Re: [PATCH] [RFC] timekeeping: Rework frequency adjustments to work better w/ nohz

[John Stultz <john.stultz@linaro.org>]

On 01/13/2014 09:51 AM, Richard Cochran wrote:
> On Mon, Jan 06, 2014 at 07:57:03PM -0800, John Stultz wrote:
>> I still think this is probably 3.15 or later material, but I'd be
>> very interested in feedback, thoughts, and testing.
> Over the weekend I did a short test of this new approach. I compiled
> two kernels, one plain v3.12.7 and one with the proposed fix. The
> test was run on three different kernel variants, the current nohz
> kernel, the same with nohz=off, and the same but with John's patch.
>
> I used a simple test script (see below). Using a PCIe express card as
> a PHC, the Intel Corporation 82574L, I simply let the this clock run
> free (not sync'ed to gps or anything), and then synchronized the Linux
> system clock to the PHC using the phc2sys program with a sample rate
> of once every 32 seconds. Each of the tests ran for at least three
> hours on a system without any other load.
>
> - Linux 3.12.7-nohz-plain-20140106                nohz-plain.log
> - Linux 3.12.7-nohz-plain-20140106 NOHZ=OFF       periodic-plain.log
> - Linux 3.12.7-nohz-fix-20140106-00001-gd753140   nohz-fix.log
>
> The performance in the log files as reflected in the clock offset is
> summarized in this table. The values are in nanoseconds.
>
>  |         | periodic-plain |      nohz-fix |    nohz-plain |
>  |---------+----------------+---------------+---------------|
>  | minimum |  -1.599000e+03 | -1.051000e+03 | -5.373700e+04 |
>  | maximum |  +1.311000e+03 | +1.048000e+03 | +6.389500e+04 |
>  | mean    |  +9.880240e-02 | -7.747305e+01 | +1.597904e+01 |
>  | stddev  |  +4.610021e+02 | +3.960978e+02 | +1.491263e+04 |
>
> I also made two graphs.
>
>    http://linuxptp.sourceforge.net/nohz-fix/current_nohz.png
>    http://linuxptp.sourceforge.net/nohz-fix/periodic_vs_fix.png
>
> (The log files and scripts are also in that directory.)
>
> So in this test, it looks like the new approach performed at least as
> well as using regular, periodic ticks.

That's great to hear! Thanks so much, I really appreciate the testing!
And this is with HZ=?

If you do get a chance to look again, I'd also be interested if running
with nohz=off w/ the fix doesn't show any regression compared to the
unmodified nohz=off case.

I've been trying to validate the ntpd case to ensure there aren't
regressions when there's more erratic steering, but unfortunately got
side-tracked with what seems to be an ntpd/virtualization issue.

thanks
-john