( )⚙ D4366 treemanifest: introduce lazy loading of subdirs

This is an archive of the discontinued Mercurial Phabricator instance.

treemanifest: introduce lazy loading of subdirs
ClosedPublic

Authored by spectral on Aug 24 2018, 1:36 AM.

Details

Summary

An earlier patch series made it so that what to load was up to the calling code,
which works fine until manifests are copied - when they're copied, they're
loaded completely and thus we lose the entire benefit.

By lazy loading everything, we can avoid having to pass in the matcher to ~every
manifest function, and handle copies correctly as well. This changeset doesn't
go as far as it could with loading only the necessary subsets, that will happen
in later changes in this series; at the moment, except in a few situations, we
just load everything the moment we want to interact with treemanifest._dirs.
This is thus most likely to be a small slowdown if treemanifests is in use
regardless of whether narrow is in use, but hopefully easier to verify
correctness and review.

This is part of a series of speedups, it is not expected to produce any real speed
improvements itself, but the numbers show that it doesn't produce a large speed
penalty in any common case, and for the cases it does provide a penalty in, it
is not a large absolute amount (even if it is a large percentage amount).

Timing numbers according to command:

hyperfine --prepare <preparation_script> 'hg status'

HGRCPATH points to a file with the following contents:

[extensions]
narrow =
strip =
rebase =

mozilla-unified (called m-u below) was at revision #468856.

      regular hash: eb39298e432d
treemanifests hash: 0553b7f29eaf

large-dir-repo (called l-d-r below) was generated with the following script:

#!/bin/bash
hg init large-dir-repo
mkdir -p large-dir-repo/third_party/rust/log
touch large-dir-repo/third_party/rust/log/foo.txt
for i in $(seq 1 30000); do
    d=$(mktemp -d large-dir-repo/third_party/XXXXXXXXX)
    touch $d/file.txt
done
hg -R large-dir-repo ci -Am 'rev0' --user test --date '0 0'
echo hi > large-dir-repo/third_party/rust/log/bar.txt
hg -R large-dir-repo ci -Am 'rev1' --user test --date '0 0'
echo hi > large-dir-repo/third_party/rust/log/baz.txt
hg -R large-dir-repo ci -Am 'rev2' --user test --date '0 0'

for the repos that use narrow, the narrowspec was this:

[include]
rootfilesin:accessible/jsat
rootfilesin:accessible/tests/mochitest/jsat
rootfilesin:mobile/android/chrome/content
rootfilesin:mobile/android/modules/geckoview
rootfilesin:third_party/rust/log
[exclude]

This narrowspec was chosen due to the size of the third_party/rust directory
(this directory was *not* modified in revision #468856 in mozilla-unified),
plus all the directories that *were* modified in revision #468856 of
mozilla-unified.

Importantly, when using narrow, these repos had everything checked out (in the
case of large-dir-repo, that means all 30,001 directories), *before* adding the
narrowspec. This is to simulate the behavior when using a virtual filesystem
that shows everything for the user even if they haven't added it to the
narrowspec yet. This is not a supported configuration, and hg update and `hg
rebase` will not really do the "correct" thing if there are mutations outside
of the narrowspec (which is not the case in these tests, due to a carefully
crafted narrowspec), but non-mutating commands should behave correctly.

I'm not claiming anything less than a 5% speed win as improvements due to this
change; these are probably eiter measurement artifacts or constant time
improvements. The numbers that aren't changing are shown primarily to prove that
this doesn't make anything worse in any case I plan on testing during this
series.

'before' is hg from commit 6268fed3
'N' indicates narrow in use
'T' indicates treemanifest in use

Please note that these commands and the narrowspec are a little different than
the ones in a similar table that I made in a3cabe9415e1.

Important: it is my understanding that these numbers below are *not super reliable*,
the large slowdowns may be artifacts of some odd interaction between GC and
python module/code complexity. Another changeset of mine (D4351) had shown large
timing differences when ~empty, uncalled functions were added to match.py,
though only when using --color=never or redirecting to /dev/null. We seem to be
on some cusp of complexity or code size that is causing, at my best guess
(according to linux perf benchmarks) GC to alter behavior and cause a
200-400ms difference in timings. I haven't had a chance to replicate these
results on another machine.

diff --git:
repo  | N | T | before (mean +- stdev) | after (mean +- stdev) | % of before
------+---+---+------------------------+-----------------------+------------
m-u   |   |   | 1.580 s +-  0.034 s    | 1.576 s +-  0.022 s   |  99.7%
m-u   |   | x | 1.568 s +-  0.025 s    | 1.584 s +-  0.044 s   | 101.0%
m-u   | x |   | 1.569 s +-  0.031 s    | 1.554 s +-  0.025 s   |  99.0%
m-u   | x | x | 107.3 ms +-   1.6 ms   | 106.3 ms +-   1.5 ms  |  99.1%
l-d-r |   |   | 232.5 ms +-   5.9 ms   | 233.5 ms +-   5.3 ms  | 100.4%
l-d-r |   | x | 236.6 ms +-   6.3 ms   | 233.6 ms +-   7.0 ms  |  98.7%
l-d-r | x |   | 118.4 ms +-   2.1 ms   | 118.4 ms +-   1.4 ms  | 100.0%
l-d-r | x | x | 116.8 ms +-   1.5 ms   | 118.9 ms +-   1.6 ms  | 101.8%

diff -c . --git:
repo  | N | T | before (mean +- stdev) | after (mean +- stdev) | % of before
------+---+---+------------------------+-----------------------+------------
m-u   |   |   | 354.4 ms +-  16.6 ms   | 351.0 ms +-   6.9 ms  |  99.0%
m-u   |   | x | 207.2 ms +-   3.0 ms   | 206.2 ms +-   2.7 ms  |  99.5%
m-u   | x |   | 422.0 ms +-  26.0 ms   | 351.2 ms +-   6.4 ms  |  83.2% <--
m-u   | x | x | 166.7 ms +-   2.1 ms   | 169.5 ms +-   4.1 ms  | 101.7%
l-d-r |   |   | 98.4 ms +-   4.5 ms    | 98.5 ms +-   2.1 ms   | 100.1%
l-d-r |   | x | 5.519 s +-  0.060 s    | 5.149 s +-  0.042 s   |  93.3% <--
l-d-r | x |   | 99.1 ms +-   3.2 ms    | 102.6 ms +-   9.7 ms  | 103.5% <--?
l-d-r | x | x | 994.9 ms +-  10.7 ms   | 1.026 s +-  0.012 s   | 103.1% <--?

rebase -r . --keep -d .^^:
repo  | N | T | before (mean +- stdev) | after (mean +- stdev) | % of before
------+---+---+------------------------+-----------------------+------------
m-u   |   |   | 6.639 s +-  0.168 s    | 6.559 s +-  0.097 s   |  98.8%
m-u   |   | x | 6.601 s +-  0.143 s    | 6.640 s +-  0.207 s   | 100.6%
m-u   | x |   | 6.582 s +-  0.098 s    | 6.543 s +-  0.098 s   |  99.4%
m-u   | x | x | 678.4 ms +-  57.7 ms   | 703.7 ms +-  52.4 ms  | 103.7% <--?
l-d-r |   |   | 780.0 ms +-  23.9 ms   | 776.0 ms +-  12.6 ms  |  99.5%
l-d-r |   | x | 7.520 s +-  0.255 s    | 7.395 s +-  0.044 s   |  98.3%
l-d-r | x |   | 331.9 ms +-  16.5 ms   | 327.0 ms +-   3.4 ms  |  98.5%
l-d-r | x | x | 6.228 s +-  0.113 s    | 5.924 s +-  0.044 s   |  95.1%

status --change . --copies:
repo  | N | T | before (mean +- stdev) | after (mean +- stdev) | % of before
------+---+---+------------------------+-----------------------+------------
m-u   |   |   | 330.8 ms +-   7.2 ms   | 329.0 ms +-   7.1 ms  |  99.5%
m-u   |   | x | 182.9 ms +-   2.7 ms   | 183.5 ms +-   2.7 ms  | 100.3%
m-u   | x |   | 330.0 ms +-   7.6 ms   | 327.1 ms +-   5.4 ms  |  99.1%
m-u   | x | x | 146.2 ms +-   2.4 ms   | 147.1 ms +-   1.3 ms  | 100.6%
l-d-r |   |   | 95.3 ms +-   1.4 ms    | 95.9 ms +-   1.5 ms   | 100.6%
l-d-r |   | x | 5.157 s +-  0.035 s    | 5.166 s +-  0.058 s   | 100.2%
l-d-r | x |   | 99.7 ms +-   3.0 ms    | 100.2 ms +-   4.4 ms  | 100.5%
l-d-r | x | x | 993.6 ms +-  13.1 ms   | 1.025 s +-  0.015 s   | 103.2% <--?

status --copies:
repo  | N | T | before (mean +- stdev) | after (mean +- stdev) | % of before
------+---+---+------------------------+-----------------------+------------
m-u   |   |   | 2.348 s +-  0.031 s    | 2.329 s +-  0.019 s   |  99.2%
m-u   |   | x | 2.337 s +-  0.026 s    | 2.346 s +-  0.034 s   | 100.4%
m-u   | x |   | 2.354 s +-  0.015 s    | 2.342 s +-  0.021 s   |  99.5%
m-u   | x | x | 120.6 ms +-   4.3 ms   | 119.2 ms +-   2.1 ms  |  98.8%
l-d-r |   |   | 731.5 ms +-  11.1 ms   | 719.6 ms +-   9.8 ms  |  98.4%
l-d-r |   | x | 729.0 ms +-  15.5 ms   | 725.7 ms +-  10.6 ms  |  99.5%
l-d-r | x |   | 211.0 ms +-   3.9 ms   | 212.8 ms +-   3.7 ms  | 100.9%
l-d-r | x | x | 211.5 ms +-   4.2 ms   | 211.0 ms +-   3.3 ms  |  99.8%

update $rev^; ~/src/hg/hg{hg}/hg update $rev:
repo  | N | T | before (mean +- stdev) | after (mean +- stdev) | % of before
------+---+---+------------------------+-----------------------+------------
m-u   |   |   | 3.910 s +-  0.055 s    | 3.920 s +-  0.075 s   | 100.3%
m-u   |   | x | 3.613 s +-  0.056 s    | 3.630 s +-  0.056 s   | 100.5%
m-u   | x |   | 3.873 s +-  0.055 s    | 3.864 s +-  0.049 s   |  99.8%
m-u   | x | x | 400.4 ms +-   7.4 ms   | 403.6 ms +-   5.0 ms  | 100.8%
l-d-r |   |   | 531.6 ms +-  10.0 ms   | 528.8 ms +-   9.6 ms  |  99.5%
l-d-r |   | x | 10.377 s +-  0.049 s   | 9.955 s +-  0.046 s   |  95.9%
l-d-r | x |   | 308.3 ms +-   4.4 ms   | 306.8 ms +-   3.7 ms  |  99.5%
l-d-r | x | x | 1.805 s +-  0.015 s    | 1.834 s +-  0.020 s   | 101.6%

Diff Detail

Repository
rHG Mercurial
Lint
Lint Skipped
Unit
Unit Tests Skipped

Event Timeline

spectral created this revision.Aug 24 2018, 1:36 AM

I'm not sure how I feel about so many methods having the if dir in self._lazydirs: self._loadlazy(dir) pattern.

On one hand, action at a distance when it involves caching can be dangerous. And doing the lookup inline will avoid a Python function call.

On the other, it seems very redundant.

Overall I'm OK with the patch. I just feel like using a collections.defaultdict or implementing __missing__ to automagically resolve missing keys might work out better. Do you think this is a reasonable request?

spectral planned changes to this revision.Aug 24 2018, 6:36 PM

I'm not sure how I feel about so many methods having the if dir in self._lazydirs: self._loadlazy(dir) pattern.
On one hand, action at a distance when it involves caching can be dangerous. And doing the lookup inline will avoid a Python function call.
On the other, it seems very redundant.
Overall I'm OK with the patch. I just feel like using a collections.defaultdict or implementing __missing__ to automagically resolve missing keys might work out better. Do you think this is a reasonable request?

I think it's a reasonable request, but I'm not sure it's feasible with how collections.defaultdict/__missing__ work. They are *only* called by getitem. We'd need a separate dict subclass for _dirs that handles contains as well, at the very least, for all the cases of if dir not in self._dirs: return <something>

I'm playing around with such a thing, it's not difficult but I think might be more awkward than helpful. Even if that doesn't bear fruit, while implementing it I noticed something: I'm not actually populating self._lazydirs in this patch, I appear to have lost it during split/merge/whatever. Oops. I have to hope that the numbers I have in the commit description were before I lost this, since otherwise it's completely incomprehensible that this patch would cause a performance benefit (well, maybe not, if dead code can cause a performance loss, why not the other way around?)

spectral edited the summary of this revision. (Show Details)Sep 6 2018, 3:01 PM
spectral updated this revision to Diff 10815.
spectral added a comment.EditedSep 6 2018, 3:08 PM

I'm not sure how I feel about so many methods having the if dir in self._lazydirs: self._loadlazy(dir) pattern.
On one hand, action at a distance when it involves caching can be dangerous. And doing the lookup inline will avoid a Python function call.
On the other, it seems very redundant.
Overall I'm OK with the patch. I just feel like using a collections.defaultdict or implementing __missing__ to automagically resolve missing keys might work out better. Do you think this is a reasonable request?

I think it's a reasonable request, but I'm not sure it's feasible with how collections.defaultdict/__missing__ work. They are *only* called by getitem. We'd need a separate dict subclass for _dirs that handles contains as well, at the very least, for all the cases of if dir not in self._dirs: return <something>
I'm playing around with such a thing, it's not difficult but I think might be more awkward than helpful. Even if that doesn't bear fruit, while implementing it I noticed something: I'm not actually populating self._lazydirs in this patch, I appear to have lost it during split/merge/whatever. Oops. I have to hope that the numbers I have in the commit description were before I lost this, since otherwise it's completely incomprehensible that this patch would cause a performance benefit (well, maybe not, if dead code can cause a performance loss, why not the other way around?)

Turns out that it got quite messy and caused some significant performance loss, though it's likely that at least some of the performance difference was something I did wrong. To handle the two dictionaries, and that things migrate from one to the other, I needed to override __missing__, get, __setitem__, __nonempty__, __iter__, items, iteritems, values, ... I basically had to reimplement all of dict, and that felt a fair amount messier than the duplication here. If you think that might still be preferable, I can try harder to figure out why there's a significant (~15-30% full runtime on some of the commands) performance difference.

This revision was automatically updated to reflect the committed changes.