_futures.py possible performance improvement

[cirospaciari]

@gi0baro taking a loop on _future.py In:
https://github.com/emmett-framework/granian/blob/master/granian/_futures.py

the code is :

import asyncio


def future_wrapper(coro, watcher):
    fut = asyncio.ensure_future(coro)
    fut.add_done_callback(future_handler(watcher))


def future_handler(watcher):
    def handler(task):
        try:
            task.result()
        except Exception:
            watcher.err()
            raise
        watcher.done()
    return handler
    

In my tests add_done_callback is really slow, in socketify I change it to do something like this:

import asyncio

async def task_wrapper(task, watcher):
    try:
        await task
    except Exception:
        watcher.err()
        raise
    watcher.done()

def future_wrapper(coro, watcher):
    return asyncio.create_task(task_wrapper(coro, watcher))

for PyPy I used a custom task in Pure python (PyPy optimizes a lot of pure Python code) and added a task factory, to reuse tasks instead of GC and create new ones every time.
In CPython I did not use the factory strategy because pure python is slower than the built-in, but I will move to the same strategy after migrating to HPy. I believe that you can create a custom task (like I did) but make use of PyO3 to optimize to CPython and use pure Python for PyPy.

You can see my code (that needs a refactor lol) here: https://github.com/cirospaciari/socketify.py/blob/main/src/socketify/loop.py

PS: this boosted the performance on CPython for socketify by 5% to 20% depending on the test and for PyPy boost up to 50%

Comparing using TechEmPower:

Before:
falcon [socketify.py asgi] | 286,395
falcon [socketify.py asgi pypy3] | 464,546

After:
falcon [socketify.py asgi] | 327,481
falcon [socketify.py asgi pypy3] | 622,844

Incomplete run:
https://www.techempower.com/benchmarks/#section=test&runid=9bb6e0fb-1619-447c-b5f4-572207cfbc72&test=plaintext&l=hra0hr-35r

After optimizations (complete run):
https://www.techempower.com/benchmarks/#section=test&runid=7ce481b2-49ec-4a4d-952d-bb1334d4a4ad&test=plaintext&l=hra0hr-35r

I did not open a PR because I did not know how to compile and test locally, but I'm happy to help. I open a PR #31

[cirospaciari]

I'm pretty sure that with this improvement you can beat uvicorn in CPython

Current performance from https://www.techempower.com/benchmarks/#section=test&runid=7ce481b2-49ec-4a4d-952d-bb1334d4a4ad&test=plaintext&l=hra0hr-35r:

uvicorn | 334,110
granian [asgi] | 322,211

even a 5% increase will get you 338k (a little edge) to 386k (in the best scenario)

Great job optimizing ASGI by the way.
221k to 322k in the last updates is an amazing improvement.

[cirospaciari]

Testing locally with

wrk -H "Host: tfb-server" -H "Accept: text/plain,text/html;q=0.9,application/xhtml+xml;q=0.9,application/xml;q=0.8,*/*;q=0.7" -H "Connection: keep-alive" --latency -d 5 -c 4096 --timeout 8 -t 8 http://localhost:8000/ -s pipeline.lua -- 16

I got 85k before and after I got 100k req/s ~17% (between 5 to 20% like I said)

[cirospaciari]

@gi0baro really happy to see the last update on benchmarks and see granian ASGI winning over uvicorn http-tools 🚀 ❤️

[cirospaciari]

@gi0baro I currently create a custom Task with I will commit today, is pure Python and perform 10% to 30% better then the default Task for CPython and PyPy, with little to no code this can gives Granian another impressive performance boost, I will do some benchmarks and create an PR

[cirospaciari]

@gi0baro #63
About 21% improvement, is really a copy past PR using code from socketify but I hope it helps you.