Bazaar performance

[Ian Clatworthy]

Guido Ostkamp wrote:
> On Thu, 15 May 2008, Matt Nordhoff wrote:
>> Did you try bzr over bzr+ssh? (You need bzr installed on the remote
>> side for it to work.)
>>
>> I doubt bzr will win, but with your current test it's at a
>> disadvantage: git and hg are using smart servers, but bzr is using
>> dumb sftp.

> Now I tried bzr+ssh once, but it turned out to use over 11 minutes to
> clone the repo. However, looking at the 'top' output which did not show
> full activity on both nodes, I am not sure if something went wrong with
> this method.
> 
> Can you or somebody else here comment?

For Bazaar 1.0 to 1.5, the protocol to use to get the fastest *initial*
clone time in Bazaar is http according to my measurements. The smart
server (bzr://) was a big improvement over http until the default format
was changed to packs in 1.0, but we're yet to optimise it for packs, and
some optimisations for earlier formats actually slow pack transfers down
(vs http).

We're working on this issue as a top priority for the next release or
two. We wanted to land a new version of the protocol (v3) first. Andrew
merged that new protocol into bzr.dev a few days ago so it's all systems
go on tuning network performance now. As soon as I land the end-of-line
conversion stuff, I'll be helping with this as well.

Ian C.