Improving packet synchronization in an NTP server

Abstract

A computer or dedicated client can use the Network Time Protocol (NTP) to synchronize an internal clock to a server that is synchronized by a 1 pulse-per-second (pps) signal from a national timing laboratory. Measuring an NTP server on a local area network can reveal timing synchronization errors and anomalies that are not nearly as likely to be recognized when NTP is utilized on a wide area network, where network delay asymmetry is the dominant source of uncertainty. We measured a commercially available NTP server by making rapid packet requests with UTC(NIST) as the common reference clock for both the server and client. Analysis of the results revealed repetitive synchronization errors in the packets transmitted by the server. Although these errors were too small for customers who deployed the server or for clients who accessed the server to detect during typical usage, improving packet synchronization would be beneficial for some applications. By collaborating with the server's manufacturer, the source of the problem was revealed, and firmware changes were made to reduce the uncertainty of the transmitted packets. This paper describes our measurements and how the results were used to improve the server's NTP packet synchronization.