[ntp:questions] Using different timebase for ntpd

David Mills mills at udel.edu
Sun Mar 29 17:30:48 UTC 2009


Danny,

Thanks for the qick pic. Indeed, the full interleaved  mode is in the 
current development version at the backroom site, but might not yet be 
in the current snapshot. The web page cited is not quite up to date, as 
I am refining it offline to include in the next revision of my book. 
Meanwhile, the interleaved symmetric mode has been proposed for the 
Proximity-1 space link protocol to distribute time in the vicinity of 
Mars and the Moon. As space links operate at far lower data rates than 
Ethernets, interleaved mode is a necessity.

Interleaved broadcast mode with hardware derived timestamps and the NTP 
calibration round is functionally equivalent to IEEE 1588. However, 
there is a showstopper in that 1588 disciplines a good quality TCXO in 
the NIC clock, while NTP disciplines a rotten quality computer clock. 
1588 profices a mechanism to wrangle a herd of instrument and computer 
NICs to a Grandmaster, but not necessarily to wrangle the computer clock 
to that herd. This would be an interesting work item for TICTOC.

Dave

Danny Mayer wrote:

>Patrick Loschmidt wrote:
>  
>
>>Hi!
>>
>>Thanks a lot for your detailed answer!
>>
>>Greg Dowd schrieb:
>>    
>>
>>>Most if not all commercial ntp appliance manufacturers have some sort of
>>>hardware support, either RTOS, custom clock, better oscillator or even
>>>modified Ethernet devices to support hardware time stamping.
>>>      
>>>
>>Well I'm mostly interested in systems doing hardware time stamping on 
>>the Ethernet interface. My intention to just use another timebase are 
>>just an intermediate step to ease my development.
>>
>>    
>>
>>>I tried  modifying one
>>>of our SyncServers and gave a presentation at ietf tictoc bof on
>>>precision frequency transfer over packet based networks. Dave Mills also
>>>did some experimentation with this.  While I used mode 3/4 with a
>>>lagging timestamp for the followup function, Dave restricted that
>>>operation to symmetric modes. I have another presentation from a sync
>>>conference last year comparing the two protocols.  Let me know if you
>>>are interested and I'll dig them up.
>>>      
>>>
>>Yes, please, I'm very much interested, since I did an extensive search 
>>via IEEE Explore, ACM Library, google, etc. and couldn't really find any 
>>results or even products explicitly stating that they use hardware support.
>>
>>    
>>
>>>In your case, you are not talking about hardware timestamping but
>>>software timestamping using a hardware clock.  It is a good first step
>>>but os/stack jitter is still likely to be dominant.
>>>      
>>>
>>Thanks for all the further hints. I have to admit, that I'm working in 
>>the IEEE 1588 community for quite a while. Since I'm currently compiling 
>>my PhD, I asked myself the basic questions like, why didn't we just 
>>enhance NTP with HW timestamping to reach the accuracy?
>>So I came up with that idea, searched for implementations or scientific 
>>papers giving results, reasons, etc. but had not much luck.
>>
>>I found "A brief history of NTP time: Confessions of an Internet Time 
>>Keeper" by David Mills, mentioning the Nanokernel and some fancy 
>>synchronization results with 50ns RMS but I couldn't dig out the cited 
>>reference, since the cite is incorrect ... Also proceedings of the PTTI 
>>didn't help much. :-(
>>
>>So, what I still would like to have is either documentation of someone 
>>who already did NTP with hardware timestamps on Ethernet or can I 
>>somewhere find a documentation of the NTP-Daemon to kernel clock 
>>interface, in order to replace all the necessary system calls?
>>
>>Anyway, I'll dig through the code an all your hints. Thank's a lot for 
>>all the helpful answers in this newsgroup.
>>
>>Regards,
>>Patrick
>>    
>>
>
>There is work going on in the TICTOC working group of the IETF to define
>standards and protocols to take advantage of of the IEEE 1588-2 standard
>but it's early days yet. Dave Mills himself has come up with ideas of
>how NTP can take advantage of 1588 to improve NTP results. Don't forget
>that 1588 assumes a strict tree and you need 1588 everywhere and I
>believe it applies only to Ethernet. Dave Mills himself has some ideas
>on implementing this in NTP and has discussed this either here on in the
>hackers mailing list. Dave's research documents can be found on his NTP
>research web site: http://www.eecis.udel.edu/~mills/ntp.html and you
>will find a reference on the left hand side to "NTP Interleaved On-Wire
>Protocol" which is what I believe you are looking for.
>
>If you are finding some of the references are wrong, please ask Dave and
>he will correct them. In addition if you are interested in getting this
>implemented in NTP I am sure that Dave would be more than happy to take
>you on to get it implemented for general use. From the look of that
>reference he has implemented a simulator (lev.c) for testing scenarios
>as well as in the NTP reference implementation but I don't think that he
>has made that available yet.
>
>Dave will say more and is very approachable in these matters.
>
>Danny
>
>  
>




More information about the questions mailing list