Re: [PATCH 00/20] Separate `ref_cache` into a separate module

On Mon, Mar 20, 2017 at 07:24:12PM +0100, Ævar Arnfjörð Bjarmason wrote:

On Mon, Mar 20, 2017 at 5:33 PM, Michael Haggerty [off-list ref] wrote:

quoted

I've completed a draft of an epic 48-patch series implementing all of
the above points on my GitHub fork [1] as branch
`wip/mmap-packed-refs`. It dramatically speeds up performance and
reduces memory usage for some tasks in repositories with very many
packed references.

Just curious, what sort of performance numbers on mainly what sort of
operations? I'd expect the fixed cost of e.g. `git log -1 >/dev/null`
to go down, which can take as long as a second on some repos with a
huge amount of refs, but anything else?

Anything that reads refs (and that includes anything that touches an
object, because it looks at refs/replace) loads the entire packed-refs
file into memory. So I'd expect you're paying that "1 second" cost on a
lot of commands.

Even commands that traverse the entire ref namespace should improve, as
we'd spend a lot less time allocating and copying data into the
in-memory refs cache. Not to mention that the memory usage is reduced,
though that probably only really matters for pathological cases.

-Peff