| Submitter | Pierre-Yves David |
|---|---|
| Date | Dec. 24, 2013, 12:32 a.m. |
| Message ID | <41a5810bb9cb062cd10f.1387845142@marginatus.fb.com> |
| Download | mbox | patch |
| Permalink | /patch/3235/ |
| State | Accepted |
| Commit | cd62532c62a1d780d7b431013a84786372481b75 |
| Headers | show |
Comments
On Mon, Dec 23, 2013 at 04:32:22PM -0800, pierre-yves.david@ens-lyon.org wrote: > # HG changeset patch > # User Pierre-Yves David <pierre-yves.david@ens-lyon.org> > # Date 1387841391 28800 > # Mon Dec 23 15:29:51 2013 -0800 > # Node ID 41a5810bb9cb062cd10f6a12b261fdd66fab8aed > # Parent 1677c9118fd971197916b7f37177b2a5c03ad802 > obsolete: order of magnitude speedup in _computebumpedset queueing these, thanks. I may make some English tweaks depending on my motivation level. > > Reminder: a changeset is said "bumped" if it tries to obsolete a immutable > changeset. > > > The previous algorithm for computing bumped changeset was: > > 1) Get all public changesets > 2) Find all they successors > 3) Search for stuff that are eligible for being "bumped" > (mutable and non obsolete) > > The entry size of this algorithm is `O(len(public))` which is mostly the same as > `O(len(repo))`. Even this this approach mean fewer obsolescence marker are > traveled, this is not very scalable. > > The new algorithm is: > > 1) For each potential bumped changesets (non obsolete mutable) > 2) iterate over precursors > 3) if a precursors is public. changeset is bumped > > We travel more obsolescence marker, but the entry size is much smaller since > the amount of potential bumped should remains mostly stable with time `O(1)`. > > On some confidential gigantic repo this move bumped computation from 15.19s to > 0.46s (×33 speedup…). On "smaller" repo (mercurial, cubicweb's review) no > significant gain were seen. The additional traversal of obsolescence marker is > probably probably counter balance the advantage of it. > > Other optimisation could be done in the future (eg: sharing precursors cache > for divergence detection) > > diff --git a/mercurial/obsolete.py b/mercurial/obsolete.py > --- a/mercurial/obsolete.py > +++ b/mercurial/obsolete.py > @@ -82,10 +82,11 @@ The header is followed by the markers. E > additional encoding. Keys cannot contain '\0' or ':' and values > cannot contain '\0'. > """ > import struct > import util, base85, node > +import phases > from i18n import _ > > _pack = struct.pack > _unpack = struct.unpack > > @@ -784,18 +785,30 @@ def _computeextinctset(repo): > > > @cachefor('bumped') > def _computebumpedset(repo): > """the set of revs trying to obsolete public revisions""" > - # get all possible bumped changesets > - tonode = repo.changelog.node > - publicnodes = (tonode(r) for r in repo.revs('public()')) > - successors = allsuccessors(repo.obsstore, publicnodes, > - ignoreflags=bumpedfix) > - # revision public or already obsolete don't count as bumped > - query = '%ld - obsolete() - public()' > - return set(repo.revs(query, _knownrevs(repo, successors))) > + bumped = set() > + # utils function (avoid attribut lookup in the loop) > + phase = repo._phasecache.phase # would be faster to grab the full list > + public = phases.public > + cl = repo.changelog > + torev = cl.nodemap.get > + obs = getrevs(repo, 'obsolete') > + for rev in repo: > + # We only evaluate mutable, non-obsolete revision > + if (public < phase(repo, rev)) and (rev not in obs): > + node = cl.node(rev) > + # (future) A cache of precursors may worth if split is very common > + for pnode in allprecursors(repo.obsstore, [node], > + ignoreflags=bumpedfix): > + prev = torev(pnode) # unfiltered! but so is phasecache > + if (prev is not None) and (phase(repo, prev) <= public): > + # we have a public precursors > + bumped.add(rev) > + break # Next draft! > + return bumped > > @cachefor('divergent') > def _computedivergentset(repo): > """the set of rev that compete to be the final successors of some revision. > """ > _______________________________________________ > Mercurial-devel mailing list > Mercurial-devel@selenic.com > http://selenic.com/mailman/listinfo/mercurial-devel
Patch
diff --git a/mercurial/obsolete.py b/mercurial/obsolete.py --- a/mercurial/obsolete.py +++ b/mercurial/obsolete.py @@ -82,10 +82,11 @@ The header is followed by the markers. E additional encoding. Keys cannot contain '\0' or ':' and values cannot contain '\0'. """ import struct import util, base85, node +import phases from i18n import _ _pack = struct.pack _unpack = struct.unpack @@ -784,18 +785,30 @@ def _computeextinctset(repo): @cachefor('bumped') def _computebumpedset(repo): """the set of revs trying to obsolete public revisions""" - # get all possible bumped changesets - tonode = repo.changelog.node - publicnodes = (tonode(r) for r in repo.revs('public()')) - successors = allsuccessors(repo.obsstore, publicnodes, - ignoreflags=bumpedfix) - # revision public or already obsolete don't count as bumped - query = '%ld - obsolete() - public()' - return set(repo.revs(query, _knownrevs(repo, successors))) + bumped = set() + # utils function (avoid attribut lookup in the loop) + phase = repo._phasecache.phase # would be faster to grab the full list + public = phases.public + cl = repo.changelog + torev = cl.nodemap.get + obs = getrevs(repo, 'obsolete') + for rev in repo: + # We only evaluate mutable, non-obsolete revision + if (public < phase(repo, rev)) and (rev not in obs): + node = cl.node(rev) + # (future) A cache of precursors may worth if split is very common + for pnode in allprecursors(repo.obsstore, [node], + ignoreflags=bumpedfix): + prev = torev(pnode) # unfiltered! but so is phasecache + if (prev is not None) and (phase(repo, prev) <= public): + # we have a public precursors + bumped.add(rev) + break # Next draft! + return bumped @cachefor('divergent') def _computedivergentset(repo): """the set of rev that compete to be the final successors of some revision. """
# HG changeset patch # User Pierre-Yves David <pierre-yves.david@ens-lyon.org> # Date 1387841391 28800 # Mon Dec 23 15:29:51 2013 -0800 # Node ID 41a5810bb9cb062cd10f6a12b261fdd66fab8aed # Parent 1677c9118fd971197916b7f37177b2a5c03ad802 obsolete: order of magnitude speedup in _computebumpedset Reminder: a changeset is said "bumped" if it tries to obsolete a immutable changeset. The previous algorithm for computing bumped changeset was: 1) Get all public changesets 2) Find all they successors 3) Search for stuff that are eligible for being "bumped" (mutable and non obsolete) The entry size of this algorithm is `O(len(public))` which is mostly the same as `O(len(repo))`. Even this this approach mean fewer obsolescence marker are traveled, this is not very scalable. The new algorithm is: 1) For each potential bumped changesets (non obsolete mutable) 2) iterate over precursors 3) if a precursors is public. changeset is bumped We travel more obsolescence marker, but the entry size is much smaller since the amount of potential bumped should remains mostly stable with time `O(1)`. On some confidential gigantic repo this move bumped computation from 15.19s to 0.46s (×33 speedup…). On "smaller" repo (mercurial, cubicweb's review) no significant gain were seen. The additional traversal of obsolescence marker is probably probably counter balance the advantage of it. Other optimisation could be done in the future (eg: sharing precursors cache for divergence detection)