Enable TCP_NODELAY for all synchronous sockets.

[plietar]

The widely documented poor interaction between the Nagle algorithm and TCP's delayed ACK feature, when making short successive writes, leads to unnecessary delays (around 50ms on Linux). This happens in httpcore whenever a POST request is made, since the headers and body are sent as two separate calls.

The TCP_NODELAY can be enabled to disable Nagle's algorithm, and therefore avoid this delay.

The option is already enabled by default for all asyncio and Trio sockets. It is also enabled by CPython's http.client module (which urllib and requests use) and by many high-level HTTP libraries found in other languages, including libcurl, Java's Netty, Rust's reqwest and Go's standard library, to name a few.

[lovelydinosaur]

Thanks, seems like this might be worth review.

I wasn't aware that urllib3 uses TCP_NODELAY by default. Are you able to point me to how I can verify this / where this is documented / where in the code stack this gets set?

[zanieb]

There was some discussion about this recently in Go; golang/go#57530 and https://withinboredom.info/blog/2022/12/29/golang-is-evil-on-shitty-networks/

[plietar]

This is the relevant bit in urllib3: https://github.com/urllib3/urllib3/blob/0b1188985b747ae50d7d4f9750f4e5ec46462517/src/urllib3/connection.py#L107https://github.com/urllib3/urllib3/pull/283/files

Similar for http.client (which I mistakenly claimed was used by requests): https://github.com/python/cpython/blame/94fc7706b7bc3d57cdd6d15bf8e8c4499ae53a69/Lib/http/client.py#L955

[lovelydinosaur]

One option here would be to start by adding a socket_options config to the ConnectionPool and proxy classes without changing the behaviour. This could be an incremental step that wouldn't initially introduce any behaviour change, and that would still leave us with the option of updating the defaults.

https://www.encode.io/httpcore/connection-pools/#other-options

I'm also curious if we'd be able to demonstrate a difference in performance with TCP_NODELAY enabled or not.

[plietar]

Here's a sample code to measure the performance:

importhttpximportrequestsimportasynciofromtimeimportperf_counterN=100URL="http://127.0.0.1:8000"defsync_version():
withhttpx.Client() asclient:
before=perf_counter()
foriinrange(N):
client.post(URL, content=b"Hello")
after=perf_counter()
return (after-before) /Nasyncdefasync_version():
asyncwithhttpx.AsyncClient() asclient:
before=perf_counter()
foriinrange(N):
awaitclient.post(URL, content=b"Hello")
after=perf_counter()
return (after-before) /Ndefrequests_version():
withrequests.Session() assession:
before=perf_counter()
foriinrange(N):
session.post(URL, data=b"Hello")
after=perf_counter()
return (after-before) /Nprint("Sync:", sync_version())
print("Async:", asyncio.run(async_version()))
print("Requests:", requests_version())

Without this change, and running against a local server, I get average request durations on the order of 2-3ms for the async and requests version, and 43ms for the sync one. After applying the change from the PR, the request duration for the sync version is about 2ms.

Here's the server code I used. I also get similar results with a Go server. Note that I've set TCP_NODELAY here, as otherwise it adds another 40ms across the board.

fromhttp.serverimportHTTPServer, BaseHTTPRequestHandlerimportsocketPORT=8000classHandler(BaseHTTPRequestHandler):
protocol_version="HTTP/1.1"defdo_POST(self):
length=int(self.headers['Content-Length'])
self.rfile.read(length)
self.send_response(200)
self.send_header('Content-Length', '14')
self.end_headers()
self.wfile.write(b"Hello, World!\n")
withHTTPServer(("", PORT), Handler) ashttpd:
print("serving at port", PORT)
httpd.socket.setsockopt(socket.IPPROTO_TCP, socket.TCP_NODELAY, 1)
httpd.serve_forever()

[zanieb]

@plietar it looks like #662 is the next step before considering changing the default.

[lovelydinosaur]

The motivation in #651 (comment) is compelling here.

Would we prefer to accept this PR as-is, or should we make the socket options configurable first? (#662, resolved via #668)

rephrased: Is having configurability actually desirable, or do we really just want to make this one change?

[lovelydinosaur]

i think we oughta okay this, yup.

needs a note in the changelog then it should be good to go.

[lovelydinosaur]

Actually can be firmer on this one.
Yes, urllib3 do this and we should too.

Thanks for nudging us on it, @plietar.

[lovelydinosaur]

Link to asyncio...

In asyncio TCP sockets on Linux are now created with TCP_NODELAY flag set by default.

In the 3.7 release notes.

Link to trio...

By default for TCP sockets, :class:SocketStream enables TCP_NODELAY,
and (on platforms where it's supported) enables TCP_NOTSENT_LOWAT with
a reasonable buffer size (currently 16 KiB) – see issue #72 <https://github.com/python-trio/trio/issues/72>__ for discussion. You can
of course override these defaults by calling :meth:setsockopt.

See the docs, and the implementation.