[darcs-devel] New performance regression?

[Simon Marlow]

David Roundy wrote:
> On Fri, Feb 08, 2008 at 10:03:52AM -0500, David Roundy wrote:
>> On Fri, Feb 08, 2008 at 10:53:57AM +0000, Simon Marlow wrote:
>>> You should be able to reproduce this pretty easily with a darcs2 ghc 
>>> repository.  I get no progress messages at all from pull for quite a while, 
>>> and the 400-patch pull now takes over 100s where it previously took 10. 
>>> Unpull is also taking about 100s on this machine, where previously it took 15.
>> Oddly enough in my quick darcs2-format test, I don't see the slowdown, but
>> on the hashed-format test I do.  This may be some sort of hysteresis
>> effect.  I'll try using your darcs2 repository, but first I want to debug
>> the darcs get slowness (or at least lack of progress).  :(
> 
> Okay, I've got a data point:  darcs obliterate --last 580 -a is fast (well,
> the old speed) while darcs obliterate --from-tag ... -a is slow (where on
> my test case there are 580 patches after the tag).  So it seems to be in
> the patch selection code.

Indeed, I can confirm that.

> On the other hand, I still see now slowness in pull!  :(

It's very strange - certainly this morning I saw very long pull times in 
both directions between my two repositories, and now I'm only seeing it in 
one direction.

To summarise, with my two repos called A and B:

   oblit in either A or B slow with --from-patch, fast with --last
   pull A -> B  fast
   pull B -> A  *slow*

Ah, but I have clues.  repo A has some strange corruption.  Spaces in 
filenames have been replaced by (literally) \32.

$ \ls InstallShield
Component\32\Definitions            README               Setup\32\Files
decyg.pl                            Registry\32\Entries  Shell\32\Objects
File\32\Groups                      runexe.c             String\32\Tables
Glasgow\32\Haskell\32\Compiler.ipr  Script\32\Files      Text\32\Substitutions

this is what it should look like:

$ \ls InstallShield
Component Definitions         README            Setup Files
decyg.pl                      Registry Entries  Shell Objects
File Groups                   runexe.c          String Tables
Glasgow Haskell Compiler.ipr  Script Files      Text Substitutions

I have to go now so I can't investigate any more, but my best guess is that 
this happened when I did 'darcs repair' yesterday using a darcs2 that was 
built with bytestring support turned on.

Cheers,
	Simon