[Twisted-Python] RE: Question about Descriptors and SelectReactor

[andrewfr_ice at yahoo.com]

Hi Glyph and Folks:

AM>I would be actually quite curious to know the rationale of choosing
AM>select() over epoll() these days. epoll() scales like O(1) with the AM>number of file descriptors, it is very performant, stable, and has no AM>limitation on the overall number of fds on linux (except for your /proc AM>and ulimit -n settings). I'd use epoll reactor, unless you have some AM>very specialized requirement.

G>Short answer: because select() is always available.

Thanks. I do most development in Linux now.

The main thing I wanted to avoid was altering file descriptor limits and bundling a new AMI and still not having the ability to test above a certain descriptor limit.

After the semester is over, I want to conduct some new tests with Stackless Python and Twisted. I am also playing with new designs. I am interested in comparing Stackless/Twisted solutions to pure Twisted solutions (just to gauge of inefficiencies). 

I am particularly interested in how channel preferences (a Stackless scheduling mechanism) affect certain performance metrics depending on the number of (and I guess duration) of connections. I will probably start looking at Reactor code again. This is why in the donate time thread, I was willing to volunteer time there (if I can actually be of help)

Cheers,
Andrew