Re: [chrony-dev] long time to re-sync with bad system clock

[Håkan Johansson <f96hajo@xxxxxxxxxxx>]

On Tue, 2 Feb 2010, Miroslav Lichvar wrote:

> On Sun, Jan 31, 2010 at 11:51:02PM +0100, Håkan Johansson wrote:
> > To avoid having to do that, I'd like to propose something like:
> >
> > If a new system (B) is to be elected reference source, taking over
> > from an old (A), then B must have made its last measurement at least
> > later than the current (half)span of A.  If it has not, a new
> > 'check' measurement by B is provoked (without it gaining reference
> > status yet).
>
> I'm planning to add some features to the source selection routine, but
> I'm not sure it should be able to initiate a measurement.
>
> In the case you describe, there are other things you can try beside
> lowering maxpoll:
> - set all NTP sources offline and make them online only when the GPS stops
> - increase the virtual delay for the GPS refclock so it's not marked
>  as falseticker so easily
> - use only stratum 2 or higher as NTP sources, so they are marked as
>  unreachable when synchronized to GPS

Ok, but I'm mostly interested in trying to devise a simple automated way. 
Bottom line would be that the current problem is sources with 'old' data 
disallowing 'newer' data sources just by being many and thereby winning 
the selection.  As soon as the old data sources actually do a new 
measurement, they realize their error and drastically reduce their poll 
intervals.  Effectively forcing a lot of new measurements, so doing an 
extra just earlier - if one can ensure that not to happen except when 
needed and limited, seems not to violate any network politeness to me.

I've looked at more transitions, and what is also happening rather often 
is that during a middle-period in the process, all sources are marked 
falsetickers, as there are two camps - old and new.  I even had some 
network sources with maxpoll 6, but those were not enough to help against 
the bigger camp of long ones.

My first suggestion was not really nice as it was 'edge-triggered', trying 
to look for the change if reference source.  Sometimes this change went in 
steps, from GPS to noone - to old-data source.  Another idea:

If there are sources which disagree, and these have a smaller jitter (i.e. 
might be good) and also have a time-span of their active linear fit that 
is smaller (or let's say half-smaller) than the age of the most recent 
measurement from others, then provoke a check measurement from the old 
disagreeing one(s).  Only if allowed by the minpoll by the old and not 
changing its poll.  It can then not happen more often than (half) the span 
of the disagreeing one(s).  Perhaps also require the disgreeing one to 
have held status as reference within the last poll interval of the 
old-data one.

Motto would be: if a source wants to disagree (with someone that was just 
recently trusted), it has to show some recent interest in verifying this.

Cheers,
Håkan