[ntpwg] REQUEST FOR TEXT - Protocol Doc Changes

[David L. Mills]

Stewart,

Yes, the switch can cause significant wiggle due to output blocking, 
retransmissions, etc., and that'w why 1588 requires special switches 
which correct for stochastic delays by retiming the clock.

There is another feature of 1588 called the two-step protocol which 
provides the transmit timestamp in a following packet. You may see my 
suggestion to incorporate this in NTP in the architecture briefing on 
the NTP project page linked from www.ntp.org. The resulting protocol has 
much in common with 1588 with the additional feature that it is symmetric.

The two-step protocol seems in principle not hard to implement as long 
as timestamps can be captured by the driver, not necessarily requiring 
1588 NICs. I lay down the challenge, but not necessarily committing my 
own elbowgrease for driver modifications. Much better somebody who knows 
the Unix internals much better than I. I will, however, committ my 
grease to the NTP protocol modifications and support.

Dave

Stewart Bryant wrote:

>
> My first thought was that the timing reference point
> should be the first bit of the NTP payload, since
> this was within the NTP scope and everything before it
> was a variable length.
>
> However on reflection after reading the thread I think
> that the only point in the packet that you can rely
> on is start of frame, and this is by far the
> easiest point in the frame for the receive hardware
> to capture and locally timestamp.
>
> The variable number of bits in the datalink header
> and the IP header (IPv4/v6) in principle be
> corrected because
>
> a) Both transmitter and receive know (at least in
> principle) how many there are and can apply corrections
>
> b) If you go through some switching equipment that
> changes any of these factors, then in general
> the errors introduced by the switching will
> be significant compared to this error, although
> in principle the switches could correct the timestamp
> accordingly as part of some sort of one step
> transparent clock.
>
> - Stewart
>
>
>