[Aptitude-devel] Experimental package and package_pool implementation

[Daniel Burrows]

On Thu, Jul 15, 2010 at 09:19:48PM +0200, Piotr Galiszewski <piotr at galiszewski.pl> was heard to say:
> 2010/7/15 Daniel Burrows <dburrows at google.com>:
> > On Thu, Jul 15, 2010 at 10:05 AM, Piotr Galiszewski
> > <piotr at galiszewski.pl> wrote:
> >> Today I have rebased my later patches to current code. Unfortunately
> >> there is one big problem with package_pool. Creating package list
> >> takes nearly 17 seconds on my machine. The problem is with
> >> boost::multi_index_container. I have changed this to std::vector and
> >> time is 0.3s now.
> >
> >  Interesting, I'd like to know why that is so we can avoid hitting it again.
> > Glad it's fixed, anyway.
> >
> 
> I have no idea. Maybe it was caused by not used hashed_unique?

  So, I saw you branched the broken code, but I'm not sure how to test it?
(--qt does nothing)

  Anyway, my two theories are:

    1) I wonder how it was hashing PkgIterators in your hashed_unique
       matcher.  The only hash I can find on package iterators doesn't
       use the boost::hash framework and isn't used anywhere.

       If it was falling back to some default hash that gave them all
       the same hash value, this could explain your performance
       problems.  (although the only way I can think of for them to
       be hashed is via conversion to Package*, which should work)

    2) I didn't think to suggest reserving space in the unordered index
       as well.  I don't think that explains nearly two orders of
       magnitude in performance loss, though.

  Daniel