[ntpwg] On preamble and trailer timestamping and 100-Mb Ethernet

[David L. Mills]

Guys,

You may recall that my advice was to leave the media timestamp positions 
out of the NTP specification and put it in a separate document specific 
to each network architecture as in IEEE 1588. But, as in the recent 
essay posted, if the general feeling is to the contrary and that a 
specific position is necessary, the best choice would be as in 1588 
following the preamble and SOF octet. Well, a closer examination 
comparing 10BASE-T and 100BASE-TX Ethernet shows the issues are a bit 
more complicated.

Since 10-Mb Ethernets were evolved from half-duplex CSMA/CD technology, 
each frame begins with a preamble to wind up a phase-lock loop and 
demodulate the Manchester code. As the number of preamble bits to lock 
varies somewhat, the natural choice for the on-time mark is the SOF 
octet. So far as I know, this is still the case even if the segment is 
full-duplex.

On the other hand, 100-Mb Ethernets are full-duplex, so the data bits 
flow all the time and no preamble is necessary. Date octets are split in 
two 4-bit nibbles and each nibble encoded in a 5-bit channel symbol. 
Channel symbols occur in pairs, a designated pair for IDLE, another pair 
JK for the beginning of the frame and another pair TR for the end of the 
frame. The frame data itself, which is encoded first in NRZI and then 
MLT-3 and then scambled (!!), includes a conventional Ethernet preamble, 
SOF, data and checksum. The NIC chip doesn't care about the encapsulated 
Ethernet frame and, in particular doesn't know if it contains a 
preamble, SOF or anything else.

According to this reasoning, a 10BASE-T 1588 chip doesn't know when the 
preamble begins and a 100BASE-TX 1588 chip doesn't know where the SOF is 
or even if it exists. It isn't clear from the chip spec sheet I have 
whether the encapsulated preamble is by convention always included or 
not, the chip has provisions to count the number of octets in the 
original frame and smash a 64-bit 1588 timestamp right in the data 
stream. Heroic, that.

I can mumble on about the T1 extended superframe, which in fact is a 
neat candidate for NTP. I once proposed to MCI they could encode NTP 
timestamps in the frame overhead. SONET and FDDI and ADSL and ISDN and 
WiFi would be fun.

Life is very complicated.

Dave