Re: [Ntp] Antw: Re: Antw: [EXT] NTPv5 Loop Detection without Stratum

[Miroslav Lichvar <[e-mail]>]

On Tue, Aug 30, 2022 at 01:51:43PM -0700, Hal Murray wrote:
> That brings up an interesting point.
> 
> What does loop mean?

I think it's when a clock is synchronizing directly or indirectly
(over one or more other clocks) to itself. As a test, you can inject a
small offset to the server's clock and see if the impulse comes back
later from one of its sources.

In the graph theory, it's a loop in a directed graph.

> The draft for NTPv5 is only the wire protocol.  It doesn't cover what goes on 
> inside a server.  If a server latches on to the best upstream server and only 
> uses any others for checking, then the topology is simple and loop would be 
> easy to define.

Such a loop is guaranteed to break when the stratum reaches maximum,
but it can repeat if some of the sources switch to something else and
back.

> Is that a reasonable assumption?  Do we need a section in the draft to collect 
> assumptions about the internals of servers?
> 
> Consider the alternative where a server does some sort of weighting on the use 
> of its upstream servers.  What does stratum mean?  Is loop well defined?

This should be specified in the draft. As David suggested, stratum could be
the maximum+1 of all sources. Maybe we can provide both minimum and
maximum. This needs to be tested.

> Do we want 2 different client/server packet modes/formats?  One for 
> leaf-clients (SNTP) and another for the client side of servers?  Or maybe the 
> same packet format but different descriptions of the way fields are used.

That sounds like an unnecessary complication to me.

> Or maybe even 2 separate RFCs.  The idea being that SNTP will be much easier 
> to understand and we can target it for a longer lifetime.  In particular, it 
> wouldn't have to say anything about loop detection.

I don't like duplicating text or code. We just need to make sure the
document is clear on what parts are not required for client-only
implementations.

I think SNTP vs NTP is more about algorithms, not the protocol.
Without the symmetric mode, NTPv5 will be very easy to implement.

-- 
Miroslav Lichvar