patchcommit-reach: fix first-parent heuristic

I originally reported this fix [1] after playing around with the trace2 series for measuring performance. Since trace2 isn't merging quickly, I pulled the performance fix patch out and am sending it on its own. The only difference here is that we don't have the tracing to verify the performance fix in the test script.

See the patch message for details about the fix.

Thanks, -Stolee

[1] https://public-inbox.org/git/20180906151309.66712-7-dstolee@microsoft.com/

[RFC PATCH 6/6] commit-reach: fix first-parent heuristic

Derrick Stolee (1):
  commit-reach: fix first-parent heuristic

 commit-reach.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

base-commit: a4b8ab5363a32f283a61ef3a962853556d136c0e
Published-As: https://github.com/gitgitgadget/git/releases/tags/pr-51%2Fderrickstolee%2Ffirst-parent-heuristic-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-51/derrickstolee/first-parent-heuristic-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/51

From: Derrick Stolee <dstolee@microsoft.com>

The algorithm in can_all_from_reach_with_flags() performs a depth- first-search, terminated by generation number, intending to use a hueristic that "important" commits are found in the first-parent history. This heuristic is valuable in scenarios like fetch negotiation.

However, there is a problem! After the search finds a target commit, it should pop all commits off the stack and mark them as "can reach". This logic is incorrect, so the algorithm instead walks all reachable commits above the generation-number cutoff.

The existing algorithm is still an improvement over the previous algorithm, as the worst-case complexity went from quadratic to linear. The performance measurement at the time was good, but not dramatic. By fixing this heuristic, we reduce the number of walked commits.

We can also re-run the performance tests from commit 4fbcca4e "commit-reach: make can_all_from_reach... linear".

Performance was measured on the Linux repository using 'test-tool reach can_all_from_reach'. The input included rows seeded by tag values. The "small" case included X-rows as v4.[0-9]* and Y-rows as v3.[0-9]*. This mimics a (very large) fetch that says "I have all major v3 releases and want all major v4 releases." The "large" case included X-rows as "v4.*" and Y-rows as "v3.*". This adds all release-candidate tags to the set, which does not greatly increase the number of objects that are considered, but does increase the number of 'from' commits, demonstrating the quadratic nature of the previous code.

Small Case:

4fbcca4e~1: 0.85 s
  4fbcca4e: 0.26 s (num_walked: 1,011,035)
      HEAD: 0.14 s (num_walked:     8,601)

Large Case:

4fbcca4e~1: 24.0  s
  4fbcca4e:  0.12 s (num_walked:  503,925)
      HEAD:  0.06 s (num_walked:  217,243)

Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
---
 commit-reach.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

"Derrick Stolee via GitGitGadget" <gitgitgadget@gmail.com> writes:

These micro-benchmarks are interesting, but we should also remember to describe the impact of the bug being fixed in the larger picture. What end-user visible operations are affected? Computing merge-base? Finding if a push fast-forwards? Something else?

On 10/19/2018 1:24 AM, Junio C Hamano wrote:

Sorry about that. Here is a description that could be inserted into the commit message:

The can_all_from_reach() method synthesizes the logic for one iteration of can_all_from_reach_with_flags() which is used by the server during fetch negotiation to determine if more haves/wants are needed. The logic is also used by is_descendant_of() which is used to check if a force-push is required and in 'git branch --contains'.

Thanks, -Stolee

Derrick Stolee <stolee@gmail.com> writes:

I am afraid that it is still not end-user serving enough. The level of "larger picture" I had in mind was those that would appear as an entry in the release notes, e.g. (this is for illustration purposes only; I do not claim its actual contents is correct).

	We started using commit generation numbers in various
	reachability computations, but due to a bug, negitiation
	between the "git fetch" and the server started to require
	30% more roundtrips than necessary, and it has become less
	efficient to see if a commit is a descendant of another
	commit in certain cases, which has been corrected in this
	release.

Thanks.

"Derrick Stolee via GitGitGadget" <gitgitgadget@gmail.com> writes:

Thanks for sending this separately. What's the current status of the tracing patch series, by the way?

On 10/18/2018 10:31 PM, Junio C Hamano wrote:

I'm still working on it. I hope to reroll and submit a new version next week. We are currently dogfooding a version of it with our gvfs users.

Jeff