What I was trying to say, I guess, is that the additional serialization overhead can't be parallelized, which means that parallelization with multiprocessing doesn't help much in some cases where GIL-free threading would.