Slower performance with --disable-gil on threaded nearest-neighbor benchmark (10M points)

[basileMarchand]

Bug report

Bug description:

Summary

I am currently benchmarking the performance of the experimental --disable-gil mode in Python 3.14b2.
In most CPU-bound cases, I observed excellent improvements (×3–×4 speedups). However, in one more realistic test case, performance significantly drops when using --disable-gil.

This issue is not really a bug report, but a performance feedback and an invitation to discuss whether this behavior is expected.

---

Reproducible test case

Nearest neighbor search using few threads over a list of 10 million 3D points (list[list[float]]).

The algorithm:

• splits the search range evenly across threads
• uses a shared dict to store the result with a threading.Lock
• each thread compares distance2(p, query_point) and updates shared data if needed

Repository with code and Dockerfile:

https://github.com/basileMarchand/benchmark_python3.14_nogil

Benchmark article:

https://dev.to/basilemarchand/benchmarks-of-python-314b2-with-disable-gil-1ml3

---

🔢 Results

Mode	Elapsed Time
Python GIL ON	2.38 s
Python GIL OFF	3.61 s ❗
C++ + threads	0.88 s ✅

---

Question

This test is the only case in my benchmarks that runs slower with --disable-gil, even though it is purely CPU-bound.

Is this performance drop expected with current implementation of nogil?

Any hints on what could explain it? (lock contention, list access overhead, memory locality, etc.)

---

Environment

• Python: 3.14.0b2, compiled with --disable-gil
• Platform: Docker / Linux (see repo for details)
• Memory: 10M Python objects (list of lists)

Thanks for your amazing work on nogil — it’s a huge leap forward 🚀

CPython versions tested on:

3.14

Operating systems tested on:

Linux

[hugovk]

This issue is not really a bug report, but a performance feedback and an invitation to discuss whether this behavior is expected.

You're better off opening a discussion at https://discuss.python.org/, more people will see it.

[basileMarchand]

You're better off opening a discussion at https://discuss.python.org/, more people will see it.

@hugovk thanks for the suggestion, I have just opened the discussion on discourse

https://discuss.python.org/t/slower-performance-with-disable-gil-on-threaded-nearest-neighbor-benchmark-10m-points/95365

[kumaraditya303]

Using list across multiple threads doesn't scale well, if you change it to use tuple it would perform much better as there would be no contention from reading the points.