patchpack-bitmap: remove trace2 region from hot path

From: Derrick Stolee <derrickstolee@github.com>

The trace2 region around the call to lazy_bitmap_for_commit() in bitmap_for_commit() was added in 28cd730680d (pack-bitmap: prepare to read lookup table extension, 2022-08-14). While adding trace2 regions is typically helpful for tracking performance, this method is called possibly thousands of times as a commit walk explores commit history looking for a matching bitmap. When trace2 output is enabled, this region is emitted many times and performance is throttled by that output.

For now, remove these regions entirely.

This is a critical path, and it would be valuable to measure that the time spent in bitmap_for_commit() does not increase when using the commit lookup table. The best way to do that would be to use a mechanism that sums the time spent in a region and reports a single value at the end of the process. This technique was introduced but not merged by [1] so maybe this example presents some justification to revisit that approach.

[1] https://lore.kernel.org/git/pull.1099.v2.git.1640720202.gitgitgadget@gmail.com/

Signed-off-by: Derrick Stolee <derrickstolee@github.com>
---
    [For 2.38.0] pack-bitmap: remove trace2 region from hot path
    
    I noticed this while trying to backport the Abhradeep's lookup table
    work into GitHub's fork. Something went wrong in that process, causing
    this region to significantly slow down. It turns out that slow down does
    not reproduce on current 'master', which is good. I must have missed
    something while I was backporting.
    
    Regardless, the use of trace2_region_enter() here should be removed or
    replaced. For the sake of 2.38.0, this simple removal is safe enough.
    However, to really dig into what was happening I had to construct a
    rebase [2] of Jeff's trace2 stopwatch work, then apply changes on top
    [3] that could replace this region with trace2_timer_*() methods.
    
    [2]
    https://github.com/git/git/compare/master...derrickstolee:trace2-stopwatch
    [3]
    https://github.com/derrickstolee/git/compare/trace2-stopwatch...bitmap-trace2
    
    As a separate discussion, it might be worth revisiting that stopwatch
    work so we have it available as a tool when doing these kinds of
    investigations.
    
    Thanks, -Stolee

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1365%2Fderrickstolee%2Fbitmap-remove-trace2-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1365/derrickstolee/bitmap-remove-trace2-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/1365

 pack-bitmap.c | 2 --
 1 file changed, 2 deletions(-)

"Derrick Stolee via GitGitGadget" <gitgitgadget@gmail.com> writes:

As "git blame" is bad in finding a piece of code that no longer exists, it may make sense to leave a comment around here why we do not have a trace2 region around this call, perhaps?

+	/* do not add trace2_region around here in the hot path */
	bitmap = lazy_bitmap_for_commit(...);

or something?

Thanks.

Junio C Hamano <gitster@pobox.com> writes:

Also, we'd need to disable t5310.150 that expects to see reading_lookup_table label in the trace.

"Derrick Stolee via GitGitGadget" <gitgitgadget@gmail.com> writes:

Well, not doing this removal is even safer, if we know that it is not hurting in the -rc code. It would be better than breaking our tests without knowing ;-) I am not upset that the patch broke the test, by the way. Compared to things silently breaking without surfacing, a loud breakage in tests is a much easier thing to handle.

My test run with the attached just finished, so that's what I am going to queue tentatively on top.

Thanks.

 pack-bitmap.c           | 1 +
 t/t5310-pack-bitmaps.sh | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

On 9/23/2022 1:36 PM, Junio C Hamano wrote:

Whoops! I'm very sorry for not noticing that there was a test checking for this trace. Your modification makes sense and I'll try to do a more careful replacement of that test, if possible.

Thanks, -Stolee

On Fri, Sep 23, 2022 at 10:36:44AM -0700, Junio C Hamano wrote:

Yeah, I would advocate for dropping this test entirely. That trace2 marker is the only indication that is so obvious that we're using the lookup table to read bitmaps.

Perhaps the pack-bitmap helper could be taught to dump which extensions it sees, and then write a test based on that? Although that isn't really testing anything new, either, since we sometimes *notice* the extension but don't actually read anything based on it (e.g., when disabled).

So I'd be content just dropping this.

Thanks, Taylor

From: Derrick Stolee <derrickstolee@github.com>

The trace2 region around the call to lazy_bitmap_for_commit() in bitmap_for_commit() was added in 28cd730680d (pack-bitmap: prepare to read lookup table extension, 2022-08-14). While adding trace2 regions is typically helpful for tracking performance, this method is called possibly thousands of times as a commit walk explores commit history looking for a matching bitmap. When trace2 output is enabled, this region is emitted many times and performance is throttled by that output.

For now, remove these regions entirely.

This is a critical path, and it would be valuable to measure that the time spent in bitmap_for_commit() does not increase when using the commit lookup table. The best way to do that would be to use a mechanism that sums the time spent in a region and reports a single value at the end of the process. This technique was introduced but not merged by [1] so maybe this example presents some justification to revisit that approach.

[1] https://lore.kernel.org/git/pull.1099.v2.git.1640720202.gitgitgadget@gmail.com/

To help with the 'git blame' output in this region, add a comment that warns against adding a trace2 region. Delete a test from t5310 that used that trace output to check that this lookup optimization was activated. To create this kind of test again in the future, the stopwatch traces mentioned earlier could be used as a signal that we activated this code path.

Helpedy-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: Derrick Stolee <derrickstolee@github.com>
---
    [For 2.38.0] pack-bitmap: remove trace2 region from hot path
    
    I noticed this while trying to backport the Abhradeep's lookup table
    work into GitHub's fork. Something went wrong in that process, causing
    this region to significantly slow down. It turns out that slow down does
    not reproduce on current 'master', which is good. I must have missed
    something while I was backporting.
    
    Regardless, the use of trace2_region_enter() here should be removed or
    replaced. For the sake of 2.38.0, this simple removal is safe enough.
    However, to really dig into what was happening I had to construct a
    rebase [2] of Jeff's trace2 stopwatch work, then apply changes on top
    [3] that could replace this region with trace2_timer_*() methods.
    
    [2]
    https://github.com/git/git/compare/master...derrickstolee:trace2-stopwatch
    [3]
    https://github.com/derrickstolee/git/compare/trace2-stopwatch...bitmap-trace2
    
    As a separate discussion, it might be worth revisiting that stopwatch
    work so we have it available as a tool when doing these kinds of
    investigations.
    
    
    Updates in v2
    =============
    
     * I removed the test that was checking for trace2 regions. In [3], I
       played around with using trace2_counter() to replace that check, so
       we could reinstate the test without the performance drop from using
       repeated trace2_region() calls.
     * I squashed in Junio's comment change.
    
    Thanks, -Stolee

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1365%2Fderrickstolee%2Fbitmap-remove-trace2-v2
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1365/derrickstolee/bitmap-remove-trace2-v2
Pull-Request: https://github.com/gitgitgadget/git/pull/1365

Range-diff vs v1:

 1:  987827eaf90 ! 1:  8ef63c9e32f pack-bitmap: remove trace2 region from hot path
     @@ Commit message
      
          [1] https://lore.kernel.org/git/pull.1099.v2.git.1640720202.gitgitgadget@gmail.com/
      
     +    To help with the 'git blame' output in this region, add a comment that
     +    warns against adding a trace2 region. Delete a test from t5310 that used
     +    that trace output to check that this lookup optimization was activated.
     +    To create this kind of test again in the future, the stopwatch traces
     +    mentioned earlier could be used as a signal that we activated this code
     +    path.
     +
     +    Helpedy-by: Junio C Hamano <gitster@pobox.com>
          Signed-off-by: Derrick Stolee <derrickstolee@github.com>
      
       ## pack-bitmap.c ##
     @@ pack-bitmap.c: struct ewah_bitmap *bitmap_for_commit(struct bitmap_index *bitmap
       			return NULL;
       
      -		trace2_region_enter("pack-bitmap", "reading_lookup_table", the_repository);
     ++		/* this is a fairly hot codepath - no trace2_region please */
       		/* NEEDSWORK: cache misses aren't recorded */
       		bitmap = lazy_bitmap_for_commit(bitmap_git, commit);
      -		trace2_region_leave("pack-bitmap", "reading_lookup_table", the_repository);
       		if (!bitmap)
       			return NULL;
       		return lookup_stored_bitmap(bitmap);
     +
     + ## t/t5310-pack-bitmaps.sh ##
     +@@ t/t5310-pack-bitmaps.sh: test_expect_success 'verify writing bitmap lookup table when enabled' '
     + 	grep "\"label\":\"writing_lookup_table\"" trace2
     + '
     + 
     +-test_expect_success 'lookup table is actually used to traverse objects' '
     +-	git repack -adb &&
     +-	GIT_TRACE2_EVENT="$(pwd)/trace3" \
     +-		git rev-list --use-bitmap-index --count --all &&
     +-	grep "\"label\":\"reading_lookup_table\"" trace3
     +-'
     +-
     + test_expect_success 'truncated bitmap fails gracefully (lookup table)' '
     + 	test_config pack.writebitmaphashcache false &&
     + 	git repack -adb &&

 pack-bitmap.c           | 3 +--
 t/t5310-pack-bitmaps.sh | 7 -------
 2 files changed, 1 insertion(+), 9 deletions(-)

On Mon, Sep 26 2022, Derrick Stolee via GitGitGadget wrote:

Just getting rid of this seems like a good thing for now.

But aside: Yes, one way to mitigate this rather than removing the tracing would be to make it really fast.

But just skimming pack-bitmap.c do we really need trace2 at the
granularity of a single commit? Looking at who calls bitmap_for_commit()
wouldn't something like this sketch-out be much more useful?:
	
	diff --git a/pack-bitmap.c b/pack-bitmap.c
	index 9d5205055a5..439aec220c7 100644
	--- a/pack-bitmap.c
	+++ b/pack-bitmap.c
	@@ -830,10 +830,8 @@ struct ewah_bitmap *bitmap_for_commit(struct bitmap_index *bitmap_git,
	 		if (!bitmap_git->table_lookup)
	 			return NULL;
	 
	-		trace2_region_enter("pack-bitmap", "reading_lookup_table", the_repository);
	 		/* NEEDSWORK: cache misses aren't recorded */
	 		bitmap = lazy_bitmap_for_commit(bitmap_git, commit);
	-		trace2_region_leave("pack-bitmap", "reading_lookup_table", the_repository);
	 		if (!bitmap)
	 			return NULL;
	 		return lookup_stored_bitmap(bitmap);
	@@ -1042,6 +1040,7 @@ static struct bitmap *find_objects(struct bitmap_index *bitmap_git,
	 	 * The ones without bitmaps in the index will be stored in the
	 	 * `not_mapped_list` for further processing.
	 	 */
	+	/* begin trace2 find roots? */
	 	while (roots) {
	 		struct object *object = roots->item;
	 		roots = roots->next;
	@@ -1055,6 +1054,7 @@ static struct bitmap *find_objects(struct bitmap_index *bitmap_git,
	 		object_list_insert(object, &not_mapped);
	 	}
	 
	+	/* end trace2 find roots? */
	 	/*
	 	 * Best case scenario: We found bitmaps for all the roots,
	 	 * so the resulting `or` bitmap has the full reachability analysis
	@@ -1100,7 +1100,7 @@ static struct bitmap *find_objects(struct bitmap_index *bitmap_git,
	 		incdata.base = base;
	 		incdata.seen = seen;
	 
	-		revs->include_check = should_include;
	+		revs->include_check = should_include; // Will call bitmap_for_commit()
	 		revs->include_check_obj = should_include_obj;
	 		revs->include_check_data = &incdata;
	 
	@@ -1110,9 +1110,11 @@ static struct bitmap *find_objects(struct bitmap_index *bitmap_git,
	 		show_data.bitmap_git = bitmap_git;
	 		show_data.base = base;
	 
	+		/* begin trace2 rev list? */
	 		traverse_commit_list(revs,
	 				     show_commit, show_object,
	 				     &show_data);
	+		/* end trace2 rev list? */
	 
	 		revs->include_check = NULL;
	 		revs->include_check_obj = NULL;
	

This is *not* the same as stricking the tracing into bitmap_for_commit(), but do we really need the tracing that far inside our loop?

On 9/26/2022 11:01 AM, Ævar Arnfjörð Bjarmason wrote:

The point of it being where it was to check that we hit the path custom to the commit lookup extension, hence the test that is removed.

You are proposing a different region entirely, meant to measure the cost of the entire walk. Maybe that's valuable, but not critical to this topic.

If you're proposing the high-level traces as an alternative to my prototype using the trace2 timers, then I'll point out that the timer approach allows us to determine how much time is being spent checking for bitmaps versus walking commits, which the regions you provided don't.

Thanks, -Stolee

On 9/23/22 9:00 AM, Derrick Stolee via GitGitGadget wrote:

I'll dust off my "timers and counters" series, address the various issues that were raised back in December, and resubmit.

Thanks for the reminder. Jeff