[Lava-users] Boot time measurement with Lava

Neil Williams neil.williams at linaro.org
Tue Aug 28 07:58:42 UTC 2018


On Tue, 28 Aug 2018 at 08:11, Denis HUMEAU <denis.humeau at st.com> wrote:

> Hello Tomeo,
>
> Sorry for this basic question, but how do you disable the serial console
> and still manage to run with Lava?


You can't disable the serial console entirely, what you can do is set the
kernel command line to quiet (but this will also hide a lot of useful
output). I would recommend using specific test job definitions which only
do a boot test when running with a quiet kernel and also ensure that the
actual kernel image correctly boots (without quiet) before running the
measurement test job. Equally, if what you are testing is kernel boot time,
be aware that you should only use ramdisk test jobs and use as small a
ramdisk as possible so that the INIT stage happens with as little fuss as
possible. As it's only a kernel boot test, for example, you do not want
INIT trying to get a DHCP address when what you want is a time measurement.
Different test jobs for different use cases. You certainly don't want the
extra time delay of communicating with NFS - that delay is not predictable
and the rest of INIT will be affected if the LAN is slow. Try to get a
cusomised ramdisk which (as near as possible) always takes the same amount
of time to run INIT.

The timing also depends on the DUT and the kernel configuration. If the
kernel is very noisy and produces a lot of debug to the console, it can
take some time for the DUT to send that information and LAVA cannot do any
measurements until the output has been sent from the DUT over serial. It is
lava-run which is doing the timing, on the worker, so it is as close as we
can get to the DUT.

The final option is to write a custom kernel module / kernel configuration
which measures and records the kernel boot time independently of the serial
port. Then access that value after boot from a Lava Test Shell. This gives
you full control over whether you are measuring only the kernel boot time
or kernel + INIT.



> I assume that serial console is mandatory for Lava to work.
>
> Regards,
>
> Denis
>
> -----Original Message-----
> From: Tomeu Vizoso <tomeu at tomeuvizoso.net>
> Sent: mercredi 22 août 2018 12:53
> To: Milosz Wasilewski <milosz.wasilewski at linaro.org>
> Cc: Denis HUMEAU <denis.humeau at st.com>; lava-users at lists.linaro.org
> Subject: Re: [Lava-users] Boot time measurement with Lava
>
> On 22 August 2018 at 10:46, Milosz Wasilewski <
> milosz.wasilewski at linaro.org> wrote:
> > On 22 August 2018 at 08:02, Denis HUMEAU <denis.humeau at st.com> wrote:
> >> Thanks Milosz,
> >>
> >> Do you have any idea on how intrusive Lava is on such boot processes?
> Based on what I see, without any further analysis, boot time is much higher
> when measured inside Lava. In my example, I measure Linux + Userland boot
> time, and I get about 20s more using Lava.
> >> As I said, I haven't analyzed yet.
> >
> > I'm suffering from the same problem. The time recorded in
> > auto-login-action might be affected by some actions LAVA performs on
> > serial. Recently there was 'kernel-messages' test result introduced.
> > It might give better match. But I guess we'll need to wait for some
> > core LAVA folks to comment.
>
> Don't know if there's anything LAVA-specific that disrupts boot time
> measurement, but when I have needed to track boot times I have always had
> to disable the serial console to get meaningful numbers. In most ARM
> boards, it can slow boot down by several seconds, depending on the amount
> of output.
>
> Regards,
>
> Tomeu
>
> > milosz
> >
> >>
> >> But, globally speaking, we have noticed so far that time measurements,
> benchmarkings, when performed via Lava, produce results significantly
> different than when performed outside Lava. Boot time is the latest
> example. I wonder if it is due to Lava (Lava may not be designed for this),
> or maybe is it due to our setup.
> >>
> >> Regards,
> >>
> >> Denis
> >> -----Original Message-----
> >> From: Milosz Wasilewski <milosz.wasilewski at linaro.org>
> >> Sent: mardi 21 août 2018 11:58
> >> To: Denis HUMEAU <denis.humeau at st.com>
> >> Cc: lava-users at lists.linaro.org
> >> Subject: Re: [Lava-users] Boot time measurement with Lava
> >>
> >> On 21 August 2018 at 10:56, Denis HUMEAU <denis.humeau at st.com> wrote:
> >>> Hello Lava users,
> >>>
> >>>
> >>>
> >>> Do you have any example or advice to share on a way to measure boot
> >>> time using Lava?
> >>>
> >>> Let’s say that I use u-boot, a minimal boot for a device running
> >>> Linux, and I want to measure the time spent between Power On
> >>> (pdu-reboot e.g) and Linux prompt.
> >>
> >> You can find it in the 'lava' suite results as 'auto-login-action'. It
> has a measurement which is roughly what you expect.
> >>
> >> milosz
> >>
> >>>
> >>>
> >>>
> >>> Best regards,
> >>>
> >>>
> >>>
> >>> Denis
> >>>
> >>>
> >>>
> >>>
> >>> _______________________________________________
> >>> Lava-users mailing list
> >>> Lava-users at lists.linaro.org
> >>> https://lists.linaro.org/mailman/listinfo/lava-users
> >>>
> > _______________________________________________
> > Lava-users mailing list
> > Lava-users at lists.linaro.org
> > https://lists.linaro.org/mailman/listinfo/lava-users
> _______________________________________________
> Lava-users mailing list
> Lava-users at lists.linaro.org
> https://lists.linaro.org/mailman/listinfo/lava-users
>


-- 

Neil Williams
=============
neil.williams at linaro.org
http://www.linux.codehelp.co.uk/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.linaro.org/pipermail/lava-users/attachments/20180828/8a0418ee/attachment-0001.html>


More information about the Lava-users mailing list