Python Concurrency on macOS: Threads vs Async vs Multiprocessing under the GIL

CS263 Project by Xinghan Yang

Specs

Machine: Apple Silicon (a.k.a. my laptop)
Endpoint: POST /score → request {"mode":"cpu"| "GIL", "TBD"} → response
Deliverables: mini-survey of GIL & models, reproducible results with scripts, visualization of result, 5-page PDF, 10–15 min video, slides
Weekly Progress check Commits: every Fri ~5pm PT
Final due: Tue Dec 9, 11:59pm PT

Overview, Goals, Questions, Hypotheses

Goal: Show how Python concurrency models behave on Apple Silicon for two types of workloads:

CPU-bound (SHA-256 chaining, pure Python)
I/O-bound (SQLite KV GET).

Research questions:

Throughput & p50/p95/p99 latency across threads / asyncio / multiprocessing
Does async improve tail latency at higher concurrency vs threads? Hypotheses: CPU-bound → threads/async plateau due to GIL; multiprocessing wins for moderate+ task sizes. I/O-bound → asyncio best tail latency at high concurrency; threads competitive at moderate; mp overkill. Memory → threads ≲ asyncio ≪ mp. Energy (opt.) → mp highest under CPU; async most efficient when I/O-wait dominated.

Repository Structure

#TODO

API contract: #TODO

Tools & Setup

Python & libs: Python 3.12+ (Apple Silicon), Flask (threads + mp), aiohttp (async), optional uvloop, sqlite3 / aiosqlite, orjson (fast JSON), matplotlib, pandas.
Bench & system tools: wrk (primary), hyperfine (sanity), /usr/bin/time -l (RSS/page faults), optional powermetrics (energy), jq (CLI JSON).
Install:

brew install wrk hyperfine jq
python3 -m venv .venv && source .venv/bin/activate
pip install flask aiohttp uvicorn uvloop orjson aiosqlite matplotlib pandas

System Design (Our code)

Interchangeable backends:

Workloads:

CPU: SHA-256 chaining, tunable iters ∈ {80k, 150k, 300k}
I/O: SQLite SELECT val FROM kv WHERE key=? (1k? keys, random access)

Fairness/Environment controls for reproducibility:

Result CSV schema: TBD

Evaluation Plan (Tbale/Plots): TBD

Timeline (Week 6 → Week 9, final due Dec 9)

Rule: push at leaset one substantial commit by every Friday ~5pm PT. Week 5 - Beginning:

Brainstormed the overall plan for the project
Narrow down the scope of work and chose the Tech Stack
Finished the README with a clear abstract and timeline
Configure the environment and prototyped the minimal experiment

Week 6 — Foundations:

Scaffold repo , add more to README.md, requirements.txt (env)
Implement one mode (threads or async) with score/metrics (CPU + I/O)
db/setup_db.py to build db/test.db
Smoke tests: baseline
Deliverables : minimal server running, DB ready, draft bench scripts, baseline p50/p95/p99 snippet

Week 7 - Backends & First Benchmark:

Add remaining modes: threads, asyncio, mp (unified API, mode switch)
Launch scripts: run_threads.sh, run_async.sh ..
Bench each mode: Get the result with metrics
Deliverables (Fri Nov 7): all modes runnable; first benchmark with CPU+I/O different modes(GIL/..

)

Week 8 - TBD:

Week 9 - TBD:

Week 10~Finals - TBD:

SexyCarrots/CS263

README