[ntp:questions] Re: NTP4 has 3 different time formats! Namly (32, 64, 128) bits wide

Terje Mathisen terje.mathisen at hda.hydro.com
Mon Jul 17 06:02:54 UTC 2006


Richard B. Gilbert wrote:
> David L. Mills wrote:
> 
>> Richard,
>>
>> I can't claim preconition, as the current NTP timestamp format was 
>> invented in 1978 when nominal accuracies were in the 16-ms range. 
>> However, the resolution limit of 232 picoseconds is likely to be 
>> exceeded when the CPU clock rate approaches 4 GHz, which might not be 
>> long off.
>>
> 
> I suppose that even a 2 GHz machine could slice time into 500 picosecond 
> increments.  But I was thinking in terms of the ability to set a clock 
> that accurately.  There's no way that I can think of that it could be 
> done over a network using today's technology.  I'm seeing a ~4us delays 
> on my 100 Mb full duplex LAN.  I think that means I can't pass time from 
>  machine A to machine B over my LAN without an uncertainty of ~2us. The 
> error is probably less than that but probably is the best we can say.
> 
> So you could get delta time measurments with 232 picosecond resolution 
> but getting absolute time accurately with that precsion is not going to 
> be easy.

If you can get _repeatable_, at least some of the time, 4 us delays, 
then you can use the same statistical methods which NTP is already using 
to get absolute accuracy close to an order of magnitude better, i.e. 
half a us or so.

The real limiter will be the need for (a) a really good local clock 
source, i.e. better than the current 10 cent (?) quartz crystals, and 
(b) a hardware method to measure the interrupt latency.

Poul-Henning of FreeBSD and NTP fame did both on his _very_ good NTP 
servers, it might be possible that some new motherboards will include a 
timing facility to handle (b).

Terje

-- 
- <Terje.Mathisen at hda.hydro.com>
"almost all programming can be viewed as an exercise in caching"




More information about the questions mailing list