Performance regression when reading files from the tree in parallel

[Guillaume Grossetie]

Reproduction steps

1. Add the following test cases in CommitFixture.cs:

[Fact]
public void CanReadCommit()
{
    var timer = new Stopwatch();
    var fileContents = new ConcurrentStack<string>();
    var path = SandboxStandardTestRepo();
    using (var repo = new Repository(path))
    {
        var latestCommit = repo.Head.Tip;
        var tree = latestCommit.Tree;
        timer.Start();
        for (var i = 0; i < 100000; i++)
        {
            fileContents.Push(ReadEntry("1.txt", tree));
            fileContents.Push(ReadEntry("README", tree));
            fileContents.Push(ReadEntry("new.txt", tree));
        }
        timer.Stop();
        testOutputHelper.WriteLine($"Took: {timer.ElapsedMilliseconds.ToString()}ms");
        Assert.Equal(300000, fileContents.Count);
    }
}

[Fact]
public void CanReadCommitParallel()
{
    var timer = new Stopwatch();
    var fileContents = new ConcurrentStack<string>();
    var path = SandboxStandardTestRepo();
    using (var repo = new Repository(path))
    {
        var latestCommit = repo.Head.Tip;
        var tree = latestCommit.Tree;
        timer.Start();
        var fileNames = new List<string>() {"1.txt", "README", "new.txt"}.AsEnumerable();
        Parallel.ForEach(fileNames, (fileName) =>
        {
            for (var i = 0; i < 100000; i++)
            {
                fileContents.Push(ReadEntry(fileName, tree));
            }
        });
        timer.Stop();
        testOutputHelper.WriteLine($"Took: {timer.ElapsedMilliseconds.ToString()}ms");
        Assert.Equal(300000, fileContents.Count);
    }
}

private static string ReadEntry(string name, Tree tree)
{
    var treeEntry = tree[name];
    if (treeEntry != null && treeEntry.Target is Blob blob)
    {
        return blob.GetContentText();
    }

    throw new InvalidOperationException($"{name} must be a Blob");
}

2. Run against version 0.27.0-preview-0119 (commit: 6329bea). On my machine, CanReadCommitParallel takes 1417ms and CanReadCommit takes 2971ms (which is fine)
3. Checkout the latest version or the latest release
4. Run the test again. On my machine, CanReadCommitParallel takes 3731ms and CanReadCommit takes 3250ms!

Expected behavior

Reading files in parallel should be faster than reading files sequentially from the git tree.

Actual behavior

It seems that reading files from the tree in parallel (multi-thread) is not faster. I did a git bisect and it seems that this regression was introduced in 21d4f13

Version of LibGit2Sharp (release number or SHA1)

Versions after 21d4f13

Operating system(s) tested; .NET runtime tested

.NET 6 and .NET 7.

[Guillaume Grossetie]

I did another git bisect (while making sure that bin was deleted) and I can reproduce it with 21d4f13 where libgit2 was updated to 1.2.0

[Edward Thomson]

That was two and a half years ago. A lot has changed since then. It's going to be tricky to isolate the problem.

One question to start - what's the perf if you have one repository per thread instead of sharing the repository across threads?

[Guillaume Grossetie]

Thanks for your reply!

One question to start - what's the perf if you have one repository per thread instead of sharing the repository across threads?

You mean one Repository object (using the same path) per thread? I can try that 👍🏻
Also, I did try to implement a similar benchmark test in libgit2 but my skills in C are insufficient 😅

[Guillaume Grossetie]

One question to start - what's the perf if you have one repository per thread instead of sharing the repository across threads?

Spot on! Using one Repository instance per thread does the trick 🙌🏻
Should we document how to properly use libgit2(sharp) in a multi-thread environment? Or maybe it's possible to (re)enable concurrent reads on a single Repository instance?

[Edward Thomson]

Right — really good question. I was mostly hoping to stem the bleeding and get you to a performant situation. Digging in to the why it's slow would be really interesting.

https://github.com/libgit2/libgit2/blob/main/docs/threading.md#sharing-objects has a bit of a discussion about libgit2's threading policy (which LibGit2Sharp should also document as it should be identical — I don't think that there's anything in LibGit2Sharp that makes threading any different).

Without actually doing a serious investigation: probably you're hitting a lock on an object cache. Why that got worse, I don't know, and we should 👀 that. But the most performant way to do what you're doing is multiple Repository instances (although there will be a bit of a memory usage increase, since you won't have a cache shared across a single repository).

[Guillaume Grossetie]

Thanks again for your insight.

libgit2/libgit2@main/docs/threading.md#sharing-objects has a bit of a discussion about libgit2's threading policy (which LibGit2Sharp should also document as it should be identical

👍🏻

I don't think that there's anything in LibGit2Sharp that makes threading any different
Without actually doing a serious investigation: probably you're hitting a lock on an object cache

Should we move this issue to https://github.com/libgit2/libgit2?

But the most performant way to do what you're doing is multiple Repository instances (although there will be a bit of a memory usage increase, since you won't have a cache shared across a single repository).

Alright!
I wasn't sure that it was indeed the most performant solution. I will update my code base accordingly. Thank you!