On Wednesday 29 Oct 2003 10:48, Frank Griffin wrote: > Jos wrote: > >I did some testing myself, working on something similar, though I still > > don't know what people prefer: that I hack init, or replace /etc/rc.d/rc > > by a binary... > > Is the problem that the necessary initializations take too long, or that > they take too long when run serially under a single thread ?
1) We do too many things. One example is: running depmod every boot is nonsense: if modules are installed by rpm, rpm can call depmod. If the user installs a kernel by hand, make modules_install calls depmod. If the user hacks the modules by hand he is on his own. Depmod -A is supposed to be fast, but it still requires scanning a directory structure which hasn't been cached yet. Hacking together a kernel.h at each boot is nice for the few guys who change kernel every five minutes, not for the mainstream systems. My Mandrake starts NFS support by default. IMHO I should see no NFS services started unless configured by the user. Now they are started with empty tables. 2) We do them serial, which consumes extra time. Parallel is faster, the few reports of slower boot times are due to incomplete service dependency tables, causing services to timeout waiting for other services. 3) Ok, many things done don't take much time. but 20x not much time is still much time. 4) We don't give the user the feeling the system is quick. The famous "Windows XP is fast" remarks hold, for Windows XP makes the user think it is fast.This is almost unresolvable unless we can drop XFree86, which is utter bloatware. For init 3 (console) I managed to show the console within 3 seconds after init entered runlevel 3. WOW that looked fast. Sure, I can't do much for my network wasn't up yet, but the psychological effect was huge. Jos Windows XP is a trademark of the company which will probably sue me for comparing it to Linux.