Duplication Delay Attribute in the Session Description Protocol

Inside an IP network, packet delivery may be interrupted due to failure of a physical link, interface, or device. To reduce the impact of such interruptions, some networks are built in a resilient manner, allowing for multiple alternative paths between two endpoints. However, if there is no resiliency in the network or the failure happens in a non-resilient part of the network, a temporary outage will occur (i.e., packets will get dropped). The outage will last until network reconvergence takes place (i.e., until connectivity is restored) around the failure. Typically, network reconvergence takes between tens and hundreds of milliseconds, depending on the size and features of the network. There are a number of network-reconvergence technologies available today, such as IP Fast Convergence, MPLS Traffic Engineering Fast Reroute, and Multicast Only Fast Reroute. These technologies can be augmented by different types of application-layer loss-repair methods such as Forward Error Correction (FEC), retransmission, temporal redundancy, and spatial redundancy to minimize (and sometimes totally eliminate) the impact of outages. Each combination has its distinct requirements in terms of bandwidth consumption and results in a different network complexity. Thus, a network operator has to carefully consider what combination to deploy for different parts of a network (e.g., core vs. edge). A detailed overview of network-convergence technologies and loss-repair methods is provided in . One of the loss-repair methods is temporal redundancy, also known as delayed duplication. A media sender using this method transmits an original source packet and transmits its duplicate after a certain delay following the original transmission. If a network outage hits the original transmission, the expectation is that the second transmission arrives at the receiver (with a high probability). Alternatively, the second transmission may be hit by an outage and so gets dropped, and the original transmission completes successfully. Also, both transmissions can arrive on the receiver side; in that case, the receiver (or the node that does the duplicate suppression) needs to identify the duplicate packets and discard them appropriately, thereby producing a duplicate-free stream. Delayed duplication can be used in a variety of multimedia applications where there is sufficient bandwidth for the duplicated traffic and the application can tolerate the introduced delay. However, it must be used with care, since it might easily result in a new series of denial-of-service attacks. Delayed duplication is harmful in cases where the primary cause of packet loss is congestion, rather than a network outage due to a temporary link or network element failure. Duplication should only be used by endpoints that want to protect against network failures; protection against congestion must be achieved through other means, as duplication will only make congestion worse. One particular use case for delayed duplication is to improve the reliability of real-time video feeds inside a core IP network where bandwidth is plentiful and maximum reliability (preferably zero loss) is desired . Compared to other redundancy approaches such as FEC and redundant data encoding (e.g., ), delayed duplication is easy to implement, since it does not require any special type of encoding or decoding. For duplicate suppression, the receiver has to be able to identify the identical packets. This is straightforward for media packets that carry one or more unique identifiers such as the sequence number field in the RTP header . In non-RTP applications, the receiver can use unique sequence numbers if available or other alternative approaches to compare the incoming packets and discard the duplicate ones. This specification introduces a new Session Description Protocol (SDP) attribute for applications/services using the delayed duplication method to indicate the relative delay for each additional duplication. The attribute is used with the duplication grouping semantics defined in . This specification does not explain how to select the duplication delay that a sender should use; the selection technique depends on the underlying network and the reconvergence technologies used inside such a network. This specification does not explain how the receiver should suppress the duplicate packets and merge the incoming streams to produce a loss-free and duplication-free output stream (a process commonly called "stream merging"), either. An application or a transport service that will use the delayed duplication method must determine its own rules about stream merging. In practice, more than two redundant streams are unlikely to be used, since the additional delay and increased overhead are not easily justified. However, we define the new attribute in a general way so that it could be used with more than two redundant streams (i.e., multiple duplications), if needed. While the primary focus in this specification is the RTP-based transport, the new attribute is applicable to both RTP and non-RTP streams. Protocol issues and details on duplicating RTP streams are presented in .

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in .

The following ABNF syntax formally describes the 'duplication-delay' attribute:

ABNF Syntax for the 'duplication-delay' Attribute The 'duplication-delay' attribute is defined as both a media-level and session-level attribute. It specifies the relative delay with respect to the previous transmission of each duplication in milliseconds (ms) at the time of transmission. The following rules apply: If used as a media-level attribute, it MUST be used with the 'ssrc-group' attribute and "DUP" grouping semantics as defined in . When used as a media-level attribute, the relative delay value(s) it specifies SHALL apply to every Synchronization Source (SSRC)-based duplication grouping in the same media description. In other words, one cannot specify different duplication delay values for different duplication groups in the same media description. If used as a session-level attribute, it MUST be used with 'group' attribute and "DUP" grouping semantics as defined in . When used as a session-level attribute, the relative delay value(s) it specifies SHALL apply to every duplication grouping in the same SDP description. In other words, one cannot specify different duplication delay values for different duplication groups in the same SDP description. If one needs to specify different duplication delay values for different duplication groups, then one MUST use different SDP descriptions for each or MUST use the 'duplication-delay' attribute at the media level. In that case, the 'duplication-delay' attribute MUST NOT be used at the session level. For offer/answer model considerations, refer to .

In the first example below, the multicast stream consists of two RTP streams, each duplicated once, resulting in two sets of two-stream groups. The same duplication delay of 100 ms is applied to each grouping. The first set's streams have SSRCs of 1000 and 1010, and the second set's streams have SSRCs of 1020 and 1030.

Note that in actual use, SSRC values, which are random 32-bit numbers, could be much larger than the ones shown in this example. In the second example below, the multicast stream is duplicated twice. 50 ms after the original transmission, the first duplicate is transmitted, and 100 ms after that, the second duplicate is transmitted. In other words, the same packet is transmitted three times over a period of 150 ms.

In the third example below, the multicast UDP stream is duplicated with a duplication delay of 50 ms. Redundant streams are sent in separate source-specific multicast (SSM) sessions, so the receiving host has to join both SSM sessions if it wants to receive both streams.

The 'duplication-delay' attribute is not believed to introduce any significant security risk to multimedia applications. A malevolent third party could use this attribute to misguide the receiver(s) about the duplication delays and/or the number of redundant streams. For example, if the malevolent third party increases the value of the duplication delay, the receiver(s) will unnecessarily incur a longer delay, since they will have to wait for the entire period. Or, if the duplication delay is reduced by the malevolent third party, the receiver(s) might not wait long enough for the duplicated transmission and incur unnecessary packet losses. However, these require intercepting and rewriting the packets carrying the SDP description; if an interceptor can do that, many more attacks are also possible. In order to avoid attacks of this sort, the SDP description needs to be integrity protected and provided with source authentication. This can, for example, be achieved on an end-to-end basis using S/MIME when SDP is used in a signaling packet using MIME types (application/sdp). Alternatively, HTTPS or the authentication method in the Session Announcement Protocol (SAP) could be used as well. Another security risk is due to possible software misconfiguration or a software bug where a large number of duplicates could be unwillingly signaled in the 'duplication-delay' attribute. Similarly, an attacker can use this attribute to start a denial-of-service attack by signaling and sending too many duplicated streams. In applications where this attribute is to be used, it is a good practice to put a hard limit on both the number of duplicate streams and the total delay introduced due to duplication, regardless of what the SDP description specifies. Since this mechanism causes duplication of media packets, if those packets are also cryptographically protected (e.g., encrypted) then such duplication could act as an accelerator if any Million Message or similar attack such as Lucky 13 exists against the security mechanism that is in use. Such acceleration could turn an otherwise infeasible attack into one that is practical; however, assuming that the amount of duplication is small and that such weak or broken security mechanisms should really not be used, the overall security impact of the duplication should be minimal. If, however, a bad actor were in control of the SDP but did not have access to the keying material used for media, then such a bad actor could potentially use the SDP to cause the media handling to use a weak or broken mechanism with a lot of duplication, in which case the duplication could be significant. Deployments where the SDP is controlled by an actor who should not have access to the media keying material should therefore be cautious in their use of this duplication mechanism. If this mechanism were used in conjunction with a source description (SDES) and if the key being used for media protection is derived from a human-memorable or otherwise dictionary-attackable secret, then the duplication done here could allow for a more efficient dictionary attack against the media. The right countermeasure is to use proper keying or, if using an SDES, to ensure that the keys used are not dictionary-attackable.

The following contact information shall be used for the registration in this document:

This document registers a new attribute name in SDP.

The authors would like to thank Colin Perkins, Paul Kyzivat, and Stephen Farrell for their suggestions and reviews.