Re: [Ntp] Smearing

[Tony Finch <[e-mail]>]

Miroslav Lichvar <[e-mail]> wrote:
> On Tue, Oct 19, 2021 at 05:15:16PM +0100, Tony Finch wrote:
> >   * 24h smear, not too long but long enough that NTP clients will notice
> >     the rate change even with a 1024 second polling interval, and the rate
> >     change is small enough that NTP clients are willing and able to follow it
>
> If there are stricter requirement on maximum offset between different
> clients, a longer interval might be needed.

I don't think anyone has implemented a longer smear?

> >   * smear from midday to midday around the leap second, to keep the phase
> >     error below 0.5 seconds
>
> A disadvantage of this approach is that it relies on the leap second
> to be known at least 12h before the leap. With some reference clocks
> it's known only much closer to the leap and it takes time to be
> distributed by NTP for a stratum >=2 server.

The history of bugs in this area suggests that an NTP implementation
should not blindly trust a reference clock to indicate a leap second
correctly. The NTP server should have a copy of the leap second table,
e.g. from the tz database, perhaps cross-checked from multiple places, so
it knows months in advance when a leap second is due. Then it can use any
smear profile, and verify whether the leap is being handled correctly by
each of its time sources. And higher-stratum servers can and should do
the same.

> >   * use a linear smear, because NTP clients are better at following
> >     constant rates (a cosine or other smoothing means the rate changes
> >     continuously during the smear, which NTP does not handle so well)
>
> A linear smear is easy to work with for humans, but not for NTP. A
> cosine smear, or even better a quadratic smear, reduces the maximum
> time error of the client's clock as it can more closely follow the
> frequency offset.

Ah, I thought Google had said their experience was the other way round,
but I remembered wrong. Their web page doesn't explain why they switched
from cosine to linear. (Maybe they wanted the kernel to be able to
unsmear?)

https://developers.google.com/time/smear)

Meinberg has a nice comparison at the link below, though it's missing a
comparison for clients with variable polling intervals and a 24 hour
smear.

https://kb.meinbergglobal.com/kb/time_sync/ntp/leap_second_smearing/ntp_leap_smearing_test_results

Tony.
-- 
f.anthony.n.finch  <[e-mail]>  https://dotat.at/
Tyne, Dogger, Fisher: West or southwest 4 to 6 veering northwest 6 to
gale 8, increasing severe gale 9 at times. In Fisher moderate or
rough, becoming very rough or high later, elsewhere slight or moderate
becoming rough or very rough. Rain or squally thundery showers.
Moderate or good, occasionally poor.