[ntp:questions] Setting the maximum rate of change

Harlan Stenn stenn at ntp.isc.org
Tue Mar 20 09:57:21 UTC 2007


>>> In article <45ffa089$0$9851$426a74cc at news.free.fr>, Spoon <devnull at localhost.com> writes:

Spoon> Here's why I ask: I'm working with a standard that deals with 27 MHz
Spoon> clocks, and that standard states that the frequency must not change
Spoon> faster than 75 mHz per second.

Harlan> If ntpd has the correct drift adjustment, you should be in good
Harlan> shape.

Spoon> drift or skew?

The value in the driftfile.

If you run 'ntpq -c rv' it will be the 'frequency=...' value.

Harlan> Under what exact conditions must the "don't change faster than
Harlan> 75mHz/sec" be met?

Spoon> Consider two systems A and B. At initialization, B's clock is set to
Spoon> A's clock (offset = 0 initially). After 30-60 seconds, it is

You mean for the first 30-60 seconds, right?

Spoon> acceptable to make a large correction to compensate for the average
Spoon> skew. From then on, B's skew correction parameter should not be
Spoon> changed faster than 75 mHz per second.

OK.

Spoon> I'll try and think aloud, in case someone can see through my
Spoon> confusion.  Consider H : a 26,999,900 Hz clock.  Thus H has an offset
Spoon> of 100 Hz = 3.7 ppm

Spoon> I am talking about a frequency offset, i.e. what NTP calls skew
Spoon> AFAIU.

OK, I think we're on the same page now.

Spoon> In other words, H "misses" 100 ticks every second.

Harlan> If left uncorrected, yes.

Spoon> So far, so good.

Spoon> I'd have to add 100 / 27e6 = 3.7 µs every second to keep H from
Spoon> drifting away from the correct time.

Harlan> Once ntpd has achieved "state 4", ntpd will calculate the drift and
Harlan> will automatically handle this.

Spoon> According to the definition's page you've mentionned, drift is the
Spoon> variation in skew, I think you meant ntpd will calculate the skew?

I'll have to reread the definitions page again.  OK, it's directly neither
skew nor drift.  It's the frequency adjustment needed to keep the clock
accurate, and it accomodates the actual frequency of the clock crystal and
the value of HZ.

Yes, ntpd will calculate it and adjust for it.

Harlan> Are you comfortable that the specs of the machines you are using
Harlan> will keep time well enough that their clocks can be kept in-sync by
Harlan> limiting adjustment to 75mHz/sec?

Spoon> I'm still testing this. Basically, the two machines will remain at
Spoon> constant temperature. I'm hoping that clock drift on either machine
Spoon> will be very small, i.e. skew between the two clocks should remain
Spoon> almost constant. Thus, once ntpd has computed the correct skew,
Spoon> everything should be OK.

Harlan> And have you read about the problems you can cause by having HZ at a
Harlan> value other than 100?  Particularly with Linux kernels?

Spoon> No. Where is that?

I just looked and I thought I saw somebody post something about this here,
although it might have been on hackers at .  I just looked, and I could not
find it.

It was a discussion about how one had to be careful to make sure that the HZ
value was "properly" divisible, and also not so fast that interrupts were
lost.

Have you just tried this yet?

H




More information about the questions mailing list