Re: Emacs repository benchmark: bzr and git

[James Westby]

On Tue, 2008-03-18 at 17:43 +0200, Teemu Likonen wrote:
> I did some benchmarking in git and bzr repositories of Emacs. Some
> numbers: 89711 revisions (by "git log --pretty=oneline | wc -l"), 2825
> files. Both repositories seem to have just linear history converted from
> CVS repo. Both have the same head revision which is
> 481c2a1e31f32c8aa0fb6d504575b75a18537788 (git) and
> revid:cvs-1:tsdh-20080318180244-lxbzttdnh6ecqbka (bzr).
> 

Hi,

Thanks for the concrete numbers that we can work from.

So, I hacked up a quick plugin to output logs in a flat style like
git. I've attached it to this mail. You can install it by dropping
it in to ~/.bazaar/plugins.

It is totally unintegrated with bzr's log framework. I plan on
rectifying that and proposing it for inclusion. It also has
some nasty UI warts like using --end revid to limit the
revisions that you want to see to ones that are not that
revid or the parents of it. This means you can't do things
like bzr flatlog -r-100..

I include the numbers from that inline with yours

> 
> Viewing history
> ---------------
> 
> 
> The complete history:
> 
> $ time git log >/dev/null
> real    0m5.741s
> 
> $ time bzr log >/dev/null
> real    3m15.708s
> 

bzr flatlog > /dev/null  45.17s user 0.75s system 97% cpu 46.883 total

> 
> Last 100 revisions:
> 
> $ time git log -100 >/dev/null
> real    0m0.011s
> 
> $ time bzr log -l100 >/dev/null
> real    2m10.270s
> 

bzr flatlog -l100 > /dev/null  12.11s user 0.32s system 97% cpu 12.714
total

> 
> Last 10 revisions:
> 
> $ time git log -10 >/dev/null
> real    0m0.007s
> 
> $ time bzr log -l10 >/dev/null
> real    2m9.163s
> 

bzr flatlog -l10 > /dev/null  12.18s user 0.25s system 98% cpu 12.582
total


I can also suggest a workaround for anyone that is using bzr and wants
to speed up log further while we work on it.

If you edit ~/.bazaar/bazaar.conf and add

[ALIASES]
flatlog = --end cvs-1:bastien1-20080217010841-op363t09ccs7pais

you can get

bzr flatlog --end cvs-1:bastien1-20080217010841-op363t09ccs7pais
> /dev/null  0.77s user 0.08s system 96% cpu 0.876 total

and still have 1000 revisions to look at.

I'll work on getting -r properly integrated which means you could
have this alias set at all times, and then if you needed full
history you could do

bzr flatlog -r1..

> 
> Creating a branch
> -----------------
> 
> With git I chose "git checkout -b" instead of "git branch" because the
> former also checks out the files as does "bzr branch". The bzr branch is
> created inside the same shared repository so that the common objects are
> shared.

bzr creates a second working tree, git replaces your first.

> 
> 
> Create new topic branch based on the head revision of the main
> development branch:
> 
> $ time git checkout -b topic master >/dev/null
> real    0m0.062s
> 
> $ time bzr branch trunk topic >/dev/null
> real    0m7.249s

to compare getting the second working tree you can use

git clone emacs temp2  1.70s user 1.29s system 31% cpu 9.395 total

(it hardlinks the objects rather than just re-using them as bzr does,
so it's still not the same)

However this is a bit silly, as you are just comparing the usual
ways to get a new branch in that particular VCS.

It is possible to emulate the git way using a couple of bzr commands
if you prefer to work in one directory with one working tree. I don't
have benchmark numbers for that currently though I'm afraid.

> 
> 
> Create new topic branch based on earlier revision of main development
> branch:
> 
> $ time git checkout -b topic master~4 >/dev/null
> real    0m0.085s
> 
> $ time bzr branch -r -5 trunk topic >/dev/null
> real    2m51.551s
> 

There was a patch proposed a couple of days to tackle an efficiency
operation in exactly this command.

> 
> 
> Compare branches' commit histories
> ----------------------------------
> 
> In above benchmark I created branch 'topic' which is based on earlier
> revision of main development branch. In this test I compared commands
> which display commits that are missing from 'topic' branch compared to
> the main development branch (four commits in total).
> 
> 
> $ time git log topic..master >/dev/null
> real    0m0.006s
> 
> $ time bzr missing --theirs-only ../trunk >/dev/null
> real    18m25.173s
> 

An inefficiency was also highlighted here a couple of days ago. I think
something was proposed that we could switch this and other commands
to that would speed them up greatly.

If you want to talk about bzr's performance or features I will be happy
to do so, but please drop the emacs list from the discussion.

Thanks,

James

flatlog.py
Description: Text Data