ProcessPool vs Sequential CPU: Localhost Lab

Intro — what this post promises

When is ProcessPoolExecutor worth it over a plain sequential loop for CPU-bound Python? This lab measures wall time for 16 identical pure-Python tasks (sum-of-squares and a PRNG-style int loop) with worker sweeps 1/2/4/8 on Linux localhost.

It is not another ThreadPool vs ProcessPool GIL bake-off — that lives in the process vs thread pool GIL localhost lab. ThreadPool for I/O is covered in the ThreadPoolExecutor vs sequential localhost lab. Here the only question is: sequential → process pool, and how many workers pay off.

Related links:

• 01shoppercove.comprocess vs thread pool GIL localhost lab
• 02shoppercove.comthreadpoolexecutor vs sequential localhost lab
• 03shoppercove.comcsv reader vs split localhost lab
• 04shoppercove.compickle vs json roundtrip localhost lab
• 05shoppercove.comhashlib md5 vs blake2b localhost lab
• 06shoppercove.comglob vs rglob vs walk localhost lab
• 07shoppercove.comjson dumps compact vs indent localhost lab
• 08shoppercove.comfork COW RSS vs spawn localhost lab

Lab honesty (1 Oct 2026 IST): Python 3.13.5, 8 cores. concurrent.futures.ProcessPoolExecutor only. Affiliates: 0. No Docker. No ThreadPool arms in this post.

Verdict up front (sum_squares, 16 tasks × 2.5M iters): sequential ~1.461 s; 4 workers ~0.45 s (~3.24×); 8 workers ~0.363 s (~4.02×). Worker=1 is slower than sequential (~0.92×) — spawn/IPC tax with no parallelism.

Arms

Lab topology

16 identical tasks · 3 rounds · p50 wall seconds
workers: 1 / 2 / 4 / 8 on 8-core box
metric: speedup = seq_p50 / pool_p50

Script: lab-evidence/80-processpool-vs-sequential/results/run_lab.py.

Lead table — sum_squares (p50 wall)

Second workload — prng_loop

Same shape, longer tasks (less relative spawn tax):

Reading it

• Process pools beat sequential on multi-core when each task is pure-Python CPU long enough to amortize fork/spawn.
• w=1 is a regression — you pay pool overhead with no parallelism. Prefer sequential unless you will scale workers.
• Scaling bends after cores are busy: ~3.2–3.5× at 4 workers, ~4.0–4.5× at 8 on this 8-core box (not linear 8× — expected).
• Do not use ThreadPool for this class of work; GIL keeps threads near 1× (see labs 24/58).

Why not ThreadPool here

ThreadPoolExecutor on the same pure-Python loops would share one GIL and typically sit at ≤1× vs sequential. Lab 58 already measured that story for I/O sleep vs CPU threads. Lab 24 compared Process vs Thread across workload types. This post stays narrow: graduate from a for-loop to ProcessPool, pick a worker count, measure wall time.

Spawn tax and task grain

Short tasks make process pools look bad. Our tasks are multi-million-iteration loops so wall time is dominated by compute, not pickling. If your real jobs are sub-millisecond, batch them before submitting — otherwise sequential or a thread/async I/O path may win for different reasons.

Pitfalls

• Forgetting if __name__ == "__main__" guards when the module is imported under spawn.
• Submitting huge closures/objects — pickling cost can erase speedups.
• Setting max_workers far above CPU count for CPU-bound work (context thrash).
• Assuming process pools help I/O-wait (threads/async usually cheaper there).

Reproduce

python3 lab-evidence/80-processpool-vs-sequential/results/run_lab.py

Evidence: summary.json, summary.txt.

Limits

One Linux box, warm, no Docker. Numbers move with core count, CPU governor, and OS scheduler. Not asyncio, not Ray/Dask. Pure-Python CPU only — C extensions that release the GIL may change the ThreadPool story (still not this post).

Takeaway

For CPU-bound pure Python, ProcessPoolExecutor with workers ≈ cores cut wall time by about ~4× vs sequential on this host. Never use w=1 as a “safe default.” Measure your task grain; keep ThreadPool for wait-bound work.