Profiling is done with the ‘readahead’ package, and subsequently tagging ‘profile’ on to your kernel params list (once (in a while)) and using that to optimize scattered reads into big ordered readahead in the boot sequence. It used to squeeze some 10-20 seconds out of a 50-110s boot for me.

Note that readahead can be used (easily) to optimize other standard sequences (like desktop login) with greate success. Especially if you switch your PC on, go for a coffee and return to log into the desktop, it will be a great help if you use readahead from, say, rc.local, to get your profile data and app binaries buffered before time.