[ntp:questions] SNTP server + ntpd 4.2.4 client

David Woolley david at ex.djwhome.demon.co.uk.invalid
Mon Mar 17 13:19:40 UTC 2008


Noob wrote:

>> Offset doesn't tell you the accuracy, it only gives you an idea of the 
>> variability of the error.  Theoretically, the error could be as much 
>> as 15 to 25ms, plus the error from the stratum one to the stratum 2.
> 
> What metric should I consider to determine accuracy?

You cannot measure that using ntpd itself.  If you really need to know 
the accuracy, you must measure the true time with something with lower 
error bounds and compare.

The error should be somewhere within about +/- (rootdelay/2 + 
rootdispersion + jitter), if I remember correctly.  Most of the time it 
will be much better than this, but you need to know the limits of 
asymmetry in your network round trip time and the worst case frequency 
errors, if you want a better figure.

The basic problem is that, if ntpd knew the error, it would compensate 
for it and that part of the error would no longer be part of the error!

With your figures, I would guess that the rootdispersion contribution is 
low, so the real question is how much propagation delay asymmetry there is.

>> It's a violation of NTP, so the result will only be compliant as an 
>> SNTP client.
> 
> What is a violation of NTP?

NTP clients must use NTP servers, not SNTP ones.

> 
>> It may or may not work, depending on whether or not early w32time 
>> implementations conformed to SNTP.  Early versions of w32time didn't 
>> set enough of the response fields to sensible values to guarantee that 
>> ntpd would work as a client.
> 
> Why do you mention w32time?

Because w32time claims to be SNTP and I've seen cases where ntpd will 
refuse to synchronise to it because it is reporting error metrics that 
exceed maximums, even though it has got an upstream time source.  (If I 
remember correctly, it was reflecting those in the request.)

> 
> Do you think the HEOL-T101 is running Windows?

It's running SNTP and w32time is the most commonly encountered 
implementation claiming to be SNTP, these days, so w32time is a model of 
what can go wrong when using SNTP.

>> If it works, you can probably expect most of the errors to be within 
>> your box.
> 
> Does this mean it is reasonable to expect 100 µs accuracy?

For ethernet communicated time, one typically expects 1ms to 2ms. 
However, you need to know network loading (a properly dimensioned 
ethernet will have significant latency due to contention) and interrupt 
and process scheduling latencies on both sides.




More information about the questions mailing list