Table of contents

• Server Runners Comparison
  • Server Runners
    • Gunicorn with UvicornWorker
    • Uvicorn (single process)
    • Uvicorn with --workers
    • FastAPI CLI
  • Measurements
    • Gunicorn thread matrix
    • All runners comparison
    • Uvicorn multiprocess
  • Verdict
  • Versions

Server Runners Comparison

This benchmark compares five ways to run a FastAPI application in production:

• Gunicorn with UvicornWorker — WSGI process manager (1w/2w × 0t/1t/2t)
• Uvicorn — single-process ASGI server
• Uvicorn with --workers — ASGI server with native multi-process mode
• FastAPI CLI (fastapi run --workers) — official CLI wrapping Uvicorn with --workers

All runners use the same underlying ASGI server (Uvicorn). The difference is in how they manage processes and handle startup.

Server Runners

Gunicorn with UvicornWorker

Gunicorn is a mature WSGI HTTP server that runs ASGI applications via the UvicornWorker class. It acts as a process manager, spawning multiple Uvicorn worker processes.

Uvicorn (single process)

Uvicorn is a lightning-fast ASGI server. Running it without --workers uses a single process.

Uvicorn with --workers

Uvicorn natively supports the --workers flag to spawn multiple worker processes. Simpler than Gunicorn but lacks its process management features.

FastAPI CLI

FastAPI CLI (fastapi run --workers N) wraps Uvicorn with --workers.

Measurements

CI run 29815274014 — Python 3.14, Ubuntu latest.

Gunicorn thread matrix

Workers	Threads	Sync RPS	Async RPS
1	0	1819	2228
1	1	1784	2195
1	2	1813	2217
2	0	2654	3415
2	1	2638	3446
2	2	2586	3383

Thread count has negligible impact. Going from 0→1→2 threads changes RPS by <3%, well within noise. The GIL prevents true parallelism for CPU-bound work, and for I/O-bound FastAPI endpoints the async event loop already handles concurrency within each worker.

Doubling workers from 1→2 gives ~46% more sync RPS and ~54% more async RPS.

All runners comparison

Runner	Config	Sync RPS	Async RPS
Gunicorn	1 worker, 0 threads	1819	2228
Gunicorn	2 workers, 0 threads	2654	3415
Uvicorn single-process	—	1485	1760
Uvicorn –workers	2 workers	2203	2739
FastAPI CLI	1 worker	1594	2039
FastAPI CLI	2 workers	2457	3338

Uvicorn multiprocess

Workers	Sync RPS	Async RPS
1	1267	1864
2	1850	2803

Verdict

Individual impact: +50-65% throughput by switching from Gunicorn (1 worker, 0 threads) to FastAPI CLI (2 workers) or Gunicorn (2 workers, 0 threads).

Config	Best Runner	Sync RPS	Async RPS
1 worker, 0 threads	Gunicorn	1819	2228
2 workers, 0 threads	Gunicorn	2654	3415

Gunicorn wins at both 1 and 2 workers. Its process management overhead is negligible and it consistently outperforms all alternatives.

FastAPI CLI is the closest competitor — within 12% of Gunicorn at 1 worker and 6% at 2 workers.

Uvicorn single-process is the slowest — roughly half the throughput of multi-worker setups.

Uvicorn --workers lags Gunicorn by 17-20% despite being pure ASGI.

Thread count is irrelevant — adding threads to Gunicorn changes RPS by less than 3%. Use 0 threads unless you have a specific reason.

Recommendation: Use gunicorn -c gconf.py -k uvicorn.workers.UvicornWorker for production. FastAPI CLI is a good alternative if you prefer simpler configuration.

Re-run the measurements yourself:

git clone git@github.com:KissPeter/fastapi-performance-optimization.git
pip3 install -r test_files/requirements.txt
pytest -vv -rP test_files/ -m server_runners

Versions

Package	Version
Python	3.14
FastAPI	0.139.2
Uvicorn	0.51.0
Gunicorn	26.0.0
orjson	3.11.9
ujson	5.13.0