'git status' on NFS performance regression in 1.7.0

'git status' on NFS performance regression in 1.7.0

[James Pickens]

Hi,

I noticed that 'git status' in version 1.7.0 is much slower than in 1.6.2.5
on large work trees on NFS - averaging ~13 seconds runtime vs. ~2 seconds.
I did a bit of debugging and found that 'git status' apparently doesn't use
the multi-threaded preload_index any more, although some other commands
like diff still use it.  Was it intentionally dropped from 'git status'?

James

Re: 'git status' on NFS performance regression in 1.7.0

[Junio C Hamano]

James Pickens <jepicken@gmail.com> writes:

> I noticed that 'git status' in version 1.7.0 is much slower than in 1.6.2.5
> on large work trees on NFS - averaging ~13 seconds runtime vs. ~2 seconds.
> I did a bit of debugging and found that 'git status' apparently doesn't use
> the multi-threaded preload_index any more, although some other commands
> like diff still use it.  Was it intentionally dropped from 'git status'?

There might be subtle breakage for doing this, but it would be worth a try
;-)

 builtin-commit.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/builtin-commit.c b/builtin-commit.c
index 55676fd..71f81c9 100644
--- a/builtin-commit.c
+++ b/builtin-commit.c
@@ -1046,7 +1046,7 @@ int cmd_status(int argc, const char **argv, const char *prefix)
 	if (*argv)
 		s.pathspec = get_pathspec(prefix, argv);
 
-	read_cache();
+	read_cache_preload();
 	refresh_index(&the_index, REFRESH_QUIET|REFRESH_UNMERGED, s.pathspec, NULL, NULL);
 	s.is_initial = get_sha1(s.reference, sha1) ? 1 : 0;
 	s.in_merge = in_merge;

Re: 'git status' on NFS performance regression in 1.7.0

[Junio C Hamano]

James Pickens <jepicken@gmail.com> writes:

> I noticed that 'git status' in version 1.7.0 is much slower than in 1.6.2.5
> on large work trees on NFS - averaging ~13 seconds runtime vs. ~2 seconds.
> I did a bit of debugging and found that 'git status' apparently doesn't use
> the multi-threaded preload_index any more, although some other commands
> like diff still use it.  Was it intentionally dropped from 'git status'?

There might be subtle breakage for doing this, but it would be worth a try
;-)

 builtin-commit.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/builtin-commit.c b/builtin-commit.c
index 55676fd..71f81c9 100644
--- a/builtin-commit.c
+++ b/builtin-commit.c
@@ -1046,7 +1046,7 @@ int cmd_status(int argc, const char **argv, const char *prefix)
 	if (*argv)
 		s.pathspec = get_pathspec(prefix, argv);
 
-	read_cache();
+	read_cache_preload(NULL);
 	refresh_index(&the_index, REFRESH_QUIET|REFRESH_UNMERGED, s.pathspec, NULL, NULL);
 	s.is_initial = get_sha1(s.reference, sha1) ? 1 : 0;
 	s.in_merge = in_merge;

Re: 'git status' on NFS performance regression in 1.7.0

[James Pickens]

On Wed, Feb 17, 2010, Junio C Hamano <gitster@pobox.com> wrote:
> There might be subtle breakage for doing this, but it would be worth a try
> ;-)

Thanks, with that patch Git 1.7.0 is faster than 1.6.2.5 - ~2 seconds
average runtime vs. ~3 seconds (1.6.2.5 got slower since my original test
last night; must be more network and/or file server traffic right now).

I'm not sure how to interpret the "subtle breakage" comment with the
winking smiley.  Do you mean that preload_index in 1.7.0 is not well tested
and may be broken?  FWIW, I didn't notice any breakage, but I didn't do
much testing.

James

Re: 'git status' on NFS performance regression in 1.7.0

[Junio C Hamano]

James Pickens <jepicken@gmail.com> writes:

> I'm not sure how to interpret the "subtle breakage" comment with the
> winking smiley.  Do you mean that preload_index in 1.7.0 is not well tested
> and may be broken?  FWIW, I didn't notice any breakage, but I didn't do
> much testing.

The new "status" codepath is different from "commit --dry-run" codepath
that was used by "git status" in 1.6.6 series.  This old codepath has been
used extensibly with preloaded index and is continued to be used when you
run "git commit" with various options.  It is not preload-index that could
be subtly broken.

However, nobody used the new "status" codepath with preloaded index, and I
haven't thought things through if anything we are doing in that codepath
is incompatible with preloaded index in some way.

By the way, the argument to read_cache_preload() should be s.pathspec, not
NULL, I think.

Thanks.

Re: 'git status' on NFS performance regression in 1.7.0

[Peter Krefting]

James Pickens:

> I noticed that 'git status' in version 1.7.0 is much slower than in 
> 1.6.2.5 on large work trees on NFS - averaging ~13 seconds runtime vs. ~2 
> seconds.

This applies to local disk as well. Where it used to be almost 
instantaneous, I now see a wait of about three seconds on a checkout of 
about 8,000 files.

-- 
\\// Peter - http://www.softwolves.pp.se/