Why are Python sockets so slow and what can be done?

[adontz]

@njs

I was all set to write something here about how this just wasn’t realistic, but then I did the math and… maybe?

On my laptop, the current epoll loop gets 16k req/sec, so each request takes ~60 µs. If you want to stay above 10k req/sec, then you have a budget of 100 µs for each request. So that leaves 40 µs to implement a friendly, full-featured async library API. A dict lookup is ~0.3 µs; an empty method call is ~1 µs; suspending/resuming a coroutine call stack is ~1-2 µs. So… it’s super tight, but maybe doable?

You overcomplicated. My actual target is easier. 10K RPS on let’s say four cores with SO_REUSEPORT. It’s just 2500 RPS per core, thus 400µs for each request. If friendly full-featured async library API fits into 100µs we have 300µs more for everything else. It’s not easy, but a bit more than maybe