An Analysis of Performance Statistics Reported by the NTPv4 Reference Implementation and Their Effect on Calibration Uncertainty

Since the early days of networked computer systems, the Network Time Protocol (NTP) has played a key role in keeping these devices synchronized to UTC. Today, many implementation of NTP exist, most of them capable of synchronizing machines over the Internet to within a few milliseconds, or to within microseconds over LANs. Some recent implementations can take advantage of hardware timestamping and, under ideal conditions, achieve synchronization to sub microsecond levels. Leveraging on NTP’s client-server model, one can use NTP not only to synchronize the local clock to an ensemble of sources, but also to estimate the offset between the local system clock and one or more remote NTP servers. The reference system’s NTP server is configured so that these remote NTP servers being calibrated do not participate in the disciplining of the local clock. This provides a simple and non-intrusive method for monitoring the performance of remote NTP servers and ultimately allows one to establish traceability of a remote NTP server to the SI (1). When this method is used to calibrate remote NTP servers over a LAN where round-trip delays are often in microseconds, uncertainties of the calibrator can become a major contributor to the overall uncertainty budget and should be minimized. By subjecting the NTPv4 reference implementation running on commodity hardware to various tests, we observed that the clock offset and round-trip delay values reported by NTP were affected by changes made to the NTP configuration of the reference system, even though these changes were not expected to have any influence on the disciplining of the local clock nor the measured remote NTP servers. In some cases, the observed change in reported values was as high as 6 µs and 35 µs for NTP offset and round-trip delay measurements over a LAN, respectively. In this presentation we will first present these observations and then compare with the values reported by another popular NTP implementation: chrony. A discussion of possible sources for these irregularities in the reported measurements will follow. References 1. Traceable Time Dissemination with NTP. Andre Charbonneau, Rob Douglas and Marina Gertsvolf. Reston, Virginia : s.n., 2018. 49th Annual Precise Time and Time Interval Systems and Applications Meeting. pp. 237 - 243.