Normally DOLFIN will free these communicators when the subfunctions are freed from memory (which would be typical for a for-loop), but in order to track inputs/outputs of operations dolfin-adjoint keeps these in memory.
Now in this case, you don’t have a deepcopy=True split, but dolfin-adjoint stores the values as deepcopies internally. However, we believe the annotation should still work without these deepcopies, so I have now pushed a commit to the master branch of dolfin-adjoint that should fix the issue in your posted code.
This can be installed using
pip install git+https://bitbucket.org/dolfin-adjoint/pyadjoint.git@master