[Psyco-devel] Re: [Zope-Coders] Psyco release 0.4.0

[Armin Rigo]

Hello Petru,

----- Original Message -----
> > Jup, Zope is about 3.5 times slower *with* psyco.
> > Any explanations for this worse behaviour ?
>
> Just a wild, wild guess: this could be related to Psyco not doing the
> regular interval polling, so PyThreadState_Swap() doesn't get called as
> often as it should, ZServer piles up requests, the worker threads get
> starved, and so on.

Morten's results give a more interesting answer. He went from a serious
slowdown to a 2x speed-up by replacing psyco.jit() with psyco.bind() on a
pair of classes. So the problem probably stands in the just-in-time
rebinder, psyco.jit(). Not only does it add a constant profiling-like
overhead (significantly more important in Python 2.1 than 2.2, by the way),
but I believe that there are quite a few problems with its current approach.

Would someone please make another check on Zope+Psyco without using
psyco.jit() ? A few carefully selected psyco.bind() could give quite more
positive speed-ups. I don't know exactly *how* to select what to bind,
however. Maybe testing various combinations and relating them with data from
profiling could give a hint at more general usage patterns that could be
eventually formalized in a more involved implementation of psyco.jit().


Thanks,

Armin