[PATCH] write-tree performance problems

Hello everyone,

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.

The primary performance problem during each commit is write-tree recalculating 
the hash of each directory, even though the contents of most directories are 
not changing.  I've attached a very quick and dirty mod of write-tree.c, it 
takes an optional tree id (sha1) and list of directories.  The hash of any 
directories not in the list are read in from existing files instead of being 
recalculated.

You have to pass each sub dir with a modified file.  So, if you change 
fs/super.c and fs/ext3/super.c, you would call "write-tree sha1 fs fs/ext3"
With this patch, the time to apply 100 commits goes down to 22 seconds.  It 
could be faster (and easier to use) if the index stored the hash of trees 
instead of just blobs, but that would be a larger change.

I was able to get the commit time down to 13 seconds by changing read-tree.c, 
update-cache.c and read-cache.c to store/read the index in tmpfs instead of 
on the local filesystem.  I haven't attached the patch for that, but it seems 
easiest to move .git/index into .git/index_dir/index, and let the user decide 
where to put index_dir.

Quick speed summary, apply/commit 100 patches
quilt push -a     :                    2s
git (unmodified):               1m5s
git (tree hash reduction)      22s
git (tree hash, tmpfs index) 13s

This patch is against pasky's tree from this morning, but also applies to 
linus' tree.  It's nasty stuff, but will hopefully get some discussion 
started on speeding things up.

-chris