Performance improvements to websockets

[tarasko]

Hi all,
At my current job I'm using aiohttp websockets to subscribe to updates from a 3rd party webserver. The traffic consists of a lot of
small update messages and is very spiky in nature. Also only the very last update usually matters, all previous updates can be often just ignored.

The client is very sensitive to message processing delays. I have already cythonized the critical path in business logic and that gave great results. The bottleneck now is within ClientWebSocketResponse + FlowControlDataQueue
I see a few performance issues there:

1. FlowControlDataQueue internals are private, there is no way to see if more message have been already received and are just waiting in the queue. If there are 20 messages in the queue, I don't need all of them, only the last one.

2. FlowControlDataQueue, can potentially pause reading. There is no way to monitor, control or turn it off.

3. ClientWebSocketResponse.receive setup timeouts, creates and set result on a future even if there are more messages in FlowControlDataQueue. That's unnecessary.

4. I believe cythonizing WebSocketReader would help. parse_frame seems to be a great candidate for cythonization. I would also cythonize and include into distribution a pxd for WSMessage type and relevant enums

Ideally, would be great if I could pass my own WSMessage handler with a simple non-async on_ws_message callback and bypass the whole ClientWebSocketResponse + FlowControlDataQueue logic. Instead of optimizing ClientWebSocketResponse, I would rather have such a low level interface.

Something similar to Protocol.data_received, Protocol.eof_received from asyncio. These are low-level non-async callbacks. Would be nice to have something similar for websockets.

[bdraco]

There have been some improvements in 3.10.x. Some of the above is still quite relevant though.

[tarasko]

Interesting, I haven't heard about mypyc.

Tbh, I'm fine with cython, I like that if something is difficult to implement in cython (due to lack of syntax for example) I can write plain c/c++ code in a separate file and just include it.

One thing that I'd like to look at in the near future is using SSE2/AVX2 instructions for masking frame payload. websockets library actually implemented it, at least for SSE2.

Another one is uvloop. I have an impression that writing side could be a little more efficient than it is now in terms of latencies especially when sending small messages. But I haven't profiled it yet

[tarasko]

So regarding uvloop, I found that just removing this last condition
self._buffer_size > self._high_water
here https://github.com/MagicStack/uvloop/blob/4083a94e5c1c3ab681ffe0b2c0b60fba16540220/uvloop/handles/stream.pyx#L426
increases websocket echo perfromance for aiohttp by 20-25% in my benchmark

So basically instead of caching and sending even small packets later, we send them right away.
I will make a PR for uvloop

[bdraco]

I kinda surprised uvloop was doing that given https://github.com/python/cpython/blob/c13e7d98fb8581014a225b900b1b88ccbfc28097/Lib/asyncio/selector_events.py#L1065

[tarasko]

Yeah, asyncio implementation makes perfect sense to me, simple and concise. I'll try to simplify this part in uvloop

[tarasko]

MagicStack/uvloop#619