Worthwhile optimization
by havoc
In case you had any doubts about how much David Zeuthen rocks, check
this out, standing on the shoulders of Owen and Ziga Mahkovec. As
I think Owen intended, this shows how important profiling before
optimizing really is. Now we understand the issues and a 40-second
boot (including GNOME login) is a simple matter of
programming. I’m looking forward to FC4.
In the boot
chart, it looks like David has gnome-session launching at about 22
seconds, so I have my hopes up that we can run gdm at that point when
using gdm instead of autologin, and have a username/password prompt in
under 30 seconds.
It does look like getting below the times David is posting will
involve simply loading less junk off the disk. It’d be interesting to
start analyzing what said junk includes and whether there’s any
low-hanging fruit.
A useful lesson also demonstrated by my D-BUS hacking: a wall-clock
sampling profiler is essential to figure out the high level areas to
optimize. Something like callgrind can be useful once you’re
optimizing a specific CPU-bound function or loop, but callgrind is
the wrong place to start.
(This post was originally found at http://log.ometer.com/2004-11.html#29)