Re: [hybi] I-D Action: draft-ietf-hybi-thewebsocketprotocol-13.txt

[Tobias Oberstein <[e-mail]>]

yes, it's plain Python, not over strings, but bytearrays, which gives some boost.

Compared to C/C++ the whole thingy is somewhere 3x slower worst-case .. not bad, but slower.

Using NumPy is a nice tip .. I'll have a look at your code.

I've been looking at

http://stackoverflow.com/questions/2119761/simple-python-challenge-fastest-bitwise-xor-on-data-buffers

also, but did not find time yet to see if it helps.

Thanks again,
Tobias

Von: [e-mail] [mailto:[e-mail]] Im Auftrag von Joel Martin
Gesendet: Mittwoch, 7. September 2011 17:32
An: Tobias Oberstein
Cc: Richard L. Barnes; Len Holgate; [e-mail]
Betreff: Re: [hybi] I-D Action: draft-ietf-hybi-thewebsocketprotocol-13.txt

On Wed, Sep 7, 2011 at 8:31 AM, Tobias Oberstein <[e-mail]<mailto:[e-mail]>> wrote:
at least for the Python/Twisted implementation I can say that - even after
quite some optimization experiments - the masking is the bottleneck.

I notice you are doing the unmasking directly using python code (https://github.com/oberstet/Autobahn/blob/master/lib/python/autobahn/websocket.py#L635). Have you tried doing the unmask by offloading it to something like numpy? If not, you might try something like this:

    numpy.bitwise_xor(data, mask)

You'll need to handle modulus 4 data (since numpy.bitwise_xor wants the data to be modulus the mask), but that's the general idea. Here is how I do it in websockify: https://github.com/kanaka/websockify/blob/6e263063c2b63642a0e17cb91d6b9f6c9365c0c0/websocket.py#L293  My main test is noVNC which doesn't send much data from client to server, but I think numpy will do the unmask significantly faster than doing the XOR in python code directly.

Regards,

Joel Martin (kanaka)