Re: Desktop Kernel Stuff

[Joel Becker <jlbec evilplan org>]

On Wed, Jul 09, 2003 at 04:57:19PM -0400, Havoc Pennington wrote:
> It turns out that a large part of the speed hit is simply paging crap
> off the disk. :-/ Having everything in shell with lots of fork() may
> be part of the rest, but don't really know.
> 
> I guess parallel could help UI-wise if we could fire up gdm before all
> the other services had started, even if it doesn't speed things up.

	We did this for the New Internet Computer.  As a linux-based
appliance, startup time was uber-important.
	One of the worst offenders is disk seeking.  We found that a
linear startup was faster than a parallel one, because if apps are
loading in parallel the disk (in this case, a CD-ROM -- which seeks
slower than modern hard disks) seeks to high heaven.
	In the end, we wrote a custom app loader.  It merely mmaped()
and loaded the libraries and applications in order before executing them
(think of the seek pattern of lazy dynamic loading).  We then booted
just as far as we could before starting X.
	In the end, the boot took 58 seconds, of which 14-20 seconds
were the crappy BIOS the appliance had.  So, ~40 seconds from lilo
starting until netscape was running (a web appliance this was).
	Just some experience.

Joel

-- 

"Every new beginning comes from some other beginning's end."

			http://www.jlbec.org/
			jlbec evilplan org