[darcs-devel] performance regression

[David Roundy]

On Sat, Dec 04, 2004 at 02:53:35PM +0000, Ian Lynagh wrote:
> On Sat, Dec 04, 2004 at 08:48:04AM -0500, David Roundy wrote:
> > It seems that there is a significant regression in darcs' speed at running
> > check:
> > ...
> Odd, I don't see a noticable difference between 1.0.0, darcs-unstable
> and my tree doing the above on a darcs repository.

Hmmm.  That's odd.  I consistently get current darcs taking 50% longer
(measured user CPU time) than darcs 1.0.0 here (freshly recompiled, so it's
not a compiler difference or anything like that).

I wonder what the difference between your system and mine is? Mine is
basically a debian sarge system.

> > I'm in the process of tracking this down (using trackdown!), but I have a
> > feeling that Ian can figure out what's causing the problem more quickly
> > than I can.
> 
> If you compile them both with profiling and run with "+RTS -p -RTS" on
> the command line then the problem may jump out. Remember to rename
> darcs.prof after the first one or the second will overwrite it.

The problem is that that takes almost an hour on my computer (recompiling
two versions of darcs), and I'm impatient! :( I'm starting now compiling a
new and old version with profiling.

> Perhaps a timesuite (like the testsuite, doubling as a memorysuite)
> would be a useful thing for someone to create? It could compare against
> a file giving previous output or another darcs binary and warn of >5% or
> 10% increases (and also notify you of significant decreases).

That is a very good idea.  It would be a bit tricky, since you'd have to
generate the repositories on which to run the tests, which probably
couldn't be small, or you'd not get good timings.  But it would be a
great tool to have at our disposal.
-- 
David Roundy
http://www.darcs.net