Suppose you bought one of our clocks and decided to test whether it really met its specifications. We claim a time accuracy of 1 us. Now imagine you don't have access to a clock that's more accurate than that.

You see how difficult this would be? You can certainly confirm that it might be as accurate as we claim -- that it isn't way off.

Now suppose you bought two clocks from us, and one was a little bit more accurate than the other. Could you tell which was which? How would you?

You could tell that the two clocks agreed to a very high degree. You could tell that both clocks were probably quite accurate, but determining which was more accurate would not be a simple task.

This is precisely the problem that ntpd faces when it has to choose what clock to use. Xntpd resolves the problem in part by asking clocks how accurate they are, but it also has to independently confirm their accuracy. This is needed both to ensure that it is not mislead by an erroneous clock and to detect a clock that is made inaccurate by network jitter and delay.

This causes a capture effect. What this means is that the clock that ntpd is currently synchronized to tends to appear a bit more accurate than it actually is, simply because ntpd is synchronized to it. Similarly, other clocks seem a bit less accurate than they really are since ntpd is not synchronized to them.

In general, this is a desirable effect. If two clocks are comparable in accuracy, it is probably better to stick with one of them rather than to switch back and forth, even if the other one seems a bit more accurate at the moment.

However, this can cause some surprising effects. For example, if you shut down a time server and then restart it, it can be several hours before some hosts begin using it. You will also find that hosts that are synchronized to a time server will report it as being more stable than those not synchronized to it.

Perhaps the most surprising effect is that adding a stable time server to your LAN will cause other timeservers to appear more stable! It is easier to show that another clock has millisecond accuracy if you have access to a clock with millisecond accuracy yourself.

Back