| Re: [chrony-dev] Poll adjust after long time unreachable? |
[ Thread Index |
Date Index
| More chrony.tuxfamily.org/chrony-dev Archives
]
On Fri, Aug 16, 2013 at 11:07:04AM -0700, Bill Unruh wrote:
> On Fri, 16 Aug 2013, Miroslav Lichvar wrote:
> >One reason for a larger number of missed polls might be overloaded
> >server. If all clients dropped to minpoll immediately, it would only
> >make it worse. I'd suggest to focus on improving the adjustment from
> >the error in the prediction instead of hardcoding a drop to minpoll.
>
> Hmm. valid point. But I suspect that the more usual case is that the internet
> connection goes down for a while. This means that it takes a long time for the
> system to respond when the internet comes up again especially as it has
> climbed in poll interval while the network was down. And I am not suggesting
> that they drop to minpoll immediately (tx_count >5 so that means 4 missed
> return packets). Thus at poll 10, it takes 20 min for the system to finally
> get another packet, and then it will take a few of those for the system to
> finally drop down in the poll interval. Ie, it could take a long time to
> recover from a dropped network. One of the nice features of chrony is the
> rapidity of its response, and this works against that.
The check error in the prediction vs peer distance should reduce the
poll interval if the connection was down for a long time.
Also, if the reason for the disconnected network is unplugged cable or
unassociated wlan, the system scripts should call chronyc offline and
chronyc online when the connection is back. If this is working
correctly, the polling will be restarted immediately.
> By the way, have you ever thought about making chrony use say a quadratic
> predictor, rather than just a linear? Not sure if that would make the
> prediciton better or worse in some cases however. In general for slow drifts
> in the rate it should make it better. But sometimes on severe changes, it
> could make things worse.
I didn't thought of that. If the changes in the frequency are random,
could a quadratic predictor really improve the error?