search

uNetworking / uSockets

Miniscule cross-platform eventing, networking & crypto for async applications
Apache License 2.0
1.29k stars 267 forks source link

New backend using io_uring #129

Closed ng5 closed 3 years ago

ng5 commented 4 years ago

io_uring - New asynchronous IO eventing added since linux kernel 5.1. Could be a possible enhancement given uSockets has already abstracted event polling backends.

Some performance benchmarks compared to epoll Provided by GitHub user https://github.com/frevib

https://github.com/frevib/io_uring-echo-server/blob/io-uring-op-provide-buffers/benchmarks/benchmarks.md

ghost commented 4 years ago

https://github.com/frevib/io_uring-echo-server/issues/8

ghost commented 4 years ago

It is said Linux 5.11 will have optimizations to io_uring so rerunning benchmarks then can be a good idea.

ghost commented 4 years ago

There are many disadvantages to io_uring, esp. when dealing with tons of mostly-idle connections like with websockets. It's the same disadvantages as Windows IOCP has - mainly that you have to bind tons of user space memory to something that may happen in the future. This has massive disadvantages to memory usage and goes against much of what the point of uWS is. Further on the supposed performance improvements have been very mild, if none at all.

IO_uring should still be experimented with and assessed, but it should not be added just because of the hype. It should only be added if a significant, scientific gain can be shown (this, I'm still waiting for). We don't do things out of hype here - this is why we skip things that are hyped like sendfile and writev when the gains are impractical or none, for the core use cases.

Nobody has really presented anything that can outperform uWS either way, so we don't really need to respond to hype just because it is hyped. I don't care about new things unless there is real, scientific gain that significantly outweighs the disadvantages.

ghost commented 4 years ago

It can also be said that we kind of already do support IO_uring since we support libuv - maybe not an ideal support but we definitely do support it (can always be improved, but still).

ghost commented 3 years ago

io_uring is slower than epoll.

LifeIsStrange commented 3 years ago

@alexhultman you should reconsider, as you know from the latests comment of your own issue on liburing, on kernel 5.12 (and even 5.11?) the performance is better vs plain epoll, even with mitigations=off

Also, I'd like to thank you for your efforts on rigorous thinking and benchmarking. This is a great example of applied rationalism and I'm sure you would like websites such as lesswrong.com Learning cognitive biases would probably interest you too https://en.m.wikipedia.org/wiki/List_of_cognitive_biases https://en.m.wikipedia.org/wiki/List_of_fallacies I'm becoming off topic but I would like to have debates with you on interesting intellectual topics as you are one of the rare examples of truth seekers. In order to do so, if you're potentially interested, please create an empty issue on this repository with your name. :) https://github.com/LifeIsStrange/CollectionOfInterestingBrains/issues

ghost commented 3 years ago

I first need a benchmark that shows a significant gain on my computer otherwise it is impossible for me to develop for this API. I need to be able to benchmark as I go to know if I am doing it correctly, can't just develop and hope it will run fast when done.