[time-nuts] Timing performance of servers

Magnus Danielson magnus at rubidium.dyndns.org
Sat Oct 27 09:56:24 EDT 2012


On 10/25/2012 07:03 AM, David J Taylor wrote:
> Magnus,
> If it helps, I have my own measurements of the Meinberg NTP port and
> later versions running on Windows here:
> http://www.satsignal.eu/mrtg/performance_ntp.php
> Strategy:
> 1 - have one FreeBSD (not Linux) server, although this is now not
> essential, but it's nice as a confirmation that the rest is working OK.
> 2 - Configure some Windows PCs as stratum-1 servers fed from GPS. On the
> plots above, PCs Alta, Bacchus, Feenix and Stamsund are acting as
> stratum-1 servers. These all have serial port connections, and cover the
> OS range Windows 2000, XP, Win-7/32 and Win-7/64. All are using the
> kernel-mode serial port driver patch developed by Dave Hart. PC Pixie is
> the FreeBSD box.
> 3 - For the client PCs, use a fixed 32-second polling interval to the
> local stratum-1 servers, with Internet servers as a backup polled at
> 1024 seconds, resulting in a configuration file something like:
> _______________________________________________
> # Use drift file
> driftfile "C:\Tools\NTP\etc\ntp.drift"
> # Use specific local NTP servers
> server iburst minpoll 5 maxpoll 5 prefer # Pixie
> server iburst minpoll 5 maxpoll 5 # Feenix
> server iburst minpoll 5 maxpoll 5 # Stamsund
> # Use pool NTP servers
> pool uk.pool.ntp.org iburst minpoll 10
> _______________________________________________
> The client performance varies, with some of the best results being on a
> Windows-8 Wi-Fi connected PC which seems to have very good drivers (PC
> Bergen). Jitter is 40 - 110 microseconds. Windows XP also shows low
> jitter, but greater offset (within 250 microseconds).
> Windows Vista was the worst performer I had, but that PC has now been
> retired. There are discussions in progress at the moment about improving
> Windows-Vista and Windows-7 as a Windows time interval setting and
> reporting bug has been discovered, particularly affecting NTP.

Lovely! I'm impressed.

What's the reasons for the offsets? Can't your tool handle negative values?

It would be good to have min, max, max-min, avg, std.dev values without 
offsets to help illustrate worst-case behaviour as well as average 
performance and noise "energy". The more advanced plotter would show 
MADEV, TDEV and MTIE plots. Ah well.

Would it be possible to set up so you could measure deviation on SNTP 
and undisciplined machines?

PS. Have my summerhouse not to far away from the town Ystad.


More information about the time-nuts mailing list