TC boot times

[Gavin McCullagh]

Hi,

On Thu, 17 May 2007, Matt Oquist wrote:

> I know ogra and co are working on this problem in general, but it
> might be worth mentioning my experience with this in here as well.

An analysis of what's taking so long might be a useful way for a few of us
admins to contribute.

This all reminds me of a few projects done in the last while including this
Debian GSOC project:

http://initscripts-ng.alioth.debian.org/soc2006-bootsystem/bootcharts.html

I wonder if we could put together a special ltsp chroot which added
detailed boot timings.  If we could gather a set of these together we might
be able to see more clearly what's slowing things down.  A wiki page on
ImprovingLTSPBootTime might be useful to pull data/ideas together.

A colleague who was involved in the above project mentioned this as useful:

http://www.bootchart.org/

> We have one collection of 12 old-ish laptops (650Mhz, 128MB) that take
> 5 minutes(!) to get to a login prompt with Edubuntu -- edgy or feisty.

> In contrast, that same laptop takes 38 seconds to get to a login
> prompt with LTSP 4.2. 

> In further contrast, we have scads of Netier thin clients (250Mhz, 64MB)
> that take only 1:50 to present an Edubuntu (edgy or feisty) login screen.

A central location to put timings and a standard way to collect them would
definitely be useful I think.  Perhaps it would be more suitable on an LTSP
site than an Edubuntu one?

> It's probably worth noting that the last 1:15 of the laptop's
> 5 minutes the 'gtk' greeter is pegging the processor, and this, at
> least, is significantly longer than the same process takes on the
> Netiers (with their much lower specs). I don't have any particular
> ideas about why this should be the case. But since ogra's working on
> a C implementation of ldm anyway, this part of the problem may just go
> away.

So, after X has started, ldm takes a further 75 seconds?  That's wierd.

AFAIR ltsp 4.2 didn't use dexconf, it had its own config method.  If you
snag the Xorg.conf on one of those really slow machines, stick it in the
chroot and use lts.conf to point it at that rather than using dexconf, I
wonder would the boot time get any better?

I seem to recall LTSP Jim and Scott saying they had an idea there might be
some nfs slowdown issue.  I'm not clear why that would be true of some
machines more than others.

Gavin