Measuring the overhead of multiprocessing v.s. the new PEP684 subinterpreter. Starting 100 sub interpreters (each with its own GIL and therefore able to run truly in parallel) is 5.6x faster. #python