Support for method cache optimization

[Yoshinori K. Okuji]

Python 2.6 has method cache optimization:

  http://bugs.python.org/issue1700288

but it is not enabled automatically in extensions:

  http://bugs.python.org/issue1878

So I attach a patch for the support in ZODB. The patch is made against the trunk (r105930). All the unit tests run successfully both before and after my modification.

[Jim Fulton (jim-zope)]

On Sat, Nov 21, 2009 at 4:06 AM, Yoshinori K. Okuji <email address hidden> wrote:
> Public bug reported:
>
> Python 2.6 has method cache optimization:
>
>  http://bugs.python.org/issue1700288
>
> but it is not enabled automatically in extensions:
>
>  http://bugs.python.org/issue1878
>
> So I attach a patch for the support in ZODB. The patch is made against
> the trunk (r105930). All the unit tests run successfully both before and
> after my modification.

Thanks. Have you done any benchmarking of this?

Jim

--
Jim Fulton

[Yoshinori K. Okuji (okuji)]

I ran ZODB's speed.py (a little modified, since cache_deactivate_after doesn't exist any longer) and collective.zodbbench, but they were affected by noise of disk I/O very much, so I didn't see much difference, although I am sure that my changes didn't decrease the performance.

I would like to see a bit more realistic test, if any. Do you have any recommendation?

[Jim Fulton (jim-zope)]

On Mon, Nov 23, 2009 at 5:50 AM, Yoshinori K. Okuji <email address hidden> wrote:
> I ran ZODB's speed.py (a little modified, since cache_deactivate_after
> doesn't exist any longer) and collective.zodbbench, but they were
> affected by noise of disk I/O very much, so I didn't see much
> difference, although I am sure that my changes didn't decrease the
> performance.
>
> I would like to see a bit more realistic test, if any. Do you have any
> recommendation?

I have a realistic benchmark that involves actual database data.
It uses an experimental branch, replay, of zc.zeoinputlog. This is a work
in progress, so only especially hardy souls should try to use this package
themselves. :)

When I get a chance, I'll use my setup to evaluate your patches.

Jim

--
Jim Fulton

[Hanno Schlichting (hannosch)]

We enabled the method cache optimization in all the "Zope 2 C extensions" like ExtensionClass and Acquisition and used the code for some months now.

There was no measurable performance difference with this change, but also no ill side effects. The claims in the original Python tracker issue of speedups between 5% and 20% are only true for synthetic benchmarks.