I have a problem in which I have performed a calculation on a remote machine, which has dolfin 2018.1.0, but I have 2019.1.0 running in a local Docker. I saved a checkpoint file on the remote using
dolfin.XDMFFile.write_checkpoint. On the remote machine, I can read that checkpoint back in successfully and continue the calculation.
When I move the xdmf/h5 files to my local machine, I can run
dolfin.XDMFFile.read_checkpoint without error, but the resulting problem does not solve.
It’s a nonlinear problem, running with Newton iterations. The checkpointed solution is used as the initial guess for a slightly modified problem (in this case, slightly larger light intensity in a semiconductor device model). When loaded inside the local Docker, the first several Newton iterations evaluate but do not converge; after 8 Newton iterations, the system develops nans. On the remote machine, the Newton process converges after reload in after 1 iteration.
I have evaluated the loaded-in quantities on both machines along line cuts (it’s a 2D problem) and haven’t seen any obvious discrepancies.
Is there any kind of known issue with dolfin2018.1.0 vs 2019.1.0 checkpoint files? I am not optimistic that there is a solution here, but I will ask anyway! Thank you.