RTP and Leap Seconds

In some media networking applications, RTP streams are referenced to a wall-clock time (absolute date and time). This is accomplished through use of the NTP timestamp field in the sender report (SR) to create a mapping between RTP timestamps and the wall clock. When a wall-clock reference is used, the playout time for RTP packets is referenced to the wall clock. Smooth and continuous media playout requires a smooth and continuous time base. The time base used by the wall clock may include leap seconds that are not rendered smoothly. This document updates RFC 3550 by providing recommendations for smoothly rendering streamed media referenced to common wall clocks that do not have smooth or continuous behavior in the presence of leap seconds.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 and indicate requirement levels for compliant implementations.

The world's scientific time standard is International Atomic Time (TAI), which is based on vibrations of cesium atoms in an atomic clock. The world's civil time is based on the rotation of the Earth. In 1972, the civil time standard, Coordinated Universal Time (UTC), was redefined in terms of TAI and the concept of leap seconds was introduced to allow UTC to remain synchronized with the rotation of the Earth. Leap seconds are scheduled by the International Earth Rotation and Reference Systems Service. Leap seconds may be scheduled at the last day of any month but are preferentially scheduled for December and June and secondarily March and September . Because Earth's rotation is unpredictable, leap seconds are typically not scheduled more than six months in advance. Leap seconds do not respect local time and always occur at the end of the UTC day. Leap seconds can be scheduled to either add or remove a second from the day. A leap second that adds an extra second is known as a positive leap second. A leap second that skips a second is known as a negative leap second. Since their introduction in 1972, all leap seconds have been scheduled in June or December, and they have all been positive. NOTE: The ITU is studying a proposal that could eventually eliminate leap seconds from UTC. As of January 2012, this proposal is expected to be decided no earlier than 2015 .

UTC clocks feature a 61st second at the end of the day when a positive leap second is scheduled. The leap second is designated "23h 59m 60s".

Under NTP , a leap second is inserted at the beginning of the last second of the day. This results in the clock freezing or slowing for one second immediately prior to the last second of the affected day. This results in the last second of the day having a real-time duration of two seconds. Timestamp accuracy is compromised during this period because the clock's rate is not well defined.

The POSIX (Portable Operating System Interface) standard requires that leap seconds be omitted from reported time. All days are defined as having 86,400 seconds, but the timebase is defined to be UTC, a leap-second-bearing reference. Implementors of POSIX systems are offered considerable latitude by the standard as to how to map POSIX time to UTC. In many systems, leap seconds are accommodated by repeating the last second of the day. A timestamp within the last second of the day is therefore ambiguous in that it can refer to a moment in time in either of the last two seconds of a day containing a leap second. Other systems use the same technique used by NTP, freezing or slowing for one second immediately prior to the last second of the affected day. In some cases, leap seconds are accommodated by warping time ; that is, the length of the second in the vicinity of a leap second is slightly altered.

illustrates the positive leap second that occurred June 30, 2012 when the offset between TAI and UTC changed from 34 to 35 seconds. The first column shows RTP timestamps for an 8 kHz audio stream. The second column shows the TAI reference. The following columns show behavior for the leap-second-bearing wall clocks described above. Time values are shown at half-second intervals. RTP TAI UTC POSIX NTP 8000 00:00:32.500 23:59:58.500 23:59:58.500 23:59:58.500 12000 00:00:33.000 23:59:59.000 23:59:59.000 23:59:59.000 16000 00:00:33.500 23:59:59.500 23:59:59.500 23:59:59.500 20000 00:00:34.000 23:59:60.000 23:59:59.000 00:00:00.000 24000 00:00:34.500 23:59:60.500 23:59:59.500 00:00:00.000 28000 00:00:35.000 00:00:00.000 00:00:00.000 00:00:00.000 32000 00:00:35.500 00:00:00.500 00:00:00.500 00:00:00.500 NOTE: Some NTP implementations do not entirely freeze the clock while the leap second is inserted. Successive calls to retrieve system time return infinitesimally larger (e.g., 1 microsecond or 1 nanosecond larger) time values. This behavior is designed to satisfy assumptions applications may make that time increases monotonically. This behavior occurs in the least-significant bits of the time value and so is not typically visible in the human-readable format shown in the table. NOTE: POSIX implementations vary. The implementation shown here repeats the last second of the affected day. Other implementations mirror NTP behavior or alter the length of a second in the vicinity of the leap second.

Timestamps generated during a leap second may be ambiguous or interpreted differently by a sender and receiver or interpreted differently by different receivers. Without prior knowledge of the leap-second schedule, NTP servers and clients may become offset by exactly one second with respect to their UTC reference. This potential discrepancy begins when a leap second occurs and ends when all participants receive a time update from a server or peer. Depending on the system implementation, the offset can last anywhere from a few seconds to a few days. A long-lived discrepancy can be particularly disruptive to operation of NTP-referenced RTP streams. These discrepancies, depending on direction, may cause receivers to think they are receiving RTP packets after they should be played or to attempt to buffer received data an additional second before playing it. Either situation can cause an interruption in playback. Some receivers may automatically recognize an unexpected offset and resynchronize to the stream to accommodate it. Once the offset is resolved, such receivers may need to resynchronize again.

Senders and receivers that are not referenced to a wall clock are not affected by issues associated with leap seconds, and no special accommodation is required. RTP implementation using a wall-clock reference is simplified by using a clock with a timescale that does not include leap seconds. IEEE 1588 , GPS , and other systems that use a TAI reference do not include leap seconds. NTP time, operating system clocks, and other systems using a UTC reference include leap seconds. Note that some TAI-based systems such as IEEE 1588 and GPS, in addition to the TAI reference clock, deliver TAI to UTC mapping information. By combining the delivered TAI reference clock and the mapping information, some receivers of these systems are able to synthesize a leap-second-bearing UTC reference clock. For the purposes of this document, it is important to recognize that it is the timescale used, not the delivery mechanism that determines whether a reference clock is leap-second bearing. Reference clock type Examples Accommodation None Self clocking None needed Non-leap-second-bearing IEEE 1588, GPS, TAI None needed Leap-second-bearing NTP Recommended All participants generating or consuming timestamps associated with a leap-second-bearing reference MUST recognize leap seconds and SHOULD have a working communications channel to receive notifications of leap-second scheduling. A working communication channel includes a protocol means of notifying clocks of an impending leap second such as the Leap Indicator in the NTP header and also a means for top-tier clocks to receive leap-second schedule information published by the International Earth Rotation and Reference Systems Service . Such a communications channel may not be available on all networks. For security or other reasons, leap-second schedules may be configured manually for some networks or clocks. When a device does not reliably receive leap-second scheduling information, failures as described in may occur. Because of the timestamp ambiguity that positive leap seconds can introduce and the inconsistent manner in which different systems accommodate positive leap seconds, generating or using NTP timestamps during the entire last second of a day on which a positive leap second has been scheduled SHOULD be avoided. Note that the period to be avoided has a real-time duration of two seconds. In the example, the region to be avoided is indicated by RTP timestamps 12000 through 28000 Negative leap seconds do not introduce timestamp ambiguity or other complications. No special treatment is needed to avoid ambiguity with respect to RTP timestamps in the presence of a negative leap second. POSIX clocks that use a warping technique to accommodate leap seconds (e.g., ) are not a good choice for an interoperable timestamp reference and SHOULD not be used to timestamp RTP streams.

In order to avoid generating or using NTP timestamps during positive leap seconds, RTP senders and receivers need to avoid sending or using sender reports to synchronize their clocks in the vicinity of a leap second and instead rely on their internal clocks to maintain synchronization until the leap second has passed. RTP Senders using a leap-second-bearing reference for timestamps SHOULD NOT generate sender reports containing an originating NTP timestamp in the vicinity of a positive leap second. To maintain a consistent RTCP schedule and avoid the risk of unintentional timeouts, such senders MAY send receiver reports in place of sender reports in the vicinity of the leap second. For the purpose of suspending sender reports in the vicinity of a leap second, senders MAY assume that a positive leap second occurs at the end of the last day of every month. Receivers consuming leap-second-bearing timestamps SHOULD ignore timestamps in any sender reports generated in the vicinity of a positive leap second. For the purpose of ignoring sender reports in the vicinity of a leap second, receivers MAY assume that a positive leap second occurs at the end of the last day of every month.

Receivers consuming leap-second-bearing timestamps SHOULD take both positive and negative leap seconds in the reference into account to determine the playout time based on RTP timestamps for data in RTP packets.

RTP streams using a wall-clock reference as discussed here present an additional attack vector compared to self-clocking streams. Manipulation of the wall clock at either the sender or receiver can potentially disrupt streaming. For an RTP stream operating to a leap-second-bearing reference to operate reliably across a leap second, the sender and receiver must both be aware of the leap second. It is possible to disrupt a stream by blocking or delaying leap second notification to one of the participants. Streaming can be similarly affected if one of the participants can be tricked into believing a leap second has been scheduled where there is not one. These vulnerabilities are present in RFC 3550 and these new recommendations neither heighten nor diminish them. Integrity of the leap-second schedule is the responsibility of the operating system and time distribution mechanism, both of which are outside the scope of RFC 3550 and these recommendations.

The authors would like to thank Steve Allen for his valuable comments that helped to improve this document.