[clug] ARS: Ubuntu 9.10 boot optimizations: 5 second startup with an SSD

[David Schoen]

Yep, it helps, you pointed out lots of stuff I didn't already know :)

I'm going to do some more reading now.

Thanks,
Dave

2009/10/2 Alex Satrapa <alexsatrapa at mac.com>:
> On 02/10/2009, at 08:30 , David Schoen wrote:
>
>> Back when I still used Gentoo someone implemented a parallel boot ...
>
>> If Ubuntu can do it in such a way that it works well enough first time
>> around it may actually be adopted by the masses.
>
> There was an article about very short boot times:
> http://lwn.net/Articles/299483/
>
> Executive Summary:
> 1) Strip kernel down to absolute minimum needed to boot and launch init
> 2) Strip init-launched services down to Dbus, stripped-down X, sReadahead,
> HAL, udev, networking
> 3) Work to a blocks-read-in budget of 5 seconds x (disk read from platter
> kB/s).
> 4) Profile the system startup and put the most interesting blocks into the
> cache file used by sReadahead
> 5) Use a fast disk to boot
>
> The modified version of X includes not invoking the C complier every time it
> boots (ie: use the keyboard mappings that worked the last time you booted
> this hardware).
>
> They also boot directly into a user's desktop, since the budget was 5
> seconds to boot a usable system.  They also wanted to avoid "dirty tricks"
> like booting the user's desktop and continuing to start services (ie:
> heavily load the system) in the background.
>
> I'll add here my observation that "sleep" or "hibernate" actually takes
> longer to boot since you're potentially reading all of RAM from disk (eg: if
> you have 4GB of RAM, and a disk that handles 30MB/s, you'll be waiting a
> touch over 2 minutes to wake from hibernate).  There was another article I
> was reading by someone who suggested that if you could make your
> applications all able to recover cleanly from a crash (ie: save the do/undo
> buffer to disk regularly), it would actually be more sensible to simply
> "crash to sleep" and reboot from scratch when you wake the computer.  After
> all, what's the point of writing all the buffers and cache from memory out
> to the hibernate file?  Fsync, then stop the processor - save two minutes of
> battery life every time the user closes the lid.
>
> A 5 second boot is not using parallel booting. Parallel booting reduces a 1
> minute boot time to 30 seconds - the bottleneck is reading heaps of
> redundant data from disk.
>
> Hope this helps.
> Alex
>
> --
> linux mailing list
> linux at lists.samba.org
> https://lists.samba.org/mailman/listinfo/linux
>