consider dropping wsaccel dependency

[belm0]

wsaccel doesn't appear to be maintained, and has broken build under pypy

This article describes a pure Python mask implementation which is near the performance of wsaccel:

https://www.willmcgugan.com/blog/tech/post/speeding-up-websockets-60x/

[Kriechi]

Amazing results! wsproto is already py3 only, so I would suggest using "Solution 3" with this snippet:

native_byteorder = sys.byteorder
def mask3(mask, data):
    datalen = len(data)
    data = int.from_bytes(data, native_byteorder)
    mask = int.from_bytes(mask * (datalen // 4) + mask[: datalen % 4], native_byteorder)
    return (data ^ mask).to_bytes(datalen, native_byteorder)

PRs are welcome!

[belm0]

It's a little confusing from the article, but solution 4 (using bytearray.translate()) is the fastest solution in Python. The author was merely looking for a python2 solution-- the twist was that he didn't expect it to be faster than python3 from_bytes().

[belm0]

I measured the bytearray.translate() implementation to actually be about 10% faster than wsaccel.

https://nbviewer.jupyter.org/gist/belm0/2b610cc405dd3dac977f34650cb187cc

[njsmith]

The translate-based solution does have some tradeoffs though: it adds something like 100 KB of extra ram usage for the lookup table, and several milliseconds of startup time to compute that table (on a fast desktop CPU, more on a rasbpi or underpowered arm router or whatever).

So I'm also leaning towards "solution 3". It's still like 50x faster than what we have right now, and doesn't have the risk of creating some pathological behavior in constrained environments.

[belm0]

Fair point about memory and startup time. Though "what we have now" is wsaccel, which is 2x faster than solution 3. So perhaps stay with wsaccel then.

[njsmith]

Note that this discussion started over here: python-trio/trio-websocket#119

See in particular python-trio/trio-websocket#119 (comment) and below, which discuss the up- and down-sides of implementing our own accelerator. (Basically: upside is that it's pretty easy to get 10x faster than wsaccel or the pure python stuff above; downside is then you have a C accelerator module to maintain.)

[njsmith]

...though, ugh, I guess wsaccel doesn't even have wheels, so right now by depending on it, trio-websocket forces everyone to have a compiler installed. (This is particularly problematic on windows.)

[bluetech]

The int.from_bytes() version has the disadvantage over the others of using 4x peak memory. Given 10 MiB of data, the peak memory is:

import tracemalloc

mask = b'\0\1\2\3'
n = 10 * 2**20
data = b'\1' * n

tracemalloc.start()
mask3(mask, data)
print(tracemalloc.get_traced_memory()[1] // 2**20, 'MiB')

• mask1: 11 MiB
• mask2: 10 MiB
• mask3: 42 MiB
• mask4: 20 MiB

The optimal result is 0 MiB, i.e. doing the masking in-place, which should be allowed given payload in payload = self.masker.process(payload) is a bytearray (unlike what the type of Buffer.consume_at_most() says).

BTW, the blog post uses data = b'\0' * data_size to generate the test data; but somewhere in mask3 there seems to be a special case for all-zeros. Changing to \1 makes it 25% slower for me (Python 3.7.3, Linux).

[bluetech]

BTW2, the entire masking business is unfortunate. It is meant to protect against bad proxies, but probably 99% of WebSocket traffic in the open internet is over TLS, which makes it unnecessary. If I had control over all WebSocket clients in the world, I would just make them use \x00\x00\x00\x00 as the mask key over wss, and the servers would detect that and skip the pointless exercise.

[pgjones]

I don't think wsaccel is (or has been) an explicit dependency of wsproto, I've added it as an extra requirement in #133 and implemented solution 4 as that is what aiohttp uses. Does #133 provide a sufficient solution for this?

[willmcgugan]

You could save the startup time by pregenerating that lookup table, and inserting a literal...

[belm0]

implemented solution 4 as that is what aiohttp uses

Doesn't the article say that aiohttp uses solution 3?

[pgjones]

It was solution 3 at the time the article was written, but now it is 4 thanks to the author of the article.

[willmcgugan]

Something to consider with the translate method is that it has to convert to and from a bytearray if the input and output is bytes.

The XOR part will still take the majority of the time, but if the payload data were a bytearray to begin with, it can avoid some extra copying. It does become more significant with smaller payload sizes.