[PATCH] performance: disable workaround for an old bug of Python

[Yuya Nishihara]

On Sat, 20 Aug 2016 23:04:06 +0200, Maciej Fijalkowski wrote:
> On Sun, Aug 14, 2016 at 4:14 AM, Yuya Nishihara <yuya at tcha.org> wrote:
> > On Fri, 12 Aug 2016 23:42:38 +0200, Maciej Fijalkowski wrote:
> >> Well, seems the comment is out of date. I know the issue - the GC got
> >> triggered every X objects in 2.6, which caused O(n^2) performance when
> >> allocating lots of objects. These days (since 2.7) it adapts the
> >> threshold, which means it always amortizes to O(n).
> >
> >> On Thu, Aug 11, 2016 at 12:36 AM, Matt Mackall <mpm at selenic.com> wrote:
> >> > It's either fixed in 2.7 or it's not. The primary users of this code are
> >> > dirstate and obsmarkers, so it should be pretty easy to test. This changeset has
> >> > a benchmark:
> >> >
> >> > changeset:   25675:5817f71c2336
> >> > user:        Pierre-Yves David <pierre-yves.david at fb.com>
> >> > date:        Wed Nov 26 16:58:31 2014 -0800
> >> > files:       mercurial/obsolete.py
> >> > description:
> >> > obsstore: disable garbage collection during initialization (issue4456)
> >
> > I can't say there's measurable win on CPython 2.7, and we have native parser
> > now. So I'll take this patch, thanks.
> >
> >   $ hg debugobsolete | wc -l
> >   106437
> >   $ hg up 5817f71c2336  # use pure python parser
> >   $ python -m timeit -r10 \
> >   -s 'from mercurial import obsolete, scmutil; svfs = scmutil.vfs(".hg/store")' \
> >   'obsolete.obsstore(svfs)'
> >
> >   (Python 2.6.9, GC disabled)
> >   10 loops, best of 10: 714 msec per loop
> >   (Python 2.6.9, GC enabled)
> >   10 loops, best of 10: 746 msec per loop
> >
> >   (Python 2.7.12+, GC disabled)
> >   10 loops, best of 10: 699 msec per loop
> >   (Python 2.7.12+, GC enabled)
> >   10 loops, best of 10: 703 msec per loop
> >
> > The result of timeit wasn't stable.
> 
> Well.... you're using TIMEIT to do GC benchmarks - it's a terrible
> idea, ever, but for GC especially. You're taking a minimum of 10 runs
> - of course it'll be random, you should take average instead and do
> more runs.
> 
> That said, this is what I would expect, but I still strongly disagree
> with the methodology.

Good point. At least, I should try to calculate average of many runs.

I thought 100k markers would be large enough to see significant difference
even by timeit, on Python 2.6.9 as described in 5817f71c2336, but it seemed
not.