epoll/kqueue

[medvednikov]

Right now the web server uses simple blocking IO. This results in terrible performance: about 1k requests/second.

It should use epoll/kqueue instead, that should bring it to 50-100k.

[vitalyster]

+HTTPAPI :) https://docs.microsoft.com/en-us/windows/win32/http/http-server-sample-application

[medvednikov]

Of course, we can't ignore Windows Server users :)

[S-YOU]

Once #1469 / #1470 or similar approach is in, you could try to test picoev, which has built-in support for epoll/kqueue/select

https://github.com/S-YOU/picoev.v/blob/master/examples/http_hello.v

Which I get 200k reqeusts/second on my machine, for following code.
And with linux PORT REUSE features, you can scale out horizontally on each CPU core with taskset, and get 200k requests / second / core as long as kernel can handle it.

https://mobile.twitter.com/__s_you/status/1157683017357250560

import picoev

pub fn callback(req byteptr) string {
	return 'Hello, World!'
}

pub fn main() {
	picoev.new(8080, &callback).serve()
}

And hyper(rust) is running par with picoev.v on my machine, so it is basically can max out performance as much as it can be like on this benchmark.

https://www.techempower.com/benchmarks/#section=data-r18&hw=ph&test=plaintext

[spytheman]

@S-YOU Good work!

[spytheman]

Wow, on my old i3, I get over 100K req/s ! There is a memory leak, but still...

[S-YOU]

Great!, and you could probably get better with less threads (other than -t 10) on wrk side unless you have more than 10 physical cpus, because wrk may probably use same cpu that is in use by web server, it could slow things down, imo.

And I have no idea which cause the memory leak :D
(may be returning v string? which may be using heap?)

[spytheman]

Using taskset definitely helped the average latency (my processor has 2 cores and fixing wrk and the http_bench process on their own cores is a good idea, 10x!) ...