On 8/19/2019 9:03 AM, SZEDER Gábor wrote: > On Mon, Aug 19, 2019 at 08:00:11AM -0400, Derrick Stolee wrote: >> On 8/18/2019 2:27 PM, SZEDER Gábor wrote: >>> Line-level log performs a preprocessing step in >>> prepare_revision_walk(), during which it filters and rewrites history >>> to keep only commits modifying the given line range. This >>> preprocessing causes significant delay before the first commit is >>> shown, wastes CPU time when the user asks only for a few commits, and >>> does parent rewriting with no way to turn it off. >>> >>> This patch series addresses these issues by integrating line-level log >>> filtering into the revision walking machinery and making it work >>> together with generation number-based topo-ordering (though for now >>> only in the case when the user doesn't explicitly asks for parent >>> rewriting, which is probably the common case). >>> >>> The first two patches are quite straightforward (and arguably somewhat >>> unrelated), but the rest deals with history traversal and parent >>> rewriting, which I don't usually do, hence the RFC. >> >> Hi Szeder, >> >> Thanks for sending this series! I'm particularly excited about it >> because we recently got a bug report from a user in the Windows OS >> repo about "git log -L" being very slow. I put it on the backlog [1] >> but haven't had time to investigate it myself. After we are done >> updating to v2.23.0 [2], I'll have time to test your changes on >> the Windows repo. I anticipate your change to provide a massive >> boost. > > Well, it depends on what you mean by "boost"... As discussed in patch > 4's commit message, this series doesn't really optimizes much, and the > total amount of work to be done remains the same, except that 'git log > -L... -<N>' will be able to return early. So when you benchmark it > with e.g. 'time git log -L... >/dev/null', then you won't see much > difference, as it will still take just about as looooong as before, > apart from the faster generation numbers-based topo-ordering. (But I > have a few WIP patches waiting to be cleaned up that might bring about > 3-4x speedup in general.) > > In the meantime, until these changes trickle into a Git release, for a > faster line-level log I would recommend to: > > - Limit the revision range up front, i.e.: > > git log -L... ^a-not-too-old-version > > because this can greatly reduce the amount of commits that that > expensive preprocessing step has to churn through. > > - Use '--no-renames'. Although it won't follow the evolution of the > line range over file renames, this might be an acceptable > compromise. (this is what those WIP patches are focusing on) > > - Or both. Usually when I'm testing something involving the --topo-order logic, I'll use a simple "-10" to limit to something similar to a "first page". In this line-log case, I'll use "-1" to just get the top result, as that is essentially how long it takes before the user gets some feedback. Here are the results using a random path I picked out from the Windows repo (it was only changed ~10 times in the 4.5 million commits): Before: real 2m7.308s real 2m8.572s With Patch 4: real 0m38.628s real 0m38.477s With Patch 5: real 0m24.685s real 0m24.310s For the specific file in the bug report from a real user, I got these numbers: real 0m32.293s (patch 4) real 0m19.362s (patch 5) Note that I don't include the "without patch" numbers. For some reason the path provided is particularly nasty and caused 20,000+ missing blobs to be downloaded one-by-one (remember: VFS for Git has many partial-clone-like behaviors). I canceled my test after 55 minutes. I'll dig in more to see what is going on, since only 37 commits actually change that path. Thanks so much for finding these "easy" wins. Your changes are nicely isolated, but give us a target to improve even further! I expect that we will incorporate your commits into microsoft/git very early. Thanks, -Stolee
