patch, 3 partsImprove abbreviation disambiguation

Hello,

My name is Derrick Stolee and I just switched teams at Microsoft from the VSTS Git Server to work on performance improvements in core Git.

This is my first patch submission, and I look forward to your feedback.

Thanks,
 Stolee

When displaying object ids, we frequently want to see an abbreviation for easier typing. That abbreviation must be unambiguous among all object ids.

The current implementation of find_unique_abbrev() performs a loop checking if each abbreviation length is unambiguous until finding one that works. This causes multiple round-trips to the disk when starting with the default abbreviation length (usually 7) but needing up to 12 characters for an unambiguous short-sha. For very large repos, this effect is pronounced and causes issues with several commands, from obvious consumers `status` and `log` to less obvious commands such as `fetch` and `push`.

This patch improves performance by iterating over objects matching the short abbreviation only once, inspecting each object id, and reporting the minimum length of an unambiguous abbreviation.

A performance helper `test-abbrev` and performance test `p0008-abbrev.sh` are added to demonstrate this performance improvement. Here are some performance results for the three included commits, using GIT_PERF_REPEAT_COUNT=10 since the first test is frequently an outlier due to the file cache being cold.

Running git on a Linux VM, we see the following gains.

| Repo    | Pack-Files | Loose Objs | Baseline | Patch 2 | Patch 3 |
|---------|------------|------------|----------|---------|---------|
| Git.git | 1          | 0          | 0.46 s   | -87.0%  | -87.0%  |
| Git.git | 5          | 0          | 1.04 s   | -84.6%  | -85.6%  |
| Git.git | 4          | 75852      | 0.88 s   | -86.4%  | -86.4%  |
| Linux   | 1          | 0          | 0.63 s   | -38.1%  | -69.8%  |
| Linux   | 24         | 0          | 5.41 s   | -69.3%  | -71.5%  |
| Linux   | 23         | 323441     | 5.41 s   | -70.6%  | -73.4%  |

Running a similar patch on Git for Windows, we see the following gains.

| Repo          | Pack-Files | Loose | Baseline | Patch 2 | Patch 3 |
|---------------|------------|-------|----------|---------|---------|
| GitForWindows | 6          | 319   | 7.19 s   | -91.1%  | -91.5%  |
| VSTS          | 3          | 38    | 7.83 s   | -88.9%  | -90.9%  |
| Linux         | 3          | 0     | 7.92 s   | -87.9%  | -90.2%  |
| Windows       | 50         | 219   | 17.8 s   | -98.6%  | -98.6%  |

Note that performance improves in all cases, but the performance gain is larger when there are multiple, large pack-files. This gain comes from the lack of in-memory caching of index files that have already been inspected.

Derrick Stolee (3):
  sha1_name: Create perf test for find_unique_abbrev()
  sha1_name: Unroll len loop in find_unique_abbrev_r
  sha1_name: Parse less while finding common prefix

 Makefile               |  1 +
 sha1_name.c            | 66 ++++++++++++++++++++++++++++++++++++++------------
 t/helper/.gitignore    |  1 +
 t/helper/test-abbrev.c | 22 +++++++++++++++++
 t/perf/p0008-abbrev.sh | 12 +++++++++
 5 files changed, 87 insertions(+), 15 deletions(-)
 create mode 100644 t/helper/test-abbrev.c
 create mode 100755 t/perf/p0008-abbrev.sh

Unroll the while loop inside find_unique_abbrev_r to avoid iterating through all loose objects and packfiles multiple times when the short name is longer than the predicted length.

Instead, inspect each object that collides with the estimated abbreviation to find the longest common prefix.

Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
---
 sha1_name.c | 57 ++++++++++++++++++++++++++++++++++++++++++---------------
 1 file changed, 42 insertions(+), 15 deletions(-)

Create helper program test-abbrev to compute the minimum length of a disambiguating short-sha for 100,000 object ids. The ids are created by iterating an unsigned int hash_base by a constant hash_delta and copying hash_base five times across the sha1. Iterating by hash_delta does not create a duplicate value for over 10,000,000 iterations.

test-abberv demonstrates the performance improvements that will be shown by later improvements to the find_unique_abberv(). The value of 100,000 is large enough to show the significance of the later improvements while only taking a few seconds on large repos.

Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
---
 Makefile               |  1 +
 t/helper/.gitignore    |  1 +
 t/helper/test-abbrev.c | 23 +++++++++++++++++++++++
 t/perf/p0008-abbrev.sh | 12 ++++++++++++
 4 files changed, 37 insertions(+)
 create mode 100644 t/helper/test-abbrev.c
 create mode 100755 t/perf/p0008-abbrev.sh

Derrick Stolee <dstolee@microsoft.com> writes:

As far as I recall, we do not (yet) allow declaration after statement in our codebase. Move this down to make it after all decls.

It probably is technically OK to assume sizeof(int) always equals to sizeof(unsigned), but because you use 'n' _only_ to work with uint and never with int, it would make more sense to match.

But I do not think we want this "clever" optimization that involves 'n' in the first place.

Does it make sense to assume that uint is always 4-byte (so this code won't work if it is 8-byte on your platform) and doing this is faster than using platform-optimized memcpy()?

I also wonder if this is measuring the right thing. I am guessing that by making many queries for a unique abbreviation of "random" (well, it is deterministic, but my point is these queries are not based on the names of objects that exist in the repository) hashes, this test measures how much time it takes for us to decide that such a random hash does not exist. In the real life use, we make the opposite query far more frequently: we have an object that we _know_ exists in the repository and we try to find a sufficient length to disambiguate it from others, and I suspect that these two use different logic. Don't you need to be measuring the time it takes to compute the shortest abbreviation of an object that exists instead?

On 9/17/2017 5:51 PM, Junio C Hamano wrote:

Will fix.

Will fix. I also notice that "n" should be const.

 >>> +	while (count++ < 100000) {

I'm not sure what you mean by using memcpy to improve this, because it would require calling memcpy in the inner loop, such as

	for (i = 0; i < n; i++)
		memcpy(oid.hash + i * sizeof(unsigned), &hash_base,
		       sizeof(unsigned));

I'm probably misunderstanding your intended use of memcpy().

First, this doesn't just measure the time it takes to determine non- existence, because it finds the len required to disambiguate an "incoming" hash from all known hashes. When testing, I put in a simple printf to report the result abbreviation so I could see how often it needed to be extended. In this sense, the test exposes the while loop that is removed by PATCH 2/3.

Second, your criticism about extant hashes is valid, and one I struggled to reconcile. I see two issues with testing known hashes:

1. By determining the hash exists, we have inspected the file that
    contains it (either a .idx or the loose-object directory). This
    has side-effects that warm up the file cache so the looped method
    is artificially faster to find matching objects. The effect is
    particularly significant on a repo with exactly one packfile.

2. If we iterate over the entire set of objects, this test takes
    O(N*t(N)) time, where t(N) is the average time to compute a
    minimum abbreviation. For large repos, this can take several
    minutes. Instead, even with the constant 100,000 test hashes, we
    have an O(t(N)) test. We could avoid the asymptotic growth by
    limiting the number of existing hashes we use, but how do we
    find a sufficiently uniform sample from them?

By looking at some other perf tests, I see that we can add a pre- requisite action. I will investigate making another helper that uniformly selects a set of hashes from the repo and writes them to stdout in a random order. p0008-abbrev.sh will run the helper as a prerequisite, saving the data to a file. test-abbrev will read the hashes from stdin to test find_unique_abbrev. This should avoid the side-effects I mentioned (test-abbrev will not warm up indexes) while also testing abbreviation lengths for existing objects.

I'll get started on these changes and send a new patch with new perf data in a couple days. Please let me know if there is a better way to measure performance for this change.

Thanks, -Stolee

Derrick Stolee <stolee@gmail.com> writes:

Sorry, I left it without saying as I thought it was obvious, but what I meant was to use a whole "struct oid", not just a single unsigned (repeated), as the hash that is tested. If you have an array of object names you use in the test, then

	for (count = 0; count < limit; count++) {
		hashcpy(&oid.hash, &samples[count]);

		... do the probing ...
	}

Sorry, my phrasing was indeed misleading. I know the time we spend to see if we have or do not have the object is the largest cycle spender in these codepaths (and even if it were, I do not think that is what you are trying to optimize in these patches anyway).

But if I recall correctly, the way we come up with the unique abbreviation for an object that exists and an object that does not exist are different? And because most of the time (think: "git log -p" output) we would be finding abbreviation for objects that we do have, benchmarking and tweaking the code that comes up with an object that does not exist is not optimizing for the right case.

Back when I wrote that initial response, I didn't check how different the code was between the two cases, but now I did. It seems that in both cases we start from the shortest-allowed length and then extend the same way, and the only difference between these two cases is that we return immediately when our candidate name is long enough not to match any existing object when the full name refers to an object we do not have, while we return only when disambiguity is resolved. I _think_ these amount to the same computation (i.e. when an object with the full name we have exists, the amount of computation we need to come up with its unique abbreviation is the same as the computation we need to find the unique abbreviation for the same name in another repository that has identical set of objects, minus that single object), so from that point of view, throwing random hashes, most of which would not name any existing object, and measuring how much time it takes to run get_short_oid() to compute the minimum length of the unique prefix should be sufficient.

Thanks.

Create get_hex_char_from_oid() to parse oids one hex character at a time. This prevents unnecessary copying of hex characters in extend_abbrev_len() when finding the length of a common prefix.

This change decreases the time to run test-abbrev by up to 40% on large repos.

Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
---
 sha1_name.c | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

Hi,

Derrick Stolee wrote:

Thanks for writing this.  Looks exciting.

[...]

[etc]

Can this (especially the performance information) go in the commit message for one of the patches?

Otherwise it is harder for people to track down when looking at these changes later with "git log". In the worst case, if the mailing list archive is down, then people would not have access to this information at all. For that reason, I try (though I often fail!) to restrict cover letters to giving hints at e.g. what changed since the patch series last visited the list, and to make the commit messages in the patches themselves as self-contained as possible.

That aside, looking forward to reading the patches themselves in more detail. Thanks for working on this.

Sincerely, Jonathan