Call for testing: generaldelta

[Matt Mackall]

On Mon, 2011-08-01 at 20:48 +0200, Isaac Jurado wrote:
> On Mon, Aug 1, 2011 at 7:02 PM, Matt Mackall <mpm at selenic.com> wrote:
> > On Sun, 2011-07-31 at 13:51 +0200, Isaac Jurado wrote:
> >> I tried with mozilla-central repository.  The Mercurial version I'm
> >> using is 1.9-1 from Debian Sid (amd64).
> >>
> >>     http://hg.mozilla.org/mozilla-central/
> >
> > Is this repo already using shrink-revlog?
> 
> I don't know.  It was a sort of "fresh" clone.  I did the following:
> 
> $ wget http://ftp.mozilla.org/pub/mozilla.org/firefox/bundles/mozilla-central.hg
> $ hg init mc
> $ cd mc
> $ hg unbundle ../mozilla-central.hg
> 
> Then a pull to bring 8 more changesets (that was two days ago).  I could
> not directly clone from the bundle because my /tmp is only 512MB in
> size.
> 
> By the way, I noticed that "hg unbundle" was much faster than "hg pull"
> from the bundle file.

Absolutely.

The "hg pull" approach treats the bundle as a "bundlerepo" - your
current (empty) repository + an uncompressed temporary copy of the
bundle and then does a normal pull.

The "hg unbundle" approach simply unpacks the bundle stream onto your
empty repo without unpacking the whole thing first.

> >> Time: real 996.090 secs (user 960.160+0.000 sys 16.160+0.000)
> >
> > Ouch. This shouldn't be that slow. But I guess it's still reordering.
> > Can you see if a gdelta clone of mc-gdelta-gdelta is fast again?
> 
> Seems even worse.
> 
> $ hg clone --time -U --pull --config format.generaldelta=1
> mc-gdelta-gdelta mc-gdelta3
> requesting all changes
> adding changesets
> adding manifests
> adding file changes
> added 73553 changesets with 337061 changes to 77253 files (+7 heads)
> Time: real 1086.660 secs (user 1002.260+0.000 sys 22.570+0.000)
> 
> I hope is not a buffer/page cache issue.  Somebody else should repeat
> the experiment in a different machine.

Doesn't look like it. You're spending all your time burning CPU in
userspace. If you were out of memory, you'd see total time much greater
than user time and lots more time in sys.

-- 
Mathematics is the supreme nostalgia of our time.