Re: [PATCH] write-tree performance problems

From: Chris Mason <mason@suse.com>
Date: Apr 19, 2005, 18:11 UTC
Message-ID: <200504191412.00227.mason@suse.com>
In-Reply-To: <Pine.LNX.4.58.0504191017300.19286@ppc970.osdl.org>

On Tuesday 19 April 2005 13:36, Linus Torvalds wrote:

Show 13 quoted lines

> On Tue, 19 Apr 2005, Chris Mason wrote:
> > I did a quick experiment with applying/commit 100 patches from the suse
> > kernel into a kernel git tree, which quilt can do in 2 seconds.  git
> > needs 1m5s.
>
> Note that I don't think you want to replace quilt with git. The approaches
> are totally different, and git does _not_ obviate the need for the quilt
> kind of "patch testing".
>
> In fact, git has all the same issues that BK had, and for the same
> fundamental reason: if you do distributed work, you have to always
> "append" stuff, and that means that you can never re-order anything after
> the fact.

Very true, you can't replace quilt with git without ruining both of them. But it would be nice to take a quilt tree and turn it into a git tree for merging purposes, or to make use of whatever visualization tools might exist someday.

Show 8 quoted lines

> What I _would_ like is the ability to re-use an old tree, though. What you
> really want to do is not pass in a set of directory names and just trust
> that they are correct, but just pass in a directory to compare with, and
> if the contents match, you don't need to write out a new one.
>
> I'll try to whip up something that does what you want done, but doesn't
> need (or take) any untrusted information from the user in the form "trust
> me, it hasn't changed".

We already have a "trust me, it hasn't changed" via update-cache. If it gets called wrong the tree won't reflect reality. The patch doesn't change the write-tree default, but does enable you to give write-tree better information about the parts of the tree you want written back to git.

With that said, I hate the patch too. I didn't see how to compare against the old tree without reading each tree object from the old tree, and that should be slower then what write-tree does now. So I wimped out and made the quick patch that demonstrates the cause of the performance hit.

The "move .git/index to a tmpfs file" change should be easier though, and has a real benefit. How do you feel about s|.git/index|.git/index_dir/index| in the sources? This gives us the flexibility to link it wherever is needed.

-chris