diff options
| author |  Runxi Yu | 2025-11-17 00:00:00 +0000 |
|---|---|---|
| committer | Runxi Yu | 2025-11-17 00:00:00 +0000 |
| commit | 83c7b6223f467cf47acc6ddec5734711b0657750 (patch) | |
| tree | c03f71b84938a1d399c4f35d7765edf8198094d9 /README.md | |
| parent | pack: Use a Go map with a mutex instead of a sync.Map for packfiles (diff) | |
| signature | ||
Update perf stats
Diffstat (limited to 'README.md')
| -rw-r--r-- | README.md | 34 |
1 files changed, 19 insertions, 15 deletions
@@ -51,25 +51,29 @@ a standard UNIX-like filesystem with ## Performance -Furgit is aggressively optimized for performance. +Furgit is being aggressively optimized for performance. -[As of November 2025](https://git.sr.ht/~runxiyu/gitbench), -for the task of `git ls-tree --long HEAD` on large repos such as -[Linux](https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git), -it is: +It is difficult to optimize Go code to be as performant as libgit2 (or +for that matter, upstream git). However, we are making tiny steps +towards it. -* approximately the same performance as [upstream git](https://git-scm.com), -* approximately 10 times faster than [libgit2](https://libgit2.org), and -* approximately 1000 times faster than [go-git](https://github.com/go-git/go-git). +The first step that has been arguably a success is the packfile parser. +By using memory-mapped I/O, relatively optimized delta resolution, and +zero-copy techniques, Furgit is able to perform the equivalent to +`git ls-tree --long HEAD` on the Linux repository in about 2ms on +a ThinkPad T14, which is comparable to Git, faster than libgit2, +and significantly faster than go-git. -This primarily measures the performance of the packfile parser, which heavily -affects the performance of most operations on large repos; however, there is no -guarantee that this generalizes to *all* Git operations, individual profiling -on specific workloads may be necessary. +However this is a microbenchmark and does not reflect all real-world +performance. For example, when recursively listing tree entires and +commits, Furgit's performance is slightly slower than libgit2, both +lack behind Git by multiple orders of magnitude. -In the future we intend to support -[commit-graph](https://git-scm.com/docs/commit-graph) and research -optimizations thereof. +Things we might consider in the future include: + +* [commit-graph](https://git-scm.com/docs/commit-graph) +* Using a custom zlib implementation to amortize decompression overhead +* More optimizations to delta resolution ## Hash algorithm |