Ubuntu boot speed fall in Hardy

[Sitsofe Wheeler]

On Sat, 2008-05-10 at 12:53 -0700, Bryce Harrington wrote:
> It's curious Fedora 9 showed such poor results compared with Ubuntu (and
> compared with Fedora 8), given that they are listing fast Xorg boot as a
> feature.  http://fedoraproject.org/wiki/Features/OneSecondX

I wouldn't say it is surprising compared to Ubuntu - if you look at the
Fedora chart ( the dates link to charts like
https://wiki.ubuntu.com/BootCharting?action=AttachFile&do=get&target=SitsofeWheeler-fedora-9-beta.png ) you can see the machine is massively CPU bound throughout and peak I/O throughput isn't that great (at least compared to distros with some sort of preload/readahead although one would need test the benefit of that versus the time it takes to do). Additionally some of its boot services seem to do a fair amount of writing to the disk causing kjournald to use up IO. Now Fedora are well known for including huge amounts of debugging in their kernels while the distro is in alpha/beta testing so perhaps this isn't yet representative of the true final speed. Compared to Fedora 8... something funny seems to be happening around udevsettle though (9s seconds versus 14s) and the time it takes for kernel to start appears to be longer in the F9 chart. Further the new gdm just isn't as fast at starting autologin as the old gdm (but you can't see that on the chart).

Unlike the chart for Hardy though, there is only one small gap of no
CPU/IO usage once userland has started so the long times don't seem to
be predominantly due to a slow X (although Xorg is started twice so X
speed will matter more in Fedora than Ubuntu). Rather, Fedora just seems
to start and do a huge amount. It's definitely a distro for higher spec
machines capable of crunching through stuff at a better. Looking at the
services it starts it seems geared towards more traditional *nix
corporate desktops / servers.

> I'll be interested to see if the fast Xorg boot stuff in the upcoming
> Xorg 1.5 will boost our boot numbers, or if the Xorg boot time just gets
> lost in the noise.

I should think it would help (at least in the time to the clock test
rather than to gdm) as GNOME tasks should be kicked off sooner. However
fake gains could be made by allowing GNOME to be responsive even when
the clock (which is often the last applet to load) hasn't finished
loading. However if more GNOME utilities need to be started any gains
will be washed away in the autologin case.

Some more boot comparisons can be seen on
http://www.harald-hoyer.de/linux/boot-time-distro-comparison .

-- 
Sitsofe | http://sucs.org/~sits/