discovery: slowly increase sampling size
Some pathological discovery runs can requires many roundtrip. When this happens
things can get very slow.
To make the algorithm more resilience again such pathological case. We slowly
increase the sample size with each roundtrip (+5%). This will have a negligible
impact on "normal" discovery with few roundtrips, but a large positive impact of
case with many roundtrips. Asking more question per roundtrip helps to reduce
the undecided set faster. Instead of reducing the undecided set a linear speed
(in the worst case), we reduce it as a guaranteed (small) exponential rate. The
data below show this slow ramp up in sample size:
round trip | 1 | 5 | 10 | 20 | 50 | 100 | 130 |
sample size | 200 | 254 | 321 | 517 | 2 199 | 25 123 | 108 549 |
covered nodes | 200 | 1 357 | 2 821 | 7 031 | 42 658 | 524 530 | 2 276 755 |
To be a bit more concrete, lets take a very pathological case as an example. We
are doing discovery from a copy of Mozilla-try to a more recent version of
mozilla-unified. Mozilla-unified heads are unknown to the mozilla-try repo and
there are over 1 million "missing" changesets. (the discovery is "local" to
avoid network interference)
Without this change, the discovery:
- last 1858 seconds (31 minutes),
- does 1700 round trip,
- asking about 340 000 nodes.
With this change, the discovery:
- last 218 seconds (3 minutes, 38 seconds a -88% improvement),
- does 94 round trip (-94%),
- asking about 344 211 nodes (+1%).
Of course, this is an extreme case (and 3 minutes is still slow). However this
give a good example of how this sample size increase act as a safety net
catching any bad situations.
We could image a steeper increase than 5%. For example 10% would give the
following number:
round trip | 1 | 5 | 10 | 20 | 50 | 75 | 100 |
sample size | 200 | 321 | 514 | 1 326 | 23 060 | 249 812 | 2 706 594 |
covered nodes | 200 | 1 541 | 3 690 | 12 671 | 251 871 | 2 746 254 | 29 770 966 |
In parallel, it is useful to understand these pathological cases and improve
them. However the current change provides a general purpose safety net to smooth
the impact of pathological cases.
To avoid issue with older http server, the increase in sample size only occurs
if the protocol has not limit on command argument size.
# Mercurial extension to provide 'hg relink' command
#
# Copyright (C) 2007 Brendan Cully <brendan@kublai.com>
#
# This software may be used and distributed according to the terms of the
# GNU General Public License version 2 or any later version.
"""recreates hardlinks between repository clones"""
from __future__ import absolute_import
import os
import stat
from mercurial.i18n import _
from mercurial import (
error,
hg,
registrar,
util,
)
from mercurial.utils import (
stringutil,
)
cmdtable = {}
command = registrar.command(cmdtable)
# Note for extension authors: ONLY specify testedwith = 'ships-with-hg-core' for
# extensions which SHIP WITH MERCURIAL. Non-mainline extensions should
# be specifying the version(s) of Mercurial they are tested with, or
# leave the attribute unspecified.
testedwith = 'ships-with-hg-core'
@command('relink', [], _('[ORIGIN]'), helpcategory=command.CATEGORY_MAINTENANCE)
def relink(ui, repo, origin=None, **opts):
"""recreate hardlinks between two repositories
When repositories are cloned locally, their data files will be
hardlinked so that they only use the space of a single repository.
Unfortunately, subsequent pulls into either repository will break
hardlinks for any files touched by the new changesets, even if
both repositories end up pulling the same changes.
Similarly, passing --rev to "hg clone" will fail to use any
hardlinks, falling back to a complete copy of the source
repository.
This command lets you recreate those hardlinks and reclaim that
wasted space.
This repository will be relinked to share space with ORIGIN, which
must be on the same local disk. If ORIGIN is omitted, looks for
"default-relink", then "default", in [paths].
Do not attempt any read operations on this repository while the
command is running. (Both repositories will be locked against
writes.)
"""
if (not util.safehasattr(util, 'samefile') or
not util.safehasattr(util, 'samedevice')):
raise error.Abort(_('hardlinks are not supported on this system'))
src = hg.repository(repo.baseui, ui.expandpath(origin or 'default-relink',
origin or 'default'))
ui.status(_('relinking %s to %s\n') % (src.store.path, repo.store.path))
if repo.root == src.root:
ui.status(_('there is nothing to relink\n'))
return
if not util.samedevice(src.store.path, repo.store.path):
# No point in continuing
raise error.Abort(_('source and destination are on different devices'))
with repo.lock(), src.lock():
candidates = sorted(collect(src, ui))
targets = prune(candidates, src.store.path, repo.store.path, ui)
do_relink(src.store.path, repo.store.path, targets, ui)
def collect(src, ui):
seplen = len(os.path.sep)
candidates = []
live = len(src['tip'].manifest())
# Your average repository has some files which were deleted before
# the tip revision. We account for that by assuming that there are
# 3 tracked files for every 2 live files as of the tip version of
# the repository.
#
# mozilla-central as of 2010-06-10 had a ratio of just over 7:5.
total = live * 3 // 2
src = src.store.path
progress = ui.makeprogress(_('collecting'), unit=_('files'), total=total)
pos = 0
ui.status(_("tip has %d files, estimated total number of files: %d\n")
% (live, total))
for dirpath, dirnames, filenames in os.walk(src):
dirnames.sort()
relpath = dirpath[len(src) + seplen:]
for filename in sorted(filenames):
if filename[-2:] not in ('.d', '.i'):
continue
st = os.stat(os.path.join(dirpath, filename))
if not stat.S_ISREG(st.st_mode):
continue
pos += 1
candidates.append((os.path.join(relpath, filename), st))
progress.update(pos, item=filename)
progress.complete()
ui.status(_('collected %d candidate storage files\n') % len(candidates))
return candidates
def prune(candidates, src, dst, ui):
def linkfilter(src, dst, st):
try:
ts = os.stat(dst)
except OSError:
# Destination doesn't have this file?
return False
if util.samefile(src, dst):
return False
if not util.samedevice(src, dst):
# No point in continuing
raise error.Abort(
_('source and destination are on different devices'))
if st.st_size != ts.st_size:
return False
return st
targets = []
progress = ui.makeprogress(_('pruning'), unit=_('files'),
total=len(candidates))
pos = 0
for fn, st in candidates:
pos += 1
srcpath = os.path.join(src, fn)
tgt = os.path.join(dst, fn)
ts = linkfilter(srcpath, tgt, st)
if not ts:
ui.debug('not linkable: %s\n' % fn)
continue
targets.append((fn, ts.st_size))
progress.update(pos, item=fn)
progress.complete()
ui.status(_('pruned down to %d probably relinkable files\n') % len(targets))
return targets
def do_relink(src, dst, files, ui):
def relinkfile(src, dst):
bak = dst + '.bak'
os.rename(dst, bak)
try:
util.oslink(src, dst)
except OSError:
os.rename(bak, dst)
raise
os.remove(bak)
CHUNKLEN = 65536
relinked = 0
savedbytes = 0
progress = ui.makeprogress(_('relinking'), unit=_('files'),
total=len(files))
pos = 0
for f, sz in files:
pos += 1
source = os.path.join(src, f)
tgt = os.path.join(dst, f)
# Binary mode, so that read() works correctly, especially on Windows
sfp = open(source, 'rb')
dfp = open(tgt, 'rb')
sin = sfp.read(CHUNKLEN)
while sin:
din = dfp.read(CHUNKLEN)
if sin != din:
break
sin = sfp.read(CHUNKLEN)
sfp.close()
dfp.close()
if sin:
ui.debug('not linkable: %s\n' % f)
continue
try:
relinkfile(source, tgt)
progress.update(pos, item=f)
relinked += 1
savedbytes += sz
except OSError as inst:
ui.warn('%s: %s\n' % (tgt, stringutil.forcebytestr(inst)))
progress.complete()
ui.status(_('relinked %d files (%s reclaimed)\n') %
(relinked, util.bytecount(savedbytes)))