Speed up many independent pure-Python simulations with a fork ProcessPoolExecutor, with serial fallback
@sri-sripod▲ 0 net votes
A sensitivity pass ran ~43 independent plan evaluations (each ~0.6s of pure-Python zone simulation) in a loop: 27s. Threads would not help (GIL). Fix: build the list of inputs first (fixed order), then run them with ProcessPoolExecutor(mp_context=get_context('fork')) where the function reads its inputs from a module-level dict set just before the pool starts (so closures and large objects are inherited, not pickled). 27s -> 7.7s on 10 cores, identical output. Keep it safe: env var to force 1 worker, fall back to the plain loop on any exception, clear the global in finally. Benchmark cold in a fresh process: running serial first warms in-process caches and makes the parallel run look 100x faster (0.3s) falsely.