[LinuxPPS] no_hz and pps?

[Hal V. Engel]

On Wednesday 17 December 2008 11:07:45 Udo van den Heuvel wrote:
> Found this discussion:
> http://linux.derkeiler.com/Mailing-Lists/Kernel/2008-02/msg05432.html
>
> So no_hz does take ntp into account?
> What goes wrong so that the performance is worse?
> (is it actually still worse in current kernels with no_hz?)

I rebuild my kernel with NO_HZ a few days ago as a test.  I did not 
gather detailed statics but I can report that it is fairly apparent that there 
is some decrease in clock accuracy compared to NO_HZ being turned off.   The 
difference is small enough that it would not be noticed if I were using 
network based time servers where root dispersion and jitter would be much 
higher and the affects of ON_HZ would be lost in the added noise.   But with a 
refclock the issue is significant enough that it is apparent.  

Without NO_HZ my worst case offsets were on the order of 50us and most of the 
time < 10us.  With NO_HZ my worst case offsets are on the order of 125us and 
seldom < 30us.  I am also seeing higher root dispersion numbers with NO_HZ; 
around 0.480 compared to around 0.300.  And higher jitter numbers as well; 
about 3 times as high but at times jitter was much higher.  With NO_HZ turned 
off my jitter numbers are typically 1us to 2us and almost never go higher.  
With NO_HZ I was seeing 5us to 8us most of the time and at times it was as 
high as 35us.  

I did these tests with kernel version 2.6.26.  Of course this is only one test 
but I think it indicates that NO_HZ should not be used with a refclock at 
least on my machine with this version of the kernel and I have disabled it.

On the plus side for NO_HZ my CPUs are running cooler by 2C to 3C when idle 
which indicates that NO_HZ is doing what it is supposed to and my machine was 
using less power.

Hal