httpx.AsyncClient has much worse performance than aiohttp.ClientSession with concurrent requests

[MarkusSintonen]

There seems to be some performance issues in httpx (0.27.0) as it has much worse performance than aiohttp (3.9.4) with concurrently running requests (in python 3.12). The following benchmark shows how running 20 requests concurrently is over 10x slower with httpx compared to aiohttp. The benchmark has very basic httpx usage for doing multiple GET requests with limited concurrency. The script outputs a figure showing how duration of each GET request has a huge duration variance with httpx.

# requirements.txt:
# httpx == 0.27.0
# aiohttp == 3.9.4
# matplotlib == 3.9.0
# 
# 1. start server: python bench.py server
# 2. run client test: python bench.py client

import asyncio
import sys
from typing import Any, Coroutine, Iterator
import aiohttp
import time
import httpx
from aiohttp import web
import matplotlib.pyplot as plt


PORT = 1234
URL = f"http://localhost:{PORT}/req"
RESP = "a" * 2000
REQUESTS = 100
CONCURRENCY = 20


def run_web_server():
    async def handle(_request):
        return web.Response(text=RESP)

    app = web.Application()
    app.add_routes([web.get('/req', handle)])
    web.run_app(app, host="localhost", port=PORT)


def duration(start: float) -> int:
    return int((time.monotonic() - start) * 1000)


async def run_requests(axis: plt.Axes):
    async def gather_limited_concurrency(coros: Iterator[Coroutine[Any, Any, Any]]):
        sem = asyncio.Semaphore(CONCURRENCY)
        async def coro_with_sem(coro):
            async with sem:
                return await coro
        return await asyncio.gather(*(coro_with_sem(c) for c in coros))

    async def httpx_get(session: httpx.AsyncClient, timings: list[int]):
        start = time.monotonic()
        res = await session.request("GET", URL)
        assert len(await res.aread()) == len(RESP)
        assert res.status_code == 200, f"status_code={res.status_code}"
        timings.append(duration(start))

    async def aiohttp_get(session: aiohttp.ClientSession, timings: list[int]):
        start = time.monotonic()
        async with session.request("GET", URL) as res:
            assert len(await res.read()) == len(RESP)
            assert res.status == 200, f"status={res.status}"
        timings.append(duration(start))

    async with httpx.AsyncClient() as session:
        # warmup
        await asyncio.gather(*(httpx_get(session, []) for _ in range(REQUESTS)))

        timings = []
        start = time.monotonic()
        await gather_limited_concurrency((httpx_get(session, timings) for _ in range(REQUESTS)))
        axis.plot([*range(REQUESTS)], timings, label=f"httpx (tot={duration(start)}ms)")

    async with aiohttp.ClientSession() as session:
        # warmup
        await asyncio.gather(*(aiohttp_get(session, []) for _ in range(REQUESTS)))

        timings = []
        start = time.monotonic()
        await gather_limited_concurrency((aiohttp_get(session, timings) for _ in range(REQUESTS)))
        axis.plot([*range(REQUESTS)], timings, label=f"aiohttp (tot={duration(start)}ms)")


def main(mode: str):
    assert mode in {"server", "client"}, f"invalid mode: {mode}"

    if mode == "server":
        run_web_server()
    else:
        fig, ax = plt.subplots()
        asyncio.run(run_requests(ax))
        plt.legend(loc="upper left")
        ax.set_xlabel("# request")
        ax.set_ylabel("[ms]")
        plt.show()

    print("DONE", flush=True)


if __name__ == "__main__":
    assert len(sys.argv) == 2, f"Usage: {sys.argv[0]} server|client"
    main(sys.argv[1])

I found the following issue but seems its not related as the workaround doesnt make a difference here #838 (comment)

[MarkusSintonen]

Found some related discussions:

• Httpx not production ready under load #3100
• httpx vs aiohttp vs requests #3206

Opening a proper issue is warranted to get better visibility for this. So the issue is easier to find for others. In its current state httpx is not a good option for highly concurrent applications. Hopefully the issue gets fixed as otherwise the library is great, so thanks for it!

[tomchristie]

Oh, interesting. There's some places I can think of where we might want to be digging into here...

• A comparison of threaded performance would also be worthwhile. requests compared against httpx, with multithreaded requests.
• A comparison of performance against a remote server would be more representative than performance against localhost.

Possibly points of interest here...

• Do we have the same socket options as aiohttp? Are we sending simple GET requests across more than one TCP packet unneccessarily, either due to socket options or due to our flow in writing the request to the stream, or both? Eg. see https://brooker.co.za/blog/2024/05/09/nagle.html
• We're currently using h11 for our HTTP construction and parsing. This is the best python option for careful spec correctness, tho it has more CPU overhead than eg. httptools.
• We're currently using anyio for our async support. We did previously have a native asyncio backend, there might be some CPU overhead to be saved here, which in this localhost case might be outweighing network overheads.
• Also worth noting here that aiohttp currently supports DNS caching where httpx does not, although not relevant in this particular case.

Also, the tracing support in both aiohttp and in httpx are likely to be extremely valuable to us here.

[MarkusSintonen]

Thank you for the good points!

A comparison of performance against a remote server would be more representative than performance against localhost.

My original benchmark hit AWS S3. There I got very similar results where httpx had a huge variance with requests timings with concurrent requests. This investigation was due to us observing some strange requests durations when servers were under heavy load in production. For now we have switched to aiohttp and it seems to have fixed the issue.

[tomchristie]

My original benchmark hit AWS S3. There I got very similar results [...]

Okay, thanks. Was that also testing small GET requests / similar approach to above?

[MarkusSintonen]

Okay, thanks. Was that also testing small GET requests / similar approach to above?

Yes pretty much, GET of a file with size of a couple KB. In the real system the sizes ofcourse vary alot.

[MarkusSintonen]

We're currently using anyio for our async support. We did previously have a native asyncio backend, there might be some CPU overhead to be saved here, which in this localhost case might be outweighing network overheads.

@tomchristie you were right, this is the issue ^!

When I just do a simple patch into httpcore to replace anyio.Lock with asyncio.Lock the performance improves greatly. Why does httpcore use AnyIO there instead of asyncio? Seems AnyIO may have some issues.

With asyncio:

With anyio:

[MarkusSintonen]

There is another hot spot in AsyncHTTP11Connection.has_expired which is called eg from AsyncConnectionPool heavily. This checks the connection status via this is_readable logic. That seems to be a particularly heavy check.

The logic in connection pool is quite heavy as it rechecks all of the connections every time requests are assigned to the connectors. It might be possible to skip the is_readable checks in the pool side if we just take a connector from the pool and take another if the picked connector was actually not healthy. Instead of checking them all every time. What do you think?

Probably it would be good idea to add some performance tests to httpx/httpcore CI.

[MarkusSintonen]

I can probably help with a PR if you give me pointers about how to proceed :)

I could eg replace the synchronization primitives to use the native asyncio.

[tomchristie]

Why does httpcore use AnyIO there instead of asyncio?

See encode/httpcore#344, #1511, and encode/httpcore#345 for where/why we switched over to anyio.

I can probably help with a PR if you give me pointers about how to proceed

A good first pass onto this would be to add an asyncio.py backend, without switching the default over.

You might want to work from the last version that had an asyncio native backend, although I think the backend API has probably changed slightly.

Docs... https://www.encode.io/httpcore/network-backends/

---

Other context...

• Added AnyIO support httpcore#169 for where/why we added anyio.
• httpcore 0.14 redesign httpcore#420 for when we dropped the asyncio-native backend completly.

[MarkusSintonen]

Thanks @tomchristie

What about this case I pointed:

When I just do a simple patch into httpcore to replace anyio.Lock with asyncio.Lock the performance improves greatly

There switching network backend won't help as the lock is not defined by the network implementation. The lock implementation is a global one. Should we just change the synchronization to use asyncio?

[MarkusSintonen]

I'm able to push the performance of httpcore to be exactly on par with aiohttp:

Previously (in httpcore master) the performance is not great and the latency behaves very randomly:

You can see the benchmark here.

Here are the changes. There are 3 things required to improve the performance to get it as fast as aiohttp (in separate commits):

1. Commit 1. Change synchronization primitives (in _synchronization.py) to use asyncio and not anyio
2. Commit 2. Bringing back asyncio-based backend which was removed in the past (AsyncIOStream)
3. Commit 3. Optimize the AsyncConnectionPool to avoid calling the socket poll every time the pool is used. Also fixing idle connection checking to have lower time complexity for it

I'm happy to open a PR from these. What do you think @tomchristie?

[tomchristie]

@MarkusSintonen - Nice one. Let's work through those as individual PRs.

Is it worth submitting a PR where we add a scripts/benchmark?

[MarkusSintonen]

Is it worth submitting a PR where we add a scripts/benchmark?

I think it would be beneficial to have benchmark run in CI so we would see the difference. Previously I have contributed to Pydantic and they use codspeed. That outputs benchmark diffs to PR when the benchmarked behaviour changes. It should be free for open-source projects.

[tomchristie]

That's an interesting idea. I'd clearly be in agreement with adding a scripts/benchmark. I'm uncertain on if we'd want the extra CI runs everytime or not. Suggest proceeding with the uncontroversial progression to start with, and then afterwards figure out if/how to tie it into CI. (Reasonable?)

[MarkusSintonen]

@tomchristie I have now opened the 2 fix PRs:

• Use asyncio synchronization instead of anyio httpcore#922
• Optimize connection pool implementation httpcore#924

Maybe Ill open the network backend addition after these as its the most complex one.

[HuiDBK]

Maybe you can refer to the implementation of aiohttp
https://docs.aiohttp.org/en/stable/http_request_lifecycle.html#why-is-aiohttp-client-api-that-way
https://stackoverflow.com/questions/78516655/httpx-vs-requests-vs-aiohttp