[ntp:questions] Onboard Local Oscillator Change Improvements

Richard B. Gilbert rgilbert88 at comcast.net
Wed Oct 10 10:38:33 UTC 2007


elio.gi at gmail.com wrote:
> Hi ,
> 
> On 8 Ott, 22:06, "David L. Mills" <mi... at udel.edu> wrote:
> 
>>The performance with ntpd running in a rusting Pentium hulk with several
>>hundred clients and PPS is surprisingly good. Our rackety.udel.edu runs
>>FreeBSD 6.1 with PPS via the parallel port. It usually shows residual
>>offset and jitter in the order of a few microseconds. It is good enough
>>that the daily variations in the air-conditioned machine room
>>temperature can be measured accurately.
>>
>>Measuring the PPS offset within a few hundred nanoseconds is brilliant;
>>however, the measured latency to read the system clock from an
>>application is in the order of a microsecond and this polluted by
>>context switches, cache misses, timer interrupts, etc.
> 
> 
> I work for a big international telecom Company (Number 1 in Mobile)
> and I think that I have to give you more details of our Company
> project that could use NTP technology.
> In this moment we're making our tests with NTP on a test pc, but our
> aim is to run NTP on our telecom blades which have onboard their own
> processor and an embedded linux OS.
> These blades will get a PPS signal derived from a GPS and, thanks to
> ntpd daemon, will synchronize their local time.
> These "Stratum 1" blades will send to other (Stratum 2) blades NTP
> messages, so the question is:
> 
> - is the precision of these "messages" (or timestemps) affected by the
> stability of the local oscillator ?
> 
> The application will run in a MAN corporate network where packet
> delays should be small compared to the Internet delays. So it could be
> worth to enhance the accuracy of the timestemps provided by the ntp
> stratum1 server; than the second question is:
> 
> - which kind of accuracy can we reach enhancing the stability of the
> oscillator ? (nanoseconds? )
> (note that we already use high stability oscillator inside the
> equipment, that could also feed the processor clock)
> 
> The same question at client side is:
> 
> - which accuracy of local clock can we expect using high stability
> oscillators? (we have measured around 60 microseconds  peak-to-peak,
> using standard crystal in *environment test* configuration- see
> below )
<snip>

The accuracy on the clients depends on, more than anything else, the 
network connection between them and the server.  With a 100 MB LAN the 
delay and, therefore, the uncertainty should be on the order of 1 
millisecond.  The stability of the oscillator has very little to do with 
it; NTP disciplines the local clock and will compensate for any drift in 
the oscillator.

If you really need synchronization to less than 1 millisecond between 
clients, you need to consider something like a PPS signal sent to each 
client and compensating for speed of light delays in each length of cable!




More information about the questions mailing list