[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Serious simulation performance drop



On Wed, Oct 01, 2008 at 07:05:44PM +0200, Ferenc Wagner wrote:
> Michael Vogt <mvo@debian.org> writes:
> > On Tue, Sep 30, 2008 at 05:48:57PM +0200, Ferenc Wagner wrote:
> >> Btw, I noticed that on Lenny:
> >> 
> >> $ time /usr/bin/apt-get -o 'Debug::NoLocking=true' -s -qq upgrade
> >> 
> >> real	0m4.351s
> >> user	0m4.328s
> >> sys	0m0.020s
> >> 
> >> takes much more time than it does on Etch:
> >> 
> >> $ time /usr/bin/apt-get -o 'Debug::NoLocking=true' -s -qq upgrade
> >> 
> >> real	0m0.800s
> >> user	0m0.776s
> >> sys	0m0.020s
> >> 
> >> even though the Etch machine has half the bogomips and more than twice
> >> the /var/lib/apt/lists than the Lenny one (which is a paravirtual Xen
> >> domain on a practically idle host, but this shouldn't matter).
> >> 
> >> Did anything change between versions 0.6.46.4-0.1 and 0.7.14 that
> >> could explain such a big difference in CPU consumption?
> >
> > Thanks for brining that up. I fixed it in my development bzr tree and
> > merged it into the apt debian-sid branch.
> 
> Way cool!  I curiously clicked on the VCS link on
> http://packages.qa.debian.org/a/apt.html, but didn't see anything.  I
> know nothing about bzr, though so it may be quite normal...  Then
> could you please summarize in a couple of words what the problem was?

bzr.debian.org does not have a web UI for bzr unfortunately. The
change is due to the fact that libapt now has the automatic dependency
information (thanks to Daniel Burrows!) calculation and that it
calculates for each operation what packages are auto installed and can
be removed. That is expensive. There is a way (via an ActionGroup) to
make it not do that and I added that to the simulator.

I uploaded the fix to experimental along with the ABI fix from Daniel
(thanks!). Feedback on this would be very welcome so that it can go
into sid ASAP.

Cheers,
 Michael


Reply to: