[ Dscho cc'd because I think he is the primary culprits for this thing, I
think. Commit e90fdc39b6903502192b2dd11e5503cea721a1ad in particular,
methinks. ]
So because I was looking at system call traces for object creation (due to
those "rename to the final resting place" patches I did that got merged
recently), I've noticed that some commands have *much* worse system call
patterns than others.
In particular, doing a "git add ." will use absolute pathnames for all git
files, while a "git diff" will not. And this is quite noticeable - the
absolute pathnames are not just longer, they have more path components in
them. Making them a lot slower to look up and use.
Of course, "a lot" depends on things a bit, but it really is noticeable.
To test, I created a kernel tree (no .git), and did a "git init" followed
by a "git add .". Here's the timings for current git head:
[torvalds@woody kernel]$ time git add .
real 0m8.377s
user 0m6.556s
sys 0m1.656s
After that, I fixed "write_loose_object()" to not unnecessarily try to
open a git object file, because every single caller has already done a
"has_sha1_file(sha1)" or "has_loose_object(sha1)" check before calling
that function, so trying to open it again is just pointless.
As a result, git add sped up a tiny bit:
[torvalds@woody kernel]$ time ~/git/git-add .
real 0m8.341s
user 0m6.588s
sys 0m1.548s
but I'll admit that it's not exactly noticeable. Half a percentage point
is not a big deal.
Then I created a hack that just made "setup_work_tree()" a no-op (because
it does all that crazy stuff that forces GIT_DIR to be an absolute path
etc). As a result I got:
[torvalds@woody kernel]$ time ~/git/git-add .
real 0m7.849s
user 0m6.420s
sys 0m1.296s
ie now we're talking about a 5%+ performance difference.
Of course, this is all for the hot-cache case, and it wouldn't be
noticeable for a cold-cache case, but it really can be a real performance
issue. And no, it's not that "setup_work_tree()" itself is expensive, but
depending on how GIT_DIR is set up, you get very different system call
patterns.
Here's what current git does for one file (sound/usb/usbmixer_maps.c):
lstat("sound/usb/usbmixer_maps.c", {st_mode=S_IFREG|0664, st_size=10230, ...}) = 0
open("sound/usb/usbmixer_maps.c", O_RDONLY) = 4
mmap(NULL, 10230, PROT_READ, MAP_PRIVATE, 4, 0) = 0x7fa247228000
close(4) = 0
access("/home/torvalds/kernel/.git/objects/d7/55be0ad8115da5c1296211678f81a7774277d5", F_OK) = -1 ENOENT (No such file or directory)
open("/home/torvalds/kernel/.git/objects/d7/55be0ad8115da5c1296211678f81a7774277d5", O_RDONLY) = -1 ENOENT (No such file or directory)
open("/home/torvalds/kernel/.git/objects/d7/tmp_obj_55be0ad8115da5c12962116_vDqXRl", O_RDWR|O_CREAT|O_EXCL, 0600) = 4
write(4, "x\1\315ZmS\333H\22\276\257\370W\364\361\t\210\1K\206$\300.\273\6\314\306\273@(l\'"..., 3402) = 3402
fchmod(4, 0444) = 0
close(4) = 0
link("/home/torvalds/kernel/.git/objects/d7/tmp_obj_55be0ad8115da5c12962116_vDqXRl", "/home/torvalds/kernel/.git/objects/d7/55be0ad8115da5c1296211678f81a7774277d5") = 0
unlink("/home/torvalds/kernel/.git/objects/d7/tmp_obj_55be0ad8115da5c12962116_vDqXRl") = 0
munmap(0x7fa247228000, 10230) = 0
and here is the fixed version:
lstat("sound/usb/usbmixer_maps.c", {st_mode=S_IFREG|0664, st_size=10230, ...}) = 0
open("sound/usb/usbmixer_maps.c", O_RDONLY) = 4
mmap(NULL, 10230, PROT_READ, MAP_PRIVATE, 4, 0) = 0x7f54c231a000
close(4) = 0
access(".git/objects/d7/55be0ad8115da5c1296211678f81a7774277d5", F_OK) = -1 ENOENT (No such file or directory)
open(".git/objects/d7/tmp_obj_55be0ad8115da5c12962116_fKy6px", O_RDWR|O_CREAT|O_EXCL, 0600) = 4
write(4, "x\1\315ZmS\333H\22\276\257\370W\364\361\t\210\1K\206$\300.\273\6\314\306\273@(l\'"..., 3402) = 3402
fchmod(4, 0444) = 0
close(4) = 0
link(".git/objects/d7/tmp_obj_55be0ad8115da5c12962116_fKy6px", ".git/objects/d7/55be0ad8115da5c1296211678f81a7774277d5") = 0
unlink(".git/objects/d7/tmp_obj_55be0ad8115da5c12962116_fKy6px") = 0
munmap(0x7f54c231a000, 10230) = 0
ie note how it does one unnecessary "open()" less, but more importantly,
notice the difference between "/home/torvalds/kernel/.git/*" and ".git/*",
and realize that that second difference was the much more noticeable one.
And no, obviously the right fix is not to just comment out all of
"setup_work_tree()" (it will break stuff that depends on GIT_WORKTREE),
but I did that as a minimal example of showing what the bad effect of that
function is.
In general, I think we've gone in the wrong direction with a lot of the
"make_absolute_path" stuff. See above. 5% performance loss is not good.
Linus
--
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
