"""Run independent Monte-Carlo trials across CPU cores. Every trial gets its own random stream spawned from one seed (numpy SeedSequence), so results are identical for any --jobs value and runs are reproducible. Import this module before numpy in a script so the BLAS thread pools stay at one thread per worker process. """ import os for _v in ("OMP_NUM_THREADS", "OPENBLAS_NUM_THREADS", "MKL_NUM_THREADS"): os.environ.setdefault(_v, "1") from concurrent.futures import ProcessPoolExecutor import numpy as np def default_jobs(): return os.cpu_count() or 1 def run_trials(fn, n, seed, jobs=1): """fn(rng) -> result, called once per trial; returns results in trial order.""" seqs = np.random.SeedSequence(seed).spawn(n) if jobs <= 1 or n <= 1: return [fn(np.random.default_rng(s)) for s in seqs] with ProcessPoolExecutor(min(jobs, n)) as ex: return list(ex.map(_call, [fn] * n, seqs, chunksize=max(1, n // (4 * jobs)))) def _call(fn, seq): return fn(np.random.default_rng(seq))