LWN: Comments on "MontaVista Achieves Ultra-Fast One Second Linux Boot Time in Embedded Industrial Applications"

Wrong goal

tkitch — 2009-07-21T22:25:48+00:00

See demo here for more clarification on the demo: http://www.youtube.com/watch?v=-l_DSZe8_F8

cold boot?

yaneti — 2009-07-17T09:44:10+00:00

> A modern bios takes between 1 and 2 seconds (if yours takes more, talk to your hw vendor...).

Hmm, then how about this DG945GCLF that I have here, that with absolutely nothing attached takes 10 seconds to tell me there is no boot device.
With a single SATA drive grub shows up in 13.
You know a modern bios, from Intel, on a hardware somewhat related to the same netbooks that you are booting moblin on in 5?

cold boot?

nye — 2009-07-16T11:04:54+00:00

Sadly this board has two SATA controllers, and one of them doesn't seem to be controllable via the main BIOS except to disable it, plus it seems that the only way to disable autodetection is to enable PATA emulation mode.

This is all the sort of thing that doesn't tend to get mentioned in motherboard reviews, which is why I'm interested in any resources that do check and give that information.

cold boot?

dlang — 2009-07-15T19:09:39+00:00

you can configure most BIOS to not do autodetection of drives, that can drasticly speed up the system.

cold boot?

tzafrir — 2009-07-15T14:27:19+00:00

How about replacing the BIOS?

cold boot?

nye — 2009-07-15T13:56:34+00:00

I've never had a BIOS that fast :-(

Actually I think most of my BIOS time is spent enumerating SATA devices, which takes *bloody ages*.

Do you have any recommendations on what (consumer-level) hardware to get which is known to have a decently fast BIOS - or know of anywhere that bothers to document that sort of thing?

cold boot?

arjan — 2009-07-15T13:13:25+00:00

I'm measuring from bios handoff.
A modern bios takes between 1 and 2 seconds (if yours takes more, talk to your hw vendor...).

I'm personally not convinced that coreboot is fundamentally faster than a decent bios; the long time taking tasks are shared (bringing up/discovering memory etc etc)....

cold boot?

pjm — 2009-07-15T12:07:36+00:00

Arjan, are you measuring from cold boot or from when Linux has loaded? I was under the impression that most netbooks take several seconds from cold boot to when Linux starts to be read. The article is quite explicit that it's measuring from cold boot, from cold power.

I believe the fact that it's measured from cold boot also goes some way to addressing Bruce's point above: it's not at all true that everyone gets [init started] in one second.

(Coreboot is somewhat relevant here, though I've conflicting evidence on how fast it is, ranging from milliseconds to 0.5 seconds to 3 seconds is the current fastest; possibly those are each measuring different things.)

Wrong goal

drag — 2009-07-15T00:14:07+00:00

> of course this is with some special purpose embedded hardware. but why is my laptop taking 20 seconds to grub? why? why? why?

'cause BIOS is teh suxor.

Blame it on technical requirements needed to be compatible with Windows XP.

Wrong "stack"

arjan — 2009-07-14T21:37:06+00:00

the kernel has been fixed since; typical is around 0.6 seconds for a kernel boot with all drivers in there....

Wrong "stack"

BrucePerens — 2009-07-14T21:29:53+00:00

Busybox got to the prompt really quickly back when I wrote it in '96. Most of the time spent in booting back then was waiting for the kernel to not detect devices.

Wrong "stack"

jeff@uclinux.org — 2009-07-14T21:10:37+00:00

"achieving a level of previously unmatched performance on any embedded Linux."

Wrong Goal and also factually incorrect. uClinux has been booting on industrial controllers in
under 1 second to functional, completely running state for 10 years. And on processors running
very low double digit MHz clock speeds (and yes, from Motorola, now Freescale too). The key is not
so much replacing init with a single purpose application (that can help), but making a proper tiny
stack that does just what you need, of which init is a part. Thankfully, just what you need is not so
different as one might initially think from application to application.

If you need to do "tuning of the entire software stack" then you're using the wrong stack.

Wrong goal

dmk — 2009-07-14T21:10:18+00:00

of course this is with some special purpose embedded hardware. but why is my laptop taking 20 seconds to grub? why? why? why?

is there some comprehensive overview of bios-boot-time for laptops?

Why?

JoeBuck — 2009-07-14T21:07:10+00:00

They are selling to designers of embedded systems, which (at least in many cases) really have one application to run.

Wrong goal

arjan — 2009-07-14T21:05:48+00:00

Doing a full boot including starting X on a single core netbook (eg slow box) is somewhere between 2 and 3 seconds (which includes udev/consolekit/etc).

It's not all that special to get that to sub-2-seconds I suppose; but to get that to 1 second will take some effort still ;)

Wrong goal

BrucePerens — 2009-07-14T20:36:14+00:00

It works this way for everyone if you replace init with your application. At least if you aren't using udev and don't need to bring up many services. It would be a lot more credible if they demonstrated a versatile, capable, and then fast start-up service than just getting in to an application quickly.