Accuracy of NTP disciplined across Geant

Following graph represents data which we collected from Feb 9 to Feb 12. As a primary NTP server we used karlovy.hea.net (located in Ireland), the site of measurement was box deneb.cesnet.cz in Cesnet LAN (Czech republic) - it is no-name Linux PC (Celeron 1 GHz) with standard quartz oscillator located in office room with no air-condition. Round-trip delay (ping) between both boxes is about 38 ms.

Figure 1: Offset of local time

Description of experiment

Site of the measurement was provided by PPS pulses (for measurement purpose only, not for synchronization). The red line (PPS offset) in the graph represents the offset between PPS time and the local time of the box. Green points (NTP offset) represent the value of item 'offset' given by the result of the command 'ntpq -c rl'. Total duration of measurement was 4 days. The accuracy of PPS offset measurement is better then 1 microsecond (we used special card for capturing PPS edge, without this card would be the accuracy better then 10 microsecond).

Comments

Our PoP lost international connectivity twice during the measurement, in both cases for about 2 hours. This situation visible in the graph as a constant value of NTP offset (ntpq reported permanently the last known value of the offset). It was between minutes 1150 and 1300 resp. 4350 and 4450. Despite this NTP source losing, secondary NTP box was able to maintain absolute offset (measured against GPS time) between -1.5 ms and +1 ms during the whole period of the experiment. The offset, which is declared by ntpd process, very well correlate with the exact offset. There is nearly constant difference 0.5 ms which is probably the difference of OWDs between Heanet and Cesnet. This important result has to be confirmed by measurement at other sites.