When a lot of commits are added to DateRevQueue, the
sort-on-insertion approach is very heavy on CPU cycles.
One approach to fix this was made by Dave Borowitz:
https://git.eclipse.org/r/#/c/5491/
But using Java's PriorityQueue seems to have brought some
extra overhead, and the desired performance could not be
reached.
This fix takes another approach to the insertion problem,
without changing the expected behaviour or bringing extra
memory overhead:
If we detect over 1000 commits in the DateRevQueue, a
"seek-index" is rebuilt every 1000th added commit.
The index keeps track of every 100th commit in the
DateRevQueue. During insertions, it will be used for a
preliminary scanning (binary search) of the queue, with
the intention of helping add() find a good starting point
to start walking from. After finding this starting point,
add() will step commit-by-commit until the correct
insertion place in the queue is found (today, the queue
is expected to be sorted at all times).
When applied to repositories with many refs, this approach
has proven to bring huge performance gains and scales quite
well.
For instance, in a repository with close to 80000 refs,
we could cut down the time a typical Gerrit replication
of 1 commit would take (just a push from JGit's point of
view) from 32sec down to 3.5sec.
Below you see some typical times to add a specific amount
of commits (with random commit times) to the DateRevQueue
and the difference the preliminary seek-index makes:
Commits | Index | No Index
1024 8ms 8ms
2048 13ms 9ms
4096 5ms 59ms
8192 11ms 595ms
16384 22ms 3058ms
32768 64ms 13811ms
65536 201ms 62677ms
131072 783ms 331585ms
Only one extra reference is needed for every 100 inserted
commits (and only when we see more than 1000 commits in
the queue), so the memory overhead should be negligible.
Various index-stepping values were tested, and 100 seemed to
scale very well and be effective from start.
In the future, it should probably be dynamic and based on
the number of refs in the queue, but this should serve well
as a starting point.
Note: While other fundamentally different data structures may
be more suitable, the DateRevQueue is extremely central to
many of the Git core operations. This approach was chosen,
since the effect of the patch is easy to predict in conjuction
with the current implementation. A totally new data structure
will make it harder to predict behaviour in many common and
uncommon cases (in terms of breaking ties, memory usage, cost
when using few elements, object creation/disposing overhead,
etc).
Change-Id: Ie7b99f40eacf6324bfb4716d82073adeda64d10f
Gustaf Lundh
84afea9179