Patchwork [1,of,2] patchcopies: give up any optimization based on `introrev`

login
register
mail settings
Submitter Pierre-Yves David
Date Oct. 10, 2019, 11:04 p.m.
Message ID <2477ba483c04067900d1.1570748652@nodosa.octobus.net>
Download mbox | patch
Permalink /patch/42213/
State Accepted
Headers show

Comments

Pierre-Yves David - Oct. 10, 2019, 11:04 p.m.
# HG changeset patch
# User Pierre-Yves David <pierre-yves.david@octobus.net>
# Date 1570672173 -7200
#      Thu Oct 10 03:49:33 2019 +0200
# Node ID 2477ba483c04067900d1e9f6523b03df68a4d545
# Parent  8ff1ecfadcd110849c47c77e31c92809eea466ab
# EXP-Topic patchcopies-regression
# Available At https://bitbucket.org/octobus/mercurial-devel/
#              hg pull https://bitbucket.org/octobus/mercurial-devel/ -r 2477ba483c04
patchcopies: give up any optimization based on `introrev`

Between 8a0136f69027 and d98fb3f42f33, we sped up the search for the
introduction revision during path copies. However, further checking show that
finding the introduction revision is still expensive and that we are better off
without it. So we simply drop it and only rely on the linkrev optimisation.

I ran `perfpathcopies` on 6989 pair of revision in the pypy
repository (`hg perfhelper-pathcopies`. The result is massively in favor of
dropping this condition. The result of the copy tracing are unchanged.

Attempt to use a smaller changes preserving linkrev usage were unsuccessful, it
can return wrong result. The following changesets broke test-mv-cp-st-diff.t

    -        if not f.isintroducedafter(limit):
    +        if limit >= 0 and f.linkrev() < limit:
                 return None


Here are various numbers (before this changeset/after this changesets)

            source       destination  before      after    saved-time  ratio
worth cases e66f24650daf 695dfb0f493b 1.062843 1.246369 -0.183526 1.172675
            c979853a3b6a 8d60fe293e79 1.036985 1.196414 -0.159429 1.153743
            22349fa2fc33 fbb1c9fd86c0 0.879926 1.038682 -0.158756 1.180420
            682b98f3e672 a4878080a536 0.909952 1.063801 -0.153849 1.169074
            5adabc9b9848 920958a93997 0.993622 1.147452 -0.153830 1.154817
worse  1%   dbfbfcf077e9 aea8f2fd3593 1.016595 1.082999 -0.066404 1.065320
worse  5%   c95f1ced15f2 7d29d5e39734 0.453694 0.471156 -0.017462 1.038488
worse 10%   3e144ed1d5b7 2aef0e942480 0.035140 0.037535 -0.002395 1.068156
worse 25%   321fc60db035 801748ba582a 0.009267 0.009325 -0.000058 1.006259
median      2088ce763fc2 e6991321d78b 0.000665 0.000651 0.000014 0.978947
best 25%    915631a97de6 385b31354be6 0.040743 0.040363 0.000380 0.990673
best 10%    ad495c36a765 19c10384d3e7 0.431658 0.411490 0.020168 0.953278
best  5%    d13ae7d283ae 813c99f810ac 1.141404 1.075346 0.066058 0.942126
best  1%    81593cb4a496 99ae11866969 1.833297 0.063823 1.769474 0.034813
best cases  c3b14617fbd7 743a0fcaa4eb 1101.811740 2.735970 1099.075770 0.002483
            c3b14617fbd7 9ba6ab77fd29 1116.753953 2.800729 1113.953224 0.002508
            058b99d6e81f 57e249b7a3ea 1246.128485 3.042762 1243.085723 0.002442
            9a8c361aab49 0354a250d371 1253.111894 3.085796 1250.026098 0.002463
            442dbbc53c68 3ec1002a818c 1261.786294 3.138607 1258.647687 0.002487


As one can see, the average case is not really impacted. However, the worth case
we get after this changeset are much better than the one we had before it. We
have 30 pairs where improvements are above 10 minutes.

This reflect in the combined time for all pairs
before: 26256s
after:   1300s (-95%)

If we remove these pathological 30 cases, we still see a significant improvements:
before:  1631s
after:   1245s (-24%)
Yuya Nishihara - Oct. 11, 2019, 12:02 p.m.
On Fri, 11 Oct 2019 01:04:12 +0200, Pierre-Yves David wrote:
> # HG changeset patch
> # User Pierre-Yves David <pierre-yves.david@octobus.net>
> # Date 1570672173 -7200
> #      Thu Oct 10 03:49:33 2019 +0200
> # Node ID 2477ba483c04067900d1e9f6523b03df68a4d545
> # Parent  8ff1ecfadcd110849c47c77e31c92809eea466ab
> # EXP-Topic patchcopies-regression
> # Available At https://bitbucket.org/octobus/mercurial-devel/
> #              hg pull https://bitbucket.org/octobus/mercurial-devel/ -r 2477ba483c04
> patchcopies: give up any optimization based on `introrev`
> 
> Between 8a0136f69027 and d98fb3f42f33, we sped up the search for the
> introduction revision during path copies. However, further checking show that
> finding the introduction revision is still expensive and that we are better off
> without it. So we simply drop it and only rely on the linkrev optimisation.
> 
> I ran `perfpathcopies` on 6989 pair of revision in the pypy
> repository (`hg perfhelper-pathcopies`. The result is massively in favor of
> dropping this condition. The result of the copy tracing are unchanged.
> 
> Attempt to use a smaller changes preserving linkrev usage were unsuccessful, it
> can return wrong result. The following changesets broke test-mv-cp-st-diff.t
> 
>     -        if not f.isintroducedafter(limit):
>     +        if limit >= 0 and f.linkrev() < limit:
>                  return None

Sure. I meant comparing linkrevs, not linkrev vs (changelog) rev. If both
arms were linkrevs, and if the limit pointed to the revision where the
linkrevs of its files are guaranteed to be the lowest, the comparison
(e.g. f.linkrev() < repo[limit][path].linkrev()) would work.

My question was whether it would make things fast or not.
Pierre-Yves David - Oct. 11, 2019, 6 p.m.
On 10/11/19 2:02 PM, Yuya Nishihara wrote:
> On Fri, 11 Oct 2019 01:04:12 +0200, Pierre-Yves David wrote:
>> # HG changeset patch
>> # User Pierre-Yves David <pierre-yves.david@octobus.net>
>> # Date 1570672173 -7200
>> #      Thu Oct 10 03:49:33 2019 +0200
>> # Node ID 2477ba483c04067900d1e9f6523b03df68a4d545
>> # Parent  8ff1ecfadcd110849c47c77e31c92809eea466ab
>> # EXP-Topic patchcopies-regression
>> # Available At https://bitbucket.org/octobus/mercurial-devel/
>> #              hg pull https://bitbucket.org/octobus/mercurial-devel/ -r 2477ba483c04
>> patchcopies: give up any optimization based on `introrev`
>>
>> Between 8a0136f69027 and d98fb3f42f33, we sped up the search for the
>> introduction revision during path copies. However, further checking show that
>> finding the introduction revision is still expensive and that we are better off
>> without it. So we simply drop it and only rely on the linkrev optimisation.
>>
>> I ran `perfpathcopies` on 6989 pair of revision in the pypy
>> repository (`hg perfhelper-pathcopies`. The result is massively in favor of
>> dropping this condition. The result of the copy tracing are unchanged.
>>
>> Attempt to use a smaller changes preserving linkrev usage were unsuccessful, it
>> can return wrong result. The following changesets broke test-mv-cp-st-diff.t
>>
>>      -        if not f.isintroducedafter(limit):
>>      +        if limit >= 0 and f.linkrev() < limit:
>>                   return None
> 
> Sure. I meant comparing linkrevs, not linkrev vs (changelog) rev. If both
> arms were linkrevs, and if the limit pointed to the revision where the
> linkrevs of its files are guaranteed to be the lowest, the comparison
> (e.g. f.linkrev() < repo[limit][path].linkrev()) would work.
> 
> My question was whether it would make things fast or not.

Ha, I understand your proposal better now.

My testing of the version in this patch versus the "wrong" version using 
linkrev too agressively did not show a significant difference (-1%). So 
I don't it to raise anything better. Can we move forward with this 
series and maybe try a different approach later ?

Cheers,
Yuya Nishihara - Oct. 13, 2019, 10:08 a.m.
On Fri, 11 Oct 2019 20:00:58 +0200, Pierre-Yves David wrote:
> On 10/11/19 2:02 PM, Yuya Nishihara wrote:
> > On Fri, 11 Oct 2019 01:04:12 +0200, Pierre-Yves David wrote:
> >> # HG changeset patch
> >> # User Pierre-Yves David <pierre-yves.david@octobus.net>
> >> # Date 1570672173 -7200
> >> #      Thu Oct 10 03:49:33 2019 +0200
> >> # Node ID 2477ba483c04067900d1e9f6523b03df68a4d545
> >> # Parent  8ff1ecfadcd110849c47c77e31c92809eea466ab
> >> # EXP-Topic patchcopies-regression
> >> # Available At https://bitbucket.org/octobus/mercurial-devel/
> >> #              hg pull https://bitbucket.org/octobus/mercurial-devel/ -r 2477ba483c04
> >> patchcopies: give up any optimization based on `introrev`
> >>
> >> Between 8a0136f69027 and d98fb3f42f33, we sped up the search for the
> >> introduction revision during path copies. However, further checking show that
> >> finding the introduction revision is still expensive and that we are better off
> >> without it. So we simply drop it and only rely on the linkrev optimisation.
> >>
> >> I ran `perfpathcopies` on 6989 pair of revision in the pypy
> >> repository (`hg perfhelper-pathcopies`. The result is massively in favor of
> >> dropping this condition. The result of the copy tracing are unchanged.
> >>
> >> Attempt to use a smaller changes preserving linkrev usage were unsuccessful, it
> >> can return wrong result. The following changesets broke test-mv-cp-st-diff.t
> >>
> >>      -        if not f.isintroducedafter(limit):
> >>      +        if limit >= 0 and f.linkrev() < limit:
> >>                   return None
> > 
> > Sure. I meant comparing linkrevs, not linkrev vs (changelog) rev. If both
> > arms were linkrevs, and if the limit pointed to the revision where the
> > linkrevs of its files are guaranteed to be the lowest, the comparison
> > (e.g. f.linkrev() < repo[limit][path].linkrev()) would work.
> > 
> > My question was whether it would make things fast or not.
> 
> Ha, I understand your proposal better now.
> 
> My testing of the version in this patch versus the "wrong" version using 
> linkrev too agressively did not show a significant difference (-1%).

Do you mean only 1% speedup or measurable speedup on 1% cases? I don't think
the former is significant, but the latter could be depending on use cases.

> So
> I don't it to raise anything better. Can we move forward with this
> series and maybe try a different approach later ?

Better to leave the _findlimit() function? Or won't it be any useful in
your future work?
Pierre-Yves David - Oct. 13, 2019, 9:02 p.m.
On 10/13/19 12:08 PM, Yuya Nishihara wrote:
> On Fri, 11 Oct 2019 20:00:58 +0200, Pierre-Yves David wrote:
>> On 10/11/19 2:02 PM, Yuya Nishihara wrote:
>>> On Fri, 11 Oct 2019 01:04:12 +0200, Pierre-Yves David wrote:
>>>> # HG changeset patch
>>>> # User Pierre-Yves David <pierre-yves.david@octobus.net>
>>>> # Date 1570672173 -7200
>>>> #      Thu Oct 10 03:49:33 2019 +0200
>>>> # Node ID 2477ba483c04067900d1e9f6523b03df68a4d545
>>>> # Parent  8ff1ecfadcd110849c47c77e31c92809eea466ab
>>>> # EXP-Topic patchcopies-regression
>>>> # Available At https://bitbucket.org/octobus/mercurial-devel/
>>>> #              hg pull https://bitbucket.org/octobus/mercurial-devel/ -r 2477ba483c04
>>>> patchcopies: give up any optimization based on `introrev`
>>>>
>>>> Between 8a0136f69027 and d98fb3f42f33, we sped up the search for the
>>>> introduction revision during path copies. However, further checking show that
>>>> finding the introduction revision is still expensive and that we are better off
>>>> without it. So we simply drop it and only rely on the linkrev optimisation.
>>>>
>>>> I ran `perfpathcopies` on 6989 pair of revision in the pypy
>>>> repository (`hg perfhelper-pathcopies`. The result is massively in favor of
>>>> dropping this condition. The result of the copy tracing are unchanged.
>>>>
>>>> Attempt to use a smaller changes preserving linkrev usage were unsuccessful, it
>>>> can return wrong result. The following changesets broke test-mv-cp-st-diff.t
>>>>
>>>>       -        if not f.isintroducedafter(limit):
>>>>       +        if limit >= 0 and f.linkrev() < limit:
>>>>                    return None
>>>
>>> Sure. I meant comparing linkrevs, not linkrev vs (changelog) rev. If both
>>> arms were linkrevs, and if the limit pointed to the revision where the
>>> linkrevs of its files are guaranteed to be the lowest, the comparison
>>> (e.g. f.linkrev() < repo[limit][path].linkrev()) would work.
>>>
>>> My question was whether it would make things fast or not.
>>
>> Ha, I understand your proposal better now.
>>
>> My testing of the version in this patch versus the "wrong" version using
>> linkrev too agressively did not show a significant difference (-1%).
> 
> Do you mean only 1% speedup or measurable speedup on 1% cases? I don't think
> the former is significant, but the latter could be depending on use cases.

only ±1% speedup for all (many) cases I looked at.

>> So
>> I don't it to raise anything better. Can we move forward with this
>> series and maybe try a different approach later ?
> 
> Better to leave the _findlimit() function? Or won't it be any useful in
> your future work?

That seems easy enough t reintroduce later if needed. My current 
optimization work focus on the changeset based copies tracing algorithm. 
I don't think I'll spend much time on the filelog based one.
Yuya Nishihara - Oct. 14, 2019, 3:57 a.m.
On Sun, 13 Oct 2019 23:02:25 +0200, Pierre-Yves David wrote:
> On 10/13/19 12:08 PM, Yuya Nishihara wrote:
> > On Fri, 11 Oct 2019 20:00:58 +0200, Pierre-Yves David wrote:
> >> On 10/11/19 2:02 PM, Yuya Nishihara wrote:
> >>> On Fri, 11 Oct 2019 01:04:12 +0200, Pierre-Yves David wrote:
> >>>> # HG changeset patch
> >>>> # User Pierre-Yves David <pierre-yves.david@octobus.net>
> >>>> # Date 1570672173 -7200
> >>>> #      Thu Oct 10 03:49:33 2019 +0200
> >>>> # Node ID 2477ba483c04067900d1e9f6523b03df68a4d545
> >>>> # Parent  8ff1ecfadcd110849c47c77e31c92809eea466ab
> >>>> # EXP-Topic patchcopies-regression
> >>>> # Available At https://bitbucket.org/octobus/mercurial-devel/
> >>>> #              hg pull https://bitbucket.org/octobus/mercurial-devel/ -r 2477ba483c04
> >>>> patchcopies: give up any optimization based on `introrev`
> >>>>
> >>>> Between 8a0136f69027 and d98fb3f42f33, we sped up the search for the
> >>>> introduction revision during path copies. However, further checking show that
> >>>> finding the introduction revision is still expensive and that we are better off
> >>>> without it. So we simply drop it and only rely on the linkrev optimisation.
> >>>>
> >>>> I ran `perfpathcopies` on 6989 pair of revision in the pypy
> >>>> repository (`hg perfhelper-pathcopies`. The result is massively in favor of
> >>>> dropping this condition. The result of the copy tracing are unchanged.
> >>>>
> >>>> Attempt to use a smaller changes preserving linkrev usage were unsuccessful, it
> >>>> can return wrong result. The following changesets broke test-mv-cp-st-diff.t
> >>>>
> >>>>       -        if not f.isintroducedafter(limit):
> >>>>       +        if limit >= 0 and f.linkrev() < limit:
> >>>>                    return None
> >>>
> >>> Sure. I meant comparing linkrevs, not linkrev vs (changelog) rev. If both
> >>> arms were linkrevs, and if the limit pointed to the revision where the
> >>> linkrevs of its files are guaranteed to be the lowest, the comparison
> >>> (e.g. f.linkrev() < repo[limit][path].linkrev()) would work.
> >>>
> >>> My question was whether it would make things fast or not.
> >>
> >> Ha, I understand your proposal better now.
> >>
> >> My testing of the version in this patch versus the "wrong" version using
> >> linkrev too agressively did not show a significant difference (-1%).
> > 
> > Do you mean only 1% speedup or measurable speedup on 1% cases? I don't think
> > the former is significant, but the latter could be depending on use cases.
> 
> only ±1% speedup for all (many) cases I looked at.
> 
> >> So
> >> I don't it to raise anything better. Can we move forward with this
> >> series and maybe try a different approach later ?
> > 
> > Better to leave the _findlimit() function? Or won't it be any useful in
> > your future work?
> 
> That seems easy enough t reintroduce later if needed. My current 
> optimization work focus on the changeset based copies tracing algorithm. 
> I don't think I'll spend much time on the filelog based one.

Fair enough. Queued the series, thanks.
via Mercurial-devel - Oct. 14, 2019, 6:34 a.m.
I've updated the queued version:

1. ran contrib/grey.py on it (I also did that on one of your
recent perf patches)
2. s/patchcopies/pathcopies/ in the commit message

On Sun, Oct 13, 2019 at 9:10 PM Yuya Nishihara <yuya@tcha.org> wrote:

> On Sun, 13 Oct 2019 23:02:25 +0200, Pierre-Yves David wrote:
> > On 10/13/19 12:08 PM, Yuya Nishihara wrote:
> > > On Fri, 11 Oct 2019 20:00:58 +0200, Pierre-Yves David wrote:
> > >> On 10/11/19 2:02 PM, Yuya Nishihara wrote:
> > >>> On Fri, 11 Oct 2019 01:04:12 +0200, Pierre-Yves David wrote:
> > >>>> # HG changeset patch
> > >>>> # User Pierre-Yves David <pierre-yves.david@octobus.net>
> > >>>> # Date 1570672173 -7200
> > >>>> #      Thu Oct 10 03:49:33 2019 +0200
> > >>>> # Node ID 2477ba483c04067900d1e9f6523b03df68a4d545
> > >>>> # Parent  8ff1ecfadcd110849c47c77e31c92809eea466ab
> > >>>> # EXP-Topic patchcopies-regression
> > >>>> # Available At https://bitbucket.org/octobus/mercurial-devel/
> > >>>> #              hg pull
> https://bitbucket.org/octobus/mercurial-devel/ -r 2477ba483c04
> > >>>> patchcopies: give up any optimization based on `introrev`
> > >>>>
> > >>>> Between 8a0136f69027 and d98fb3f42f33, we sped up the search for the
> > >>>> introduction revision during path copies. However, further checking
> show that
> > >>>> finding the introduction revision is still expensive and that we
> are better off
> > >>>> without it. So we simply drop it and only rely on the linkrev
> optimisation.
> > >>>>
> > >>>> I ran `perfpathcopies` on 6989 pair of revision in the pypy
> > >>>> repository (`hg perfhelper-pathcopies`. The result is massively in
> favor of
> > >>>> dropping this condition. The result of the copy tracing are
> unchanged.
> > >>>>
> > >>>> Attempt to use a smaller changes preserving linkrev usage were
> unsuccessful, it
> > >>>> can return wrong result. The following changesets broke
> test-mv-cp-st-diff.t
> > >>>>
> > >>>>       -        if not f.isintroducedafter(limit):
> > >>>>       +        if limit >= 0 and f.linkrev() < limit:
> > >>>>                    return None
> > >>>
> > >>> Sure. I meant comparing linkrevs, not linkrev vs (changelog) rev. If
> both
> > >>> arms were linkrevs, and if the limit pointed to the revision where
> the
> > >>> linkrevs of its files are guaranteed to be the lowest, the comparison
> > >>> (e.g. f.linkrev() < repo[limit][path].linkrev()) would work.
> > >>>
> > >>> My question was whether it would make things fast or not.
> > >>
> > >> Ha, I understand your proposal better now.
> > >>
> > >> My testing of the version in this patch versus the "wrong" version
> using
> > >> linkrev too agressively did not show a significant difference (-1%).
> > >
> > > Do you mean only 1% speedup or measurable speedup on 1% cases? I don't
> think
> > > the former is significant, but the latter could be depending on use
> cases.
> >
> > only ±1% speedup for all (many) cases I looked at.
> >
> > >> So
> > >> I don't it to raise anything better. Can we move forward with this
> > >> series and maybe try a different approach later ?
> > >
> > > Better to leave the _findlimit() function? Or won't it be any useful in
> > > your future work?
> >
> > That seems easy enough t reintroduce later if needed. My current
> > optimization work focus on the changeset based copies tracing algorithm.
> > I don't think I'll spend much time on the filelog based one.
>
> Fair enough. Queued the series, thanks.
> _______________________________________________
> Mercurial-devel mailing list
> Mercurial-devel@mercurial-scm.org
> https://www.mercurial-scm.org/mailman/listinfo/mercurial-devel
>

Patch

diff --git a/mercurial/copies.py b/mercurial/copies.py
--- a/mercurial/copies.py
+++ b/mercurial/copies.py
@@ -155,7 +155,11 @@  def _chain(a, b):
 
 def _tracefile(fctx, am, basemf, limit):
     """return file context that is the ancestor of fctx present in ancestor
-    manifest am, stopping after the first ancestor lower than limit"""
+    manifest am
+
+    Note: we used to try and stop after a given limit, however checking if that
+    limit is reached turned out to be very expensive. we are better off
+    disabling that feature."""
 
     for f in fctx.ancestors():
         path = f.path()
@@ -163,9 +167,6 @@  def _tracefile(fctx, am, basemf, limit):
             return path
         if basemf and basemf.get(path, None) == f.filenode():
             return path
-        if not f.isintroducedafter(limit):
-            return None
-
 
 def _dirstatecopies(repo, match=None):
     ds = repo.dirstate