[ntp:questions] Second attempt at GPS-18 based NTP server

unruh unruh at invalid.ca
Fri Feb 3 01:36:14 UTC 2012


On 2012-02-02, David Woolley <david at ex.djwhome.demon.invalid> wrote:
> David Lord wrote:
>
>> 
>> I've just added the "minpoll 4 maxpoll 4" back to my ntp.conf
>> as it could reduce the 30 usec offset blips I get when logs are
>> rotated.
>
> That's almost certainly totally the wrong thing to do.  If you have 
> short periods of increased interrupt latency or network latency, you 
> want a long poll interval so that the samples with the high offsets (and 
> delays) are ignored.

The delays from refclocks are almost always all zero. Thus there will be
no info in the delay as whther a problem occured. ntpd will simply
assume it is as good as any other sample, and use it (the clockfilter
uses delays to select samples)

>
> Forcing a short interval will make ntpd track the temporary errors in 
> the time measurements.

Which is of course the right thing to do, since by assumption the
refclock errors are negligible. And there is no way that the system
knows that it serviced the interrupt 30us late. Now, if ntpd did a least
squares fit over a number of samples, it could see that these particlar
ones are outliers and perhaps get rid of them. But it doesn't and thus
cannot do so. 
. 
>
> Offset is not a direct measure of the timekeeping error.

It my not be a direct measure, but it is the only measure you have. And
since ntpd is blind to history, it cannot even use history to determine
such errors. 



More information about the questions mailing list