RFC: startup time - again

[Russel Winder]

On Fri, 2008-09-12 at 20:02 +1000, Robert Collins wrote:
> On Fri, 2008-09-12 at 09:45 +0100, Russel Winder wrote:
> 
> > The so what about things not being stable is that to focus on a supposed
> > 5% improvement when the variance is 10% is a waste of everybody's time.
> 
> Actually, its not. I made a very large improvement in commit performance
> by a series of improvements, nearly every one of which was under the
> variance of any set of runs.
> 
> Performance depends on two basic things:
>  - doing as little work as possible
>  - not waiting when work could be done
> 
> (the later point applies to networking, threading, etc)

Threading in Python is a bit of a disaster in these multicore days -- I
am switching to using the processing package (renamed multiprocessing
and standard in 2.6) or Parallel Python.  However that is a very
different thread!

> If you run a workload ten times, once your caches are primed, each run
> does the same work. So the lowest time indicates when you spent the
> least time waiting when work could be done. Thus a small improvement
> within the variance of the runs is still probably representative of less
> work being done.

I fully agree that you can undertake small percentage (less than the
variance) improvement work and still progress.  On this I have no
quibble -- I accept that getting 10 instance of a 2% improvement is real
progress.  What worries me is that 5% of 0.5s is a minor issue compared
to trimming 5% of the 90s it takes to do many of the things that Bazaar
is being asked to do on an hour by hour basis for real programmers using
real repositories.  It is a question of priorities.

Whether it is liked or not, performance is the issue of the moment, Git
has it, Mercurial sort of has it, and Bazaar has some problems -- start
up time being 0.2s rather than 0.4s is the least of the worries.

I prefer the model of Bazaar to the model of Git (and Mercurial) so can
forgive a lot, but there are breaking points.  I am hoping that the work
on indexing and repository format can bring in the 25%-50% performance
improvement overall that will establish Bazaar as a player.  Currently
Git (especially) and Mercurial have the zeitgeist.  

However Alexander has declared my contributions useless so I guess I
should just shut up.

-- 
Russel.
====================================================
Dr Russel Winder                 Partner

Concertant LLP                   t: +44 20 7585 2200, +44 20 7193 9203
41 Buckmaster Road,              f: +44 8700 516 084
London SW11 1EN, UK.             m: +44 7770 465 077
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: This is a digitally signed message part
Url : https://lists.ubuntu.com/archives/bazaar/attachments/20080913/ee8164ff/attachment.pgp