[LinuxPPS] Linux 2.6.28 - serial time string jumps

[Hal V. Engel]

On Tuesday 20 January 2009 07:42:45 Udo van den Heuvel wrote:
> Thorsten Mühlfelder wrote:
> > As far as I know:
> >
> > On Tue, 20 Jan 2009 16:08:08 +0100
> >
> > Udo van den Heuvel <udovdh at xs4all.nl> wrote:
> >> Pre-emptibale stuff settings
> >
> > good

Yes I agree this seems to be good.  At the very least it is not harmful.

> >
> >> HPET stuff
> >
> > bad

I have had good experiences with my HPET timer.  In my case the results were 
clearly better than the acpi_pm timer.  But this could be dependent on the 
specific hardware.  In most cases the HPET timer and other timers are using 
the same quartz oscillator but other than the CPU counter timer (tsc timer) 
the other timers will operate at lower frequencies and therefore have lower 
resolution.  Most HPET timers are 10 MHz (this is the minimum allowed by the 
HPET spec) or 25 MHz where as the acpi_pm timer is a little over 1 MHz.

In machines that have an invariant tsc timer it is probably best to use the 
tsc timer.  Based on stuff I have seen on the ntp and time nuts mailing lists 
if your tsc timer is not invariant then favor timers in this order if you have 
them:

hpet
acpi_pm 
jiffies 
PIT

>
> ?
> More high res timers == bad?
> COuld you please explain?

I have high res timers enabled and other sources recommend this.

>
> >> NO_HZ stuff
> >
> > bad
>
> So far (2.6.28) I see no worse performance, but my reception isn't that
> good.
> (working on it)

I noticed a slight decrease in the stability of my clock with NO_HZ enabled.  
But this could be dependent on other settings in the kernel configuration.  
For example NO_HZ might do better with higher settings for HZ.  I only tested 
with HZ=100.

>
> > And use 100 Hz!

With NO_HZ disabled I noticed decreased clock accuracy with HZ settings other 
than 100 but again this might be dependent of the specific hardware in use or 
perhaps other kernel config settings.

Hal