[PATCH 00 of 10] RFC: Light-Weight Copy

[Dirkjan Ochtman]

On Thu, Sep 9, 2010 at 10:35, Sune Foldager <cryo at cyanite.org> wrote:
>> Do we have any benchmarks for this?
>
> For the above paragraph or for the patch queue as a whole? And what
> benchmarks, speed, space..?
>
> For very small files, the repository size actually increases by using
> lwcopy, but, coming to think of it, it would be easy to store such entries
> without lwcopy to avoid this slight regression; it mainly affects the test
> suite, I think.

I'd assume the simple model mostly works (i.e. for copies, we now need
space which is negligible compared to the file size for all but the
smallest of files), so those benchmarks likely wouldn't be that
interesting.

> As for performance in general, there are a few more indirections in the
> general revlog code path, and a bit more still for filelog, although for
> non-copy entries it amounts to about 2-3 indirect calls. For copy entries
> there is some repacking when receiving changegroups, and some partial
> repacking when adding local changesets. Among other things this can require
> twice the memory for file data storage. I believe there is room for slight
> improvement in this area.

It would be interesting to have some hard numbers on the performance
of clone --pull, maybe, or other revlog-intensive actions.

Cheers,

Dirkjan