When reachability bitmap coverage exists in a repository, Git will use a
different (and hopefully faster) traversal to compute revision walks.
Consider a set of positive and negative tips (which we'll refer to with
their standard bitmap parlance by "wants", and "haves"). In order to
figure out what objects exist between the tips, the existing traversal
in `prepare_bitmap_walk()` does something like:
1. Consider if we can even compute the set of objects with bitmaps,
and fall back to the usual traversal if we cannot. For example,
pathspec limiting traversals can't be computed using bitmaps (since
they don't know which objects are at which paths). The same is true
of certain kinds of non-trivial object filters.
2. If we can compute the traversal with bitmaps, partition the
(dereferenced) tips into two object lists, "haves", and "wants",
based on whether or not the objects have the UNINTERESTING flag,
respectively.
3. Fall back to the ordinary object traversal if either (a) there are
more than zero haves, none of which are in the bitmapped pack or
MIDX, or (b) there are no wants.
4. Construct a reachability bitmap for the "haves" side by walking
from the revision tips down to any existing bitmaps, OR-ing in any
bitmaps as they are found.
5. Then do the same for the "wants" side, stopping at any objects that
appear in the "haves" bitmap.
6. Filter the results if any object filter (that can be easily
computed with bitmaps alone) was given, and then return back to the
caller.
When there is good bitmap coverage relative to the traversal tips, this
walk is often significantly faster than an ordinary object traversal
because it can visit far fewer objects.
But in certain cases, it can be significantly *slower* than the usual
object traversal. Why? Because we need to compute complete bitmaps on
either side of the walk. If either one (or both) of the sides require
walking many (or all!) objects before they get to an existing bitmap,
the extra bitmap machinery is mostly or all overhead.
One of the benefits, however, is that even if the walk is slower, bitmap
traversals are guaranteed to provide an *exact* answer. Unlike the
traditional object traversal algorithm, which can over-count the results
by not opening trees for older commits, the bitmap walk builds an exact
reachability bitmap for either side, meaning the results are never
over-counted.
But producing non-exact results is OK for our traversal here (both in
the bitmap case and not), as long as the results are over-counted, not
under.
Relaxing the bitmap traversal to allow it to produce over-counted
results gives us the opportunity to make some significant improvements.
Instead of the above, the new algorithm only has to walk from the
*boundary* down to the nearest bitmap, instead of from each of the
UNINTERESTING tips.
The boundary-based approach still has degenerate cases, but we'll show
in a moment that it is often a significant improvement.
The new algorithm works as follows:
1. Build a (partial) bitmap of the haves side by first OR-ing any
bitmap(s) that already exist for UNINTERESTING commits between the
haves and the boundary.
2. For each commit along the boundary, add it as a fill-in traversal
tip (where the traversal terminates once an existing bitmap is
found), and perform fill-in traversal.
3. Build up a complete bitmap of the wants side as usual, stopping any
time we intersect the (partial) haves side.
4. Return the results.
And is more-or-less equivalent to using the *old* algorithm with this
invocation:
$ git rev-list --objects --use-bitmap-index $WANTS --not \
$(git rev-list --objects --boundary $WANTS --not $HAVES |
perl -lne 'print $1 if /^-(.*)/')
The new result performs significantly better in many cases, particularly
when the distance from the boundary commit(s) to an existing bitmap is
shorter than the distance from (all of) the have tips to the nearest
bitmapped commit.
Note that when using the old bitmap traversal algorithm, the results can
be *slower* than without bitmaps! Under the new algorithm, the result is
computed faster with bitmaps than without (at the cost of over-counting
the true number of objects in a similar fashion as the non-bitmap
traversal):
# (Computing the number of tagged objects not on any branches
# without bitmaps).
$ time git rev-list --count --objects --tags --not --branches
20
real 0m1.388s
user 0m1.092s
sys 0m0.296s
# (Computing the same query using the old bitmap traversal).
$ time git rev-list --count --objects --tags --not --branches --use-bitmap-index
19
real 0m22.709s
user 0m21.628s
sys 0m1.076s
# (this commit)
$ time git.compile rev-list --count --objects --tags --not --branches --use-bitmap-index
19
real 0m1.518s
user 0m1.234s
sys 0m0.284s
The new algorithm is still slower than not using bitmaps at all, but it
is nearly a 15-fold improvement over the existing traversal.
In a more realistic setting (using my local copy of git.git), I can
observe a similar (if more modest) speed-up:
$ argv="--count --objects --branches --not --tags"
hyperfine \
-n 'no bitmaps' "git.compile rev-list $argv" \
-n 'existing traversal' "git.compile rev-list --use-bitmap-index $argv" \
-n 'boundary traversal' "git.compile -c pack.useBitmapBoundaryTraversal=true rev-list --use-bitmap-index $argv"
Benchmark 1: no bitmaps
Time (mean ± σ): 124.6 ms ± 2.1 ms [User: 103.7 ms, System: 20.8 ms]
Range (min … max): 122.6 ms … 133.1 ms 22 runs
Benchmark 2: existing traversal
Time (mean ± σ): 368.6 ms ± 3.0 ms [User: 325.3 ms, System: 43.1 ms]
Range (min … max): 365.1 ms … 374.8 ms 10 runs
Benchmark 3: boundary traversal
Time (mean ± σ): 167.6 ms ± 0.9 ms [User: 139.5 ms, System: 27.9 ms]
Range (min … max): 166.1 ms … 169.2 ms 17 runs
Summary
'no bitmaps' ran
1.34 ± 0.02 times faster than 'boundary traversal'
2.96 ± 0.05 times faster than 'existing traversal'
Here, the new algorithm is also still slower than not using bitmaps, but
represents a more than 2-fold improvement over the existing traversal in
a more modest example.
Since this algorithm was originally written (nearly a year and a half
ago, at the time of writing), the bitmap lookup table shipped, making
the new algorithm's result more competitive. A few other future
directions for improving bitmap traversal times beyond not using bitmaps
at all:
- Decrease the cost to decompress and OR together many bitmaps
together (particularly when enumerating the uninteresting side of
the walk). Here we could explore more efficient bitmap storage
techniques, like Roaring+Run and/or use SIMD instructions to speed
up ORing them together.
- Store pseudo-merge bitmaps, which could allow us to OR together
fewer "summary" bitmaps (which would also help with the above).
Helped-by: Jeff King <peff@peff.net>
Helped-by: Derrick Stolee <derrickstolee@github.com>
Signed-off-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
b0afdce5da