When a changed-path Bloom filter has either zero, or more than a
certain number (commonly 512) of entries, the commit-graph machinery
encodes it as "missing". More specifically, it sets the indices adjacent
in the BIDX chunk as equal to each other to indicate a "length 0"
filter; that is, that the filter occupies zero bytes on disk.
This has heretofore been fine, since the commit-graph machinery has no
need to care about these filters with too few or too many changed paths.
Both cases act like no filter has been generated at all, and so there is
no need to store them.
In a subsequent commit, however, the commit-graph machinery will learn
to only compute Bloom filters for some commits in the current
commit-graph layer. This is a change from the current implementation
which computes Bloom filters for all commits that are in the layer being
written. Critically for this patch, only computing some of the Bloom
filters means adding a third state for length 0 Bloom filters: zero
entries, too many entries, or "hasn't been computed".
It will be important for that future patch to distinguish between "not
representable" (i.e., zero or too-many changed paths), and "hasn't been
computed". In particular, we don't want to waste time recomputing
filters that have already been computed.
To that end, change how we store Bloom filters in the "computed but not
representable" category:
- Bloom filters with no entries are stored as a single byte with all
bits low (i.e., all queries to that Bloom filter will return
"definitely not")
- Bloom filters with too many entries are stored as a single byte with
all bits set high (i.e., all queries to that Bloom filter will
return "maybe").
These rules are sufficient to not incur a behavior change by changing
the on-disk representation of these two classes. Likewise, no
specification changes are necessary for the commit-graph format, either:
- Filters that were previously empty will be recomputed and stored
according to the new rules, and
- old clients reading filters generated by new clients will interpret
the filters correctly and be none the wiser to how they were
generated.
Clients will invoke the Bloom machinery in more cases than before, but
this can be addressed by returning a NULL filter when all bits are set
high. This can be addressed in a future patch.
Note that this does increase the size of on-disk commit-graphs, but far
less than other proposals. In particular, this is generally more
efficient than storing a bitmap for which commits haven't computed their
Bloom filters. Storing a bitmap incurs a penalty of one bit per commit,
whereas storing explicit filters as above incurs a penalty of one byte
per too-large or empty commit.
In practice, these boundary commits likely occupy a small proportion of
the overall number of commits, and so the size penalty is likely smaller
than storing a bitmap for all commits.
See, for example, these relative proportions of such boundary commits
(collected by SZEDER Gábor):
| Percentage of | commit-graph | |
| commits modifying | file size | |
├────────┬──────────────┼───────────────────┤ pct. |
| 0 path | >= 512 paths | before | after | change |
┌────────────────┼────────┼──────────────┼─────────┼─────────┼───────────┤
| android-base | 13.20% | 0.13% | 37.468M | 37.534M | +0.1741 % |
| cmssw | 0.15% | 0.23% | 17.118M | 17.119M | +0.0091 % |
| cpython | 3.07% | 0.01% | 7.967M | 7.971M | +0.0423 % |
| elasticsearch | 0.70% | 1.00% | 8.833M | 8.835M | +0.0128 % |
| gcc | 0.00% | 0.08% | 16.073M | 16.074M | +0.0030 % |
| gecko-dev | 0.14% | 0.64% | 59.868M | 59.874M | +0.0105 % |
| git | 0.11% | 0.02% | 3.895M | 3.895M | +0.0020 % |
| glibc | 0.02% | 0.10% | 3.555M | 3.555M | +0.0021 % |
| go | 0.00% | 0.07% | 3.186M | 3.186M | +0.0018 % |
| homebrew-cask | 0.40% | 0.02% | 7.035M | 7.035M | +0.0065 % |
| homebrew-core | 0.01% | 0.01% | 11.611M | 11.611M | +0.0002 % |
| jdk | 0.26% | 5.64% | 5.537M | 5.540M | +0.0590 % |
| linux | 0.01% | 0.51% | 63.735M | 63.740M | +0.0073 % |
| llvm-project | 0.12% | 0.03% | 25.515M | 25.516M | +0.0050 % |
| rails | 0.10% | 0.10% | 6.252M | 6.252M | +0.0027 % |
| rust | 0.07% | 0.17% | 9.364M | 9.364M | +0.0033 % |
| tensorflow | 0.09% | 1.02% | 7.009M | 7.010M | +0.0158 % |
| webkit | 0.05% | 0.31% | 17.405M | 17.406M | +0.0047 % |
(where the above increase is determined by computing a non-split
commit-graph before and after this patch).
Given that these projects are all "large" by commit count, the storage
cost by writing these filters explicitly is negligible. In the most
extreme example, android-base (which has 494,848 commits at the time of
writing) would have its commit-graph increase by a modest 68.4 KB.
Finally, a test to exercise filters which contain too many changed path
entries will be introduced in a subsequent patch.
Suggested-by: SZEDER Gábor <szeder.dev@gmail.com>
Suggested-by: Jakub Narębski <jnareb@gmail.com>
Helped-by: Derrick Stolee <dstolee@microsoft.com>
Helped-by: SZEDER Gábor <szeder.dev@gmail.com>
Helped-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When 'get_or_compute_bloom_filter()' needs to compute a Bloom filter
from scratch, it looks to the default 'struct bloom_filter_settings' in
order to determine the maximum number of changed paths, number of bits
per entry, and so on.
All of these values have so far been constant, and so there was no need
to pass in a pointer from the caller (eg., the one that is stored in the
'struct write_commit_graph_context').
Start passing in a 'struct bloom_filter_settings *' instead of using the
default values to respect graph-specific settings (eg., in the case of
setting 'GIT_TEST_BLOOM_SETTINGS_MAX_CHANGED_PATHS').
In order to have an initialized value for these settings, move its
initialization to earlier in the commit-graph write.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
'get_bloom_filter' takes a flag to control whether it will compute a
Bloom filter if the requested one is missing. In the next patch, we'll
add yet another parameter to this method, which would force all but one
caller to specify an extra 'NULL' parameter at the end.
Instead of doing this, split 'get_bloom_filter' into two functions:
'get_bloom_filter' and 'get_or_compute_bloom_filter'. The former only
looks up a Bloom filter (and does not compute one if it's missing,
thus dropping the 'compute_if_not_present' flag). The latter does
compute missing Bloom filters, with an additional parameter to store
whether or not it needed to do so.
This simplifies many call-sites, since the majority of existing callers
to 'get_bloom_filter' do not want missing Bloom filters to be computed
(so they can drop the parameter entirely and use the simpler version of
the function).
While we're at it, instrument the new 'get_or_compute_bloom_filter()'
with counters in the 'write_commit_graph_context' struct which store
the number of filters that we did and didn't compute, as well as filters
that were truncated.
It would be nice to drop the 'compute_if_not_present' flag entirely,
since all remaining callers of 'get_or_compute_bloom_filter' pass it as
'1', but this will change in a future patch and hence cannot be removed.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
For now, we assume that there is a fixed constant describing the
maximum number of changed paths we are willing to store in a Bloom
filter.
Prepare for that to (at least partially) not be the case by making it a
member of the 'struct bloom_filter_settings'. This will be helpful in
the subsequent patches by reducing the size of test cases that exercise
storing too many changed paths, as well as preparing for an eventual
future in which this value might change.
This patch alone does not cause newly generated Bloom filters to use
a custom upper-bound on the maximum number of changed paths a single
Bloom filter can hold, that will occur in a later patch.
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
"git log -L..." now takes advantage of the "which paths are touched
by this commit?" info stored in the commit-graph system.
* ds/line-log-on-bloom:
line-log: integrate with changed-path Bloom filters
line-log: try to use generation number-based topo-ordering
line-log: more responsive, incremental 'git log -L'
t4211-line-log: add tests for parent oids
line-log: remove unused fields from 'struct line_log_data'
Code cleanup and typofixes
* ds/bloom-cleanup:
completion: offer '--(no-)patch' among 'git log' options
bloom: use num_changes not nr for limit detection
bloom: de-duplicate directory entries
Documentation: changed-path Bloom filters use byte words
bloom: parse commit before computing filters
test-bloom: fix usage typo
bloom: fix whitespace around tab length
The previous changes to the line-log machinery focused on making the
first result appear faster. This was achieved by no longer walking the
entire commit history before returning the early results. There is still
another way to improve the performance: walk most commits much faster.
Let's use the changed-path Bloom filters to reduce time spent computing
diffs.
Since the line-log computation requires opening blobs and checking the
content-diff, there is still a lot of necessary computation that cannot
be replaced with changed-path Bloom filters. The part that we can reduce
is most effective when checking the history of a file that is deep in
several directories and those directories are modified frequently. In
this case, the computation to check if a commit is TREESAME to its first
parent takes a large fraction of the time. That is ripe for improvement
with changed-path Bloom filters.
We must ensure that prepare_to_use_bloom_filters() is called in
revision.c so that the bloom_filter_settings are loaded into the struct
rev_info from the commit-graph. Of course, some cases are still
forbidden, but in the line-log case the pathspec is provided in a
different way than normal.
Since multiple paths and segments could be requested, we compute the
struct bloom_key data dynamically during the commit walk. This could
likely be improved, but adds code complexity that is not valuable at
this time.
There are two cases to care about: merge commits and "ordinary" commits.
Merge commits have multiple parents, but if we are TREESAME to our first
parent in every range, then pass the blame for all ranges to the first
parent. Ordinary commits have the same condition, but each is done
slightly differently in the process_ranges_[merge|ordinary]_commit()
methods. By checking if the changed-path Bloom filter can guarantee
TREESAME, we can avoid that tree-diff cost. If the filter says "probably
changed", then we need to run the tree-diff and then the blob-diff if
there was a real edit.
The Linux kernel repository is a good testing ground for the performance
improvements claimed here. There are two different cases to test. The
first is the "entire history" case, where we output the entire history
to /dev/null to see how long it would take to compute the full line-log
history. The second is the "first result" case, where we find how long
it takes to show the first value, which is an indicator of how quickly a
user would see responses when waiting at a terminal.
To test, I selected the paths that were changed most frequently in the
top 10,000 commits using this command (stolen from StackOverflow [1]):
git log --pretty=format: --name-only -n 10000 | sort | \
uniq -c | sort -rg | head -10
which results in
121 MAINTAINERS
63 fs/namei.c
60 arch/x86/kvm/cpuid.c
59 fs/io_uring.c
58 arch/x86/kvm/vmx/vmx.c
51 arch/x86/kvm/x86.c
45 arch/x86/kvm/svm.c
42 fs/btrfs/disk-io.c
42 Documentation/scsi/index.rst
(along with a bogus first result). It appears that the path
arch/x86/kvm/svm.c was renamed, so we ignore that entry. This leaves the
following results for the real command time:
| | Entire History | First Result |
| Path | Before | After | Before | After |
|------------------------------|--------|--------|--------|--------|
| MAINTAINERS | 4.26 s | 3.87 s | 0.41 s | 0.39 s |
| fs/namei.c | 1.99 s | 0.99 s | 0.42 s | 0.21 s |
| arch/x86/kvm/cpuid.c | 5.28 s | 1.12 s | 0.16 s | 0.09 s |
| fs/io_uring.c | 4.34 s | 0.99 s | 0.94 s | 0.27 s |
| arch/x86/kvm/vmx/vmx.c | 5.01 s | 1.34 s | 0.21 s | 0.12 s |
| arch/x86/kvm/x86.c | 2.24 s | 1.18 s | 0.21 s | 0.14 s |
| fs/btrfs/disk-io.c | 1.82 s | 1.01 s | 0.06 s | 0.05 s |
| Documentation/scsi/index.rst | 3.30 s | 0.89 s | 1.46 s | 0.03 s |
It is worth noting that the least speedup comes for the MAINTAINERS file
which is
* edited frequently,
* low in the directory heirarchy, and
* quite a large file.
All of those points lead to spending more time doing the blob diff and
less time doing the tree diff. Still, we see some improvement in that
case and significant improvement in other cases. A 2-4x speedup is
likely the more typical case as opposed to the small 5% change for that
file.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* We need a `final_new_line` to make our source code as text file, per
POSIX and C specification.
* `bloom_filters` should be limited to interal linkage only
Signed-off-by: Đoàn Trần Công Danh <congdanhqx@gmail.com>
Signed-off-by: Ramsay Jones <ramsay@ramsayjones.plus.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fix alignment issues that were likely introduced due to an editor
using tab lengths of 4 instead of 8.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Revision walk will now use Bloom filters for commits to speed up
revision walks for a particular path (for computing history for
that path), if they are present in the commit-graph file.
We load the Bloom filters during the prepare_revision_walk step,
currently only when dealing with a single pathspec. Extending
it to work with multiple pathspecs can be explored and built on
top of this series in the future.
While comparing trees in rev_compare_trees(), if the Bloom filter
says that the file is not different between the two trees, we don't
need to compute the expensive diff. This is where we get our
performance gains. The other response of the Bloom filter is '`:maybe',
in which case we fall back to the full diff calculation to determine
if the path was changed in the commit.
We do not try to use Bloom filters when the '--walk-reflogs' option
is specified. The '--walk-reflogs' option does not walk the commit
ancestry chain like the rest of the options. Incorporating the
performance gains when walking reflog entries would add more
complexity, and can be explored in a later series.
Performance Gains:
We tested the performance of `git log -- <path>` on the git repo, the linux
and some internal large repos, with a variety of paths of varying depths.
On the git and linux repos:
- we observed a 2x to 5x speed up.
On a large internal repo with files seated 6-10 levels deep in the tree:
- we observed 10x to 20x speed ups, with some paths going up to 28 times
faster.
Helped-by: Derrick Stolee <dstolee@microsoft.com
Helped-by: SZEDER Gábor <szeder.dev@gmail.com>
Helped-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Garima Singh <garima.singh@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add logic to
a) parse Bloom filter information from the commit graph file and,
b) re-use existing Bloom filters.
See Documentation/technical/commit-graph-format for the format in which
the Bloom filter information is written to the commit graph file.
To read Bloom filter for a given commit with lexicographic position
'i' we need to:
1. Read BIDX[i] which essentially gives us the starting index in BDAT for
filter of commit i+1. It is essentially the index past the end
of the filter of commit i. It is called end_index in the code.
2. For i>0, read BIDX[i-1] which will give us the starting index in BDAT
for filter of commit i. It is called the start_index in the code.
For the first commit, where i = 0, Bloom filter data starts at the
beginning, just past the header in the BDAT chunk. Hence, start_index
will be 0.
3. The length of the filter will be end_index - start_index, because
BIDX[i] gives the cumulative 8-byte words including the ith
commit's filter.
We toggle whether Bloom filters should be recomputed based on the
compute_if_not_present flag.
Helped-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Garima Singh <garima.singh@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add the core implementation for computing Bloom filters for
the paths changed between a commit and it's first parent.
We fill the Bloom filters as (const char *data, int len) pairs
as `struct bloom_filters" within a commit slab.
Filters for commits with no changes and more than 512 changes,
is represented with a filter of length zero. There is no gain
in distinguishing between a computed filter of length zero for
a commit with no changes, and an uncomputed filter for new commits
or for commits with more than 512 changes. The effect on
`git log -- path` is the same in both cases. We will fall back to
the normal diffing algorithm when we can't benefit from the
existence of Bloom filters.
Helped-by: Jeff King <peff@peff.net>
Helped-by: Derrick Stolee <dstolee@microsoft.com>
Reviewed-by: Jakub Narębski <jnareb@gmail.com>
Signed-off-by: Garima Singh <garima.singh@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Introduce the constructs for Bloom filters, Bloom filter keys
and Bloom filter settings.
For details on what Bloom filters are and how they work, refer
to Dr. Derrick Stolee's blog post [1]. It provides a concise
explanation of the adoption of Bloom filters as described in
[2] and [3].
Implementation specifics:
1. We currently use 7 and 10 for the number of hashes and the
size of each entry respectively. They served as great starting
values, the mathematical details behind this choice are
described in [1] and [4]. The implementation, while not
completely open to it at the moment, is flexible enough to allow
for tweaking these settings in the future.
Note: The performance gains we have observed with these values
are significant enough that we did not need to tweak these
settings. The performance numbers are included in the cover letter
of this series and in the commit message of the subsequent commit
where we use Bloom filters to speed up `git log -- path`.
2. As described in [1] and [3], we do not need 7 independent hashing
functions. We use the Murmur3 hashing scheme, seed it twice and
then combine those to procure an arbitrary number of hash values.
3. The filters will be sized according to the number of changes in
each commit, in multiples of 8 bit words.
[1] Derrick Stolee
"Supercharging the Git Commit Graph IV: Bloom Filters"
https://devblogs.microsoft.com/devops/super-charging-the-git-commit-graph-iv-Bloom-filters/
[2] Flavio Bonomi, Michael Mitzenmacher, Rina Panigrahy, Sushil Singh, George Varghese
"An Improved Construction for Counting Bloom Filters"
http://theory.stanford.edu/~rinap/papers/esa2006b.pdfhttps://doi.org/10.1007/11841036_61
[3] Peter C. Dillinger and Panagiotis Manolios
"Bloom Filters in Probabilistic Verification"
http://www.ccs.neu.edu/home/pete/pub/Bloom-filters-verification.pdfhttps://doi.org/10.1007/978-3-540-30494-4_26
[4] Thomas Mueller Graf, Daniel Lemire
"Xor Filters: Faster and Smaller Than Bloom and Cuckoo Filters"
https://arxiv.org/abs/1912.08258
Helped-by: Derrick Stolee <dstolee@microsoft.com>
Reviewed-by: Jakub Narębski <jnareb@gmail.com>
Signed-off-by: Garima Singh <garima.singh@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In preparation for computing changed paths Bloom filters,
implement the Murmur3 hash algorithm as described in [1].
It hashes the given data using the given seed and produces
a uniformly distributed hash value.
[1] https://en.wikipedia.org/wiki/MurmurHash#Algorithm
Helped-by: Derrick Stolee <dstolee@microsoft.com>
Helped-by: Szeder Gábor <szeder.dev@gmail.com>
Reviewed-by: Jakub Narębski <jnareb@gmail.com>
Signed-off-by: Garima Singh <garima.singh@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Taylor Blau
Junio C Hamano
Derrick Stolee
Đoàn Trần Công Danh
Garima Singh