Re: slow git-cherry-pick.

3 messages, 3 authors, 2016-06-15 · open the first message on its own page

Re: slow git-cherry-pick.

From: Thomas Rast <hidden>
Date: 2016-06-15 22:59:22

Paweł Sikora [off-list ref] writes:
Umm, there's a gem here that the thread missed so far:
my git repo isn't very big[1] but it's checked out on the linear lvm
where random i/o generally hurts and strace shows that current git version
performs 2x{lstat}+1x{open,read,close} [2] on whole checkout before
           ^^^^^^^^^

There's no reason why it should do the lstat() *twice* for every file.
But Paweł is right; the code path roughly goes like this:

int cmd_cherry_pick(int argc, const char **argv, const char *prefix)
{
[...]
	res = sequencer_pick_revisions(&opts);

int sequencer_pick_revisions(struct replay_opts *opts)
{
[...]
	read_and_refresh_cache(opts);
[...]
	return pick_commits(todo_list, opts);
}

static int pick_commits(struct commit_list *todo_list, struct replay_opts *opts)
{
[...]
	read_and_refresh_cache(opts);


I'm too tired to dig further, but AFAICT it's just a rather obvious case
of duplication of effort.

-- 
Thomas Rast
tr@thomasrast.ch

Re: slow git-cherry-pick.

From: Duy Nguyen <hidden>
Date: 2016-06-15 22:59:22

On Wed, Dec 4, 2013 at 3:13 AM, Thomas Rast [off-list ref] wrote:
Paweł Sikora [off-list ref] writes:
quoted
Umm, there's a gem here that the thread missed so far:
quoted
my git repo isn't very big[1] but it's checked out on the linear lvm
where random i/o generally hurts and strace shows that current git version
performs 2x{lstat}+1x{open,read,close} [2] on whole checkout before
           ^^^^^^^^^

There's no reason why it should do the lstat() *twice* for every file.
But Paweł is right; the code path roughly goes like this:

int cmd_cherry_pick(int argc, const char **argv, const char *prefix)
{
[...]
        res = sequencer_pick_revisions(&opts);

int sequencer_pick_revisions(struct replay_opts *opts)
{
[...]
        read_and_refresh_cache(opts);
[...]
        return pick_commits(todo_list, opts);
}

static int pick_commits(struct commit_list *todo_list, struct replay_opts *opts)
{
[...]
        read_and_refresh_cache(opts);


I'm too tired to dig further, but AFAICT it's just a rather obvious case
of duplication of effort.
That's something to optimize, but it's single commit picking,
sequencer_pick_revisions() should call single_pick() instead of
pick_commits().

The read+close on the whole checkout looks like there's problem with
refresh operation and git decides to read up and verify sha-1 by
content. Pawel, if you run "strace git update-index --refresh" twice,
does it still show 1 stat + 1 read for every entry on the second try?
-- 
Duy

Re: slow git-cherry-pick.

From: Paweł Sikora <hidden>
Date: 2016-06-15 22:59:23

On Wednesday 04 of December 2013 08:07:23 Duy Nguyen wrote:
On Wed, Dec 4, 2013 at 3:13 AM, Thomas Rast [off-list ref] wrote:
quoted
Paweł Sikora [off-list ref] writes:

Umm, there's a gem here that the thread missed so far:
quoted
my git repo isn't very big[1] but it's checked out on the linear lvm
where random i/o generally hurts and strace shows that current git
version
performs 2x{lstat}+1x{open,read,close} [2] on whole checkout before
           ^^^^^^^^^

There's no reason why it should do the lstat() *twice* for every file.
But Paweł is right; the code path roughly goes like this:

int cmd_cherry_pick(int argc, const char **argv, const char *prefix)
{
[...]

        res = sequencer_pick_revisions(&opts);

int sequencer_pick_revisions(struct replay_opts *opts)
{
[...]

        read_and_refresh_cache(opts);

[...]

        return pick_commits(todo_list, opts);

}

static int pick_commits(struct commit_list *todo_list, struct replay_opts
*opts) {
[...]

        read_and_refresh_cache(opts);

I'm too tired to dig further, but AFAICT it's just a rather obvious case
of duplication of effort.
That's something to optimize, but it's single commit picking,
sequencer_pick_revisions() should call single_pick() instead of
pick_commits().

The read+close on the whole checkout looks like there's problem with
refresh operation and git decides to read up and verify sha-1 by
content. Pawel, if you run "strace git update-index --refresh" twice,
does it still show 1 stat + 1 read for every entry on the second try?
the 'git update-index --refresh' runs quickly and strace shows only lstat()
on every file. i see no massive open/read actions in this case.

$ strace -o strace-try1.log git update-index --refresh
hmdb: needs update
$ strace -o strace-try2.log git update-index --refresh
hmdb: needs update

$ grep -c lstat strace-try1.log 
33793
$ grep -c lstat strace-try2.log
33793

-- 
gpg key fingerprint = 60B4 9886 AD53 EB3E 88BB 1EB5 C52E D01B 683B 9411
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help