Re: [Ntp] Antw: Re: Antw: [EXT] Re: CLOCK_TAI (was NTPv5: big picture)

[FUSTE Emmanuel <[e-mail]>]

Le 11/01/2021 à 11:05, Martin Burnicki a écrit :
> Am 11.01.21 um 08:48 schrieb Hal Murray:
>> [e-mail] said:
>>> Why not a week before and a week after? I mean there are many arguments for
>>> and against different intervalks.
>> It has to be long enough and the transition from normal-to-smear gentle enough
>> so that ntpd will follow the smearing servers.  As long as the interval is
>> long enough, how long doesn't matter.  The big boys agreed on a number.  There
>> is no reason to pick a different one.
> Right.
>
> On the one hand, it has to be long enough that clients can easily follow
> the changing clock frequency from the server, but on the other hand, the
> smeared time is off the real current time during the smear interval, so
> you want an interval as short as possible.
>
> Finally, the length of the smear interval has to be a compromise.
Why pursuing with a sub-optimal server side implementation ?
We are defining NTPv5. v5 client are to be implemented an deployed. 
There is no install base.
v4 smearing servers where done to cope with a v4 base where no provision 
was made for smearing instead of jumping.

On the client side, you could implement a very very short optimal smear 
interval on top of the main algo.
If this smear "profile" is defined in the v5 spec. All client side 
smearing clients will behave the same, and even the mapping between 
smearing and non smearing clients will be know.
No more lies, no more client following problems, no more smearing/non 
smearing servers. It 's purely a client capability or policy problem and 
should be done on the client side.

Emmanuel.