can't reproduce benchmark, get only 10% of RPS using bench/socketify_plaintext.py

[raceychan]

Describe the bug
A clear and concise description of what the bug is.

To Reproduce
Steps to reproduce the behavior:

1. copy socketify_plaintext.py file from the main branch
2. uv add socketify
3. uv run python sock.py

Expected behavior
on benchmark it shows that even the slowest way to run socketify through python3, should reach 500K RPS, however, I only get 73K RPS.

import os

from socketify import App

def run_app():
    app = App(request_response_factory_max_items=200_000)
    router = app.router()

    @router.get("/")
    async def home(res, req):
        res.send(b"Hello, World!")

    app.listen(
        8000,
        lambda config: print(
            "PID %d Listening on port http://localhost:%d now\n"
            % (os.getpid(), config.port)
        ),
    )
    app.run()


def create_fork():
    n = os.fork()
    # n greater than 0 means parent process
    if not n > 0:
        run_app()


# fork limiting the cpu count - 1
# for i in range(1, multiprocessing.cpu_count()):
#     create_fork()
if __name__ == "__main__":
    run_app()  # run app on the main process too :)

wrk -t4 -c64 http://localhost:8000

Running 10s test @ http://localhost:8000
  4 threads and 64 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     0.87ms  221.20us  11.43ms   96.59%
    Req/Sec    18.56k     1.36k   32.00k    79.75%
  738552 requests in 10.06s, 97.90MB read
Requests/sec:  73406.05
Transfer/sec:      9.73MB

compared to a naive asgi

from starlette.responses import Response

async def naive(scope, receive, send):
    content =  "hello, world".encode()
    await Response(content)(scope, receive, send)

wrk -t4 -c64 http://localhost:8000
Running 10s test @ http://localhost:8000
  4 threads and 64 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     0.91ms  840.99us  35.45ms   99.39%
    Req/Sec    18.10k     0.95k   29.63k    93.03%
  724043 requests in 10.10s, 35.22MB read
Requests/sec:  71691.43
Transfer/sec:      3.49MB

Desktop (please complete the following information):
Distributor ID: Ubuntu
Description: Ubuntu 20.04.6 LTS
Release: 20.04
Codename: focal

[cirospaciari]

what is your hardware? python version, pypy version? I would stay with TechEmPower benchmarks and use they hardware to see performance.

[cirospaciari]

have no Ideia where did you find that numbers and the ideia that the libuv is the fastest async IO lib, or that every single implementation using it would have the same performance or worse.

[cirospaciari]

Any information about the benchmarks take a look in https://github.com/TechEmpower/FrameworkBenchmarks there is the code, how to run and even live results using the hardware provided by TechEmPower: https://tfb-status.techempower.com/

[raceychan]

have no Ideia where did you find that numbers and the ideia that the libuv is the fastest async IO lib, or that every single implementation using it would have the same performance or worse.

There is no guarantee that libuv is the fastest asyncio IO lib, I should say that it is the fastest I have seen, and from my tests as well as others tests, as I paste the benchmark above, all asyncio libs, across languages, seem to have performance lower than 150k RPS, so this is why I am so surprised to see there is something faster than that.

https://github.com/uNetworking/uSockets/blob/master/src/internal/eventing/libuv.h

It seems that usockets uses libuv internally, wouldn't it be a safe assumption that usockets can only add abstractions to it and therefore being slower?

[cirospaciari]

https://github.com/uNetworking/uSockets/blob/master/src/internal/eventing/libuv.h

You are also ignoring this part: https://github.com/uNetworking/uSockets/blob/master/src/internal/eventing/epoll_kqueue.h
again no ideia where the 150K RPS number is coming from, like I said before the benchmarks depends on multiple variables, like hardware, operation system, load/pipelining. If is there a problem with TechEmPower numbers please open an Issue in they repo, running they tools locally on your hardware properly will compare all frameworks that you wanna using the same script and you will see a number that matchs your hardware, if you run it in other hardware the numbers will be different, just running worker without pipelining, without the same payloads dont necessary will translate in same numbers that in Round 22 or Round 23 because is hardware dependent and they use a custom script to properly call the benchmark in the same way with they own configurations.

[raceychan]

admittedly, I do not have access to techempower's hardware,I test using the dockerfile from their repo
https://github.com/TechEmpower/FrameworkBenchmarks/blob/master/frameworks/Python/socketify.py/socketify.py-asgi.dockerfile
with my local machine, so it is true that it might make some significant difference there.

Thanks for you replies, I know it is unfair to say nothing can't pass 150K RPS just because I can't find anything pass that number by testing on my local machine, I understand your point, thank you for having this conversation with me.

[cirospaciari]

If you wanna have a more dynamic call with me feel free to call me on discord, if you feel any number is miss represented in any form let me know.