When "cat-file --batch-all-objects" iterates over each object, it knows
where to find each one. But when we look up details of the object, we
don't use that information at all.
This patch teaches it to use the pack/offset pair when we're iterating
over objects in a pack. This yields a measurable speed improvement
(timings on a fully packed clone of linux.git):
Benchmark #1: ./git.old cat-file --batch-all-objects --unordered --batch-check="%(objecttype) %(objectname)"
Time (mean ± σ): 8.128 s ± 0.118 s [User: 7.968 s, System: 0.156 s]
Range (min … max): 8.007 s … 8.301 s 10 runs
Benchmark #2: ./git.new cat-file --batch-all-objects --unordered --batch-check="%(objecttype) %(objectname)"
Time (mean ± σ): 4.294 s ± 0.064 s [User: 4.167 s, System: 0.125 s]
Range (min … max): 4.227 s … 4.457 s 10 runs
Summary
'./git.new cat-file --batch-all-objects --unordered --batch-check="%(objecttype) %(objectname)"' ran
1.89 ± 0.04 times faster than './git.old cat-file --batch-all-objects --unordered --batch-check="%(objecttype) %(objectname)"
The implementation is pretty simple: we just call packed_object_info()
instead of oid_object_info_extended() when we can. Most of the changes
are just plumbing the pack/offset pair through the callstack. There is
one subtlety: replace lookups are not handled by packed_object_info().
But since those are disabled for --batch-all-objects, and since we'll
only have pack info when that option is in effect, we don't have to
worry about that.
There are a few limitations to this optimization which we could address
with further work:
- I didn't bother recording when we found an object loose. Technically
this could save us doing a fruitless lookup in the pack index. But
opening and mmap-ing a loose object is so expensive in the first
place that this doesn't matter much. And if your repository is large
enough to care about per-object performance, most objects are going
to be packed anyway.
- This works only in --unordered mode. For the sorted mode, we'd have
to record the pack/offset pair as part of our oid-collection. That's
more code, plus at least 16 extra bytes of heap per object. It would
probably still be a net win in runtime, but we'd need to measure.
- For --batch, this still helps us with getting the object metadata,
but we still do a from-scratch lookup for the object contents. This
probably doesn't matter that much, because the lookup cost will be
much smaller relative to the cost of actually unpacking and printing
the objects.
For small objects, we could probably swap out read_object_file() for
using packed_object_info() with a "object_info.contentp" to get the
contents. But we'd still need to deal with streaming for larger
objects. A better path forward here is to teach the initial
oid_object_info_extended() / packed_object_info() calls to retrieve
the contents of smaller objects while they are already being
accessed. That would save the extra lookup entirely. But it's a
non-trivial feature to add to the object_info code, so I left it for
now.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
bf972896d7