They are indeed faster, and my first attempt with atomics was extremely fast, I only saw the issues when I had to push harder for correctness and extreme concurrency where my pool ended up allowing multiple processes to use the same connection at the same time.

So I had to lean into the BEAM for correctness which also reduced cache misses, I could properly have kept the original design and just built a queue, but it got to a place where I had to trade simplicity vs complexity.

All that being said, I want to see more of this, since we can get bare metal performance on the BEAM without writing low level code.

This is awesome to see, although be careful about atomics as they can be slow, I had to completely avoid them for my SOTA pool implementation as they got extremely slow due to how schedulers work in conjunction with processes that can move around freely.

If you can keep your atomics on each scheduler then you can avoid global contention.

I just found out that Donald Knuth had his birthday yesterday.
I have to say that a sizeable part of my motivation to do this project was to have more of Knuth algos in Elixir :-).