In the NTP model a number of primary reference sources, synchronized by wire or radio to national standards, are connected to widely accessible resources, such as backbone gateways, and operated as primary time servers. The purpose of NTP is to convey timekeeping information from these servers to other time servers via the Internet and also to cross-check clocks and mitigate errors due to equipment or propagation failures. Some number of local-net hosts or gateways, acting as secondary time servers, run NTP with one or more of the primary servers. In order to reduce the protocol overhead, the secondary servers distribute time via NTP to the remaining local-net hosts. In the interest of reliability, selected hosts can be equipped with less accurate but less expensive radio clocks and used for backup in case of failure of the primary and/or secondary servers or communication paths between them.
Throughout this document a standard nomenclature has been adopted: the stability of a clock is how well it can maintain a constant frequency, the accuracy is how well its frequency and time compare with national standards and the precision is how precisely these quantities can be maintained within a particular timekeeping system. Unless indicated otherwise, the offset of two clocks is the time difference between them, while the skew is the frequency difference (first derivative of offset with time) between them. Real clocks exhibit some variation in skew (second derivative of offset with time), which is called drift; however, in this version of the specification the drift is assumed zero.
NTP is designed to produce three products: clock offset, roundtrip delay and dispersion, all of which are relative to a selected reference clock. Clock offset represents the amount to adjust the local clock to bring it into correspondence with the reference clock. Roundtrip delay provides the capability to launch a message to arrive at the reference clock at a specified time. Dispersion represents the maximum error of the local clock relative to the reference clock. Since most host time servers will synchronize via another peer time server, there are two components in each of these three products, those determined by the peer relative to the primary reference source of standard time and those measured by the host relative to the peer. Each of these components are maintained separately in the protocol in order to facilitate error control and management of the subnet itself. They provide not only precision measurements of offset and delay, but also definitive maximum error bounds, so that the user interface can determine not only the time, but the quality of the time as well.
There is no provision for peer discovery or virtual-circuit management in NTP. Data integrity is provided by the IP and UDP checksums. No flow-control or retransmission facilities are provided or necessary. Duplicate detection is inherent in the processing algorithms. The service can operate in a symmetric mode, in which servers and clients are indistinguishable, yet maintain a small amount of state information, or in client/server mode, in which servers need maintain no state other than that contained in the client request. A lightweight association-management capability, including dynamic reachability and variable poll-rate mechanisms, is included only to manage the state information and reduce resource requirements. Since only a single NTP message format is used, the protocol is easily implemented and can be used in a variety of solicited or unsolicited polling mechanisms.
It should be recognized that clock synchronization requires by its nature long periods and multiple comparisons in order to maintain accurate timekeeping. While only a few measurements are usually adequate to reliably determine local time to within a second or so, periods of many hours and dozens of measurements are required to resolve oscillator skew and maintain local time to the order of a millisecond. Thus, the accuracy achieved is directly dependent on the time taken to achieve it. Fortunately, the frequency of measurements can be quite low and almost always non-intrusive to normal net operations.