[systemd-devel] bootchart issues on slow hardware

[Frederic Crozat]

Le lundi 18 mars 2013 à 14:34 +0100, Lennart Poettering a écrit :
> On Mon, 18.03.13 09:39, Frederic Crozat (fcrozat at suse.com) wrote:
> 
> > Le dimanche 17 mars 2013 à 14:54 +0100, Kay Sievers a écrit :
> > > Here is a chart:
> > >   http://people.freedesktop.org/~kay/bootchart-20130317-1434.svg
> > > 
> > > Rotating media and really cheap hardware looks very sad, and we take
> > > like 5 times longer to boot than Windows 8.
> > > 
> > > Why does bootchart stop before all the *really* slow desktop stuff starts?
> > > 
> > > Why does bootchart claim an idle time of ~5 sec, not that I wouldn't
> > > like that, but .... :)
> > 
> > Another thing we lost with current readahead (compared to what we had
> > with meego's readahead) is prefetching user session files (say
> > gnome-shell, files in /usr/share/applications, etc..), which, on
> > workstation used by a single user is really having an impact on
> > perceived system "speed" (ie when user can start working).
> 
> Hmm, how would we have lost that? the readahead stuff stays running for
> a while after the system job queue is empty, in order to cover the user
> session and suchlike.

I've checked readahead data on my test "netbook" and unfortunately, the
system is taking soo much time to start GNOME 3 after the job queue
empty that it doesn't catch enough readahead data on the user session
(and it can only work if you enable autologin). I've checked what we
were doing on Meego and the timer was 15s instead of 10s (not sure it
would gain much more). 

What is frustrating is to see IO going near 0 once the readhead replay
is done and then, going up just for the user session, and in between, a
nice gap which would love to be filled with IO prefetching ;) And when
you start digging in the readahead data, you can see ugly prefetch, like
stuff in /var/log/* :( 

-- 
Frederic Crozat <fcrozat at suse.com>
SUSE