Re: [Lennart Poettering] Re: A few observations about systemd
On Mon, Jul 18, 2011 at 04:03:41PM -0700, Steve Langasek wrote:
> On Mon, Jul 18, 2011 at 11:14:39PM +0200, Mike Hommey wrote:
> > > I'm sure that systemd does much better than a traditional sysvinit boot with
> > > /bin/bash and no dependency-based booting. But then, so does Debian's
> > > current boot system, and so does upstart; and neither of the latter two
> > > involve grandiose claims of a "shell-free boot". Trying to take the shell
> > > completely out of the boot means a definite tradeoff here between boot speed
> > > and configurability/maintainability, and in the absence of hard numbers, I
> > > suspect this is a false optimization and not a trade-off that we actually
> > > want to make in a general distribution. Which then calls into question the
> > > use of such claims as a justification for a switch to systemd at all...
> > I'd expect some important differences between shell script based init
> > and systemd-type init by the simple fact that there are (or at least
> > should be, I don't know how systemd actually works) less files to read.
> > Less files to read == less disk seeks. Disks seeks hurt startup performance.
> Yes, I've read your blog entries on the subject. :-) That's true, but I
> think the reduction in the number of files being accessed, for systemd vs.
> sysvinit or upstart, is rather small; aside from some things in /etc/rcS.d,
> most init scripts would have approximately a 1:1 correlation with upstart
> jobs or systemd config files, and if you've read the shell off disk once
> it's in cache and there's not likely to be any more seeking. So I do expect
> that most of the shell penalty will be CPU rather than disk in the context
> of boot.
But you don't only load the /etc/rcS.d scripts, you also need to load
all the executables the scripts use, and their dependent libraries, and
the files they need to read. But yeah, that would still need to be
accounted and compared.