On 10/3/2017 6:49 AM, Junio C Hamano wrote:
Derrick Stolee <dsto...@microsoft.com> writes:
p0008.1: find_unique_abbrev() for existing objects
--------------------------------------------------
For 10 repeated tests, each checking 100,000 known objects, we find the
following results when running in a Linux VM:
| | Pack | Packed | Loose | Base | New | |
| Repo | Files | Objects | Objects| Time | Time | Rel% |
|-------|-------|---------|--------|--------|--------|---------|
| Git | 1 | 230078 | 0 | 0.09 s | 0.06 s | - 33.3% |
| Git | 5 | 230162 | 0 | 0.11 s | 0.08 s | - 27.3% |
| Git | 4 | 154310 | 75852 | 0.09 s | 0.07 s | - 22.2% |
| Linux | 1 | 5606645 | 0 | 0.12 s | 0.32 s | +146.2% |
| Linux | 24 | 5606645 | 0 | 1.12 s | 1.12 s | - 0.9% |
| Linux | 23 | 5283204 | 323441 | 1.08 s | 1.05 s | - 2.8% |
| VSTS | 1 | 4355923 | 0 | 0.12 s | 0.23 s | + 91.7% |
| VSTS | 32 | 4355923 | 0 | 1.02 s | 1.08 s | + 5.9% |
| VSTS | 31 | 4276829 | 79094 | 2.25 s | 2.08 s | - 7.6% |
The above does not look so good, especially in cases where a
repository is well maintained by packing into smaller number of
packs, we get much worse result?
I understand that this patch on its own does not have good numbers. I
split the
patches 3 and 4 specifically to highlight two distinct changes:
Patch 3: Unroll the len loop that may inspect all files multiple times.
Patch 4: Parse less while disambiguating.
Patch 4 more than makes up for the performance hits in this patch.
p0008.2: find_unique_abbrev() for missing objects
-------------------------------------------------
For 10 repeated tests, each checking 100,000 missing objects, we find
the following results when running in a Linux VM:
| | Pack | Packed | Loose | Base | New | |
| Repo | Files | Objects | Objects| Time | Time | Rel% |
|-------|-------|---------|--------|--------|--------|--------|
| Git | 1 | 230078 | 0 | 0.66 s | 0.08 s | -87.9% |
| Git | 5 | 230162 | 0 | 0.90 s | 0.13 s | -85.6% |
| Git | 4 | 154310 | 75852 | 0.79 s | 0.10 s | -87.3% |
| Linux | 1 | 5606645 | 0 | 0.48 s | 0.32 s | -33.3% |
| Linux | 24 | 5606645 | 0 | 4.41 s | 1.09 s | -75.3% |
| Linux | 23 | 5283204 | 323441 | 4.11 s | 0.99 s | -75.9% |
| VSTS | 1 | 4355923 | 0 | 0.46 s | 0.25 s | -45.7% |
| VSTS | 32 | 4355923 | 0 | 5.40 s | 1.15 s | -78.7% |
| VSTS | 31 | 4276829 | 79094 | 5.88 s | 1.18 s | -79.9% |
The question is if this is even measuring a relevant workload. How
often would we have a full 40-hex object name for which we actually
do not have the object, and ask its name to be abbreviated?
Compared to it, the result from p0008.1 feels a lot more important.
We know we make tons of "abbreviate the object name for this object
we have" and we see them every day in our "git log -p" output.
Seeing a lot more impressive result from p0008.2 than p0008.1 makes
me unsure if this patch is optimizing for the right case.
I'll have to see the code a bit deeper before I can comment on it.
Thanks.
I agree that p0008.1 is more important. p0008.2 covers an important case
of the
previous implementation. The line
exists = has_sha1_file(sha1);
will inspect all packfiles and scan the full loose-object directory that
would contain
the object. The only reason for this call is to determine how to inspect
the result of
get_short_oid(), but is a significant portion of the time that is gained
here.
Thanks,
-Stolee
