[ntp:questions] NTP offset doesn't change.

brian utterback brian.utterback at oracle.com
Wed Feb 11 13:22:48 UTC 2015


On 2/11/2015 2:12 AM, Harlan Stenn wrote:
> William Unruh writes:
>> > On 2015-02-10, Terje Mathisen <terje.mathisen at tmsw.no> wrote:
>>> > > William Unruh wrote:
>>>> > >> No. It only does that for "offsets from Hades". The Ones from Hell, ntpd
>>>> > >> abandons all hope and quits. ( Hades is 128ms to 1000 sec, Hell is
>>>> > >>       >1000 sec)
>>>> > >> Ie, for <128ms, ntp will try to slew the clock ( at a max of 500PPM- as
>>>> > >> far as I can see a completely arbitrary limit Mills decided on decades
>>> > >
>>> > > The 500 ppm limit is not at all arbitrary!
>>> > >
>>> > > In fact, it was originally just 100 ppm, but when too many systems 
>>> > > turned up with a system clock which was a bit too far out, Prof Mills 
>>> > > redid the control loop to allow a 500 ppm range.
>>> > >
>>> > > It could have been a lot more, but the ultimate stability of the control 
>>> > > loop is supposed to be better this way.
>>> > >
>>> > > My own control theory math was back around 1980, so I have forgotten 
>>> > > most of it. :-(
>>> > >
>> > 
>> > As you state, it is arbitrary.
> Is not! Is so! ...
>
>> > If it can be changed from 100 to 500
>> > after complaints, it indicates that the number was not picked to
>> > optimise anything.
> I'm not sure that logically follows...
>

Well, it indicates that it can be changed, not how easy it is nor what
it entails nor the consequences. As I recall, the choice of PPM limit
directly affects the maximum error budget. In Mills book he explains
that the control loop is carefully crafted to simplify the calculations
and to be provably correct and stable. With today's systems I am not
sure simplifying implementation at the expense of clarity is necessary
and provably correct is a goal that most of us worry about.

Dr. Mills crafted a wonderful piece of software, amazing for its time,
but he no longer actively engages us much at all. I understand, that
isn't his fault. But no one who does actively engage really understands
it or knows how to improve it. Unruh has a point, we don't know if there
isn't a better way built on statistical analysis. Perhaps a hybrid
between the two approaches would be better still. But we don't even know
the consequences of changing a single constant with any degree of
certainty.

-- 
Brian Utterback
Solaris RPE, Oracle Corporation.
Ph:603-262-3916, Em:brian.utterback at oracle.com



More information about the questions mailing list