Google

Kungsbron 2 Stockholm 11122 Sweden harald@alvestrand.no

This document is not intended for publication as an RFC. It gives the underpinning arguments for why the idea that RTP sessions and MIME top level types are related is a deeply broken paradigm, and that we need to get away from it. These arguments are solely the opinion of the listed author. 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.

The RTP universe of functionality can, for the purposes of this argument, be reduced to two components: The RTP wire protocol (consisting of the RTP packet format, the RTCP reporting format, and the handling rules for RTP sessions), and the SDP session description language. For the purposes of this argument, the SDP functionality for describing non-RTP sessions is ignored, as is the ability to negotiate RTP sessions by other means than SDP. This document argues that the RTP mechanisms of multiple RTP sessions make sense for a lot of purposes, but does NOT make sense for a mandated separation between different top-level MIME media types.

RTP, according to its own description an application layer framework component, is a suitable protocol for framing data that needs to travel across the network in a time-sensitive fashion, with the idea that it is going to be presented at the receiving end in a time sequence. Normally, the data (usually called "media") is streamed across the network at a rate approximately equal to the speed at which it is intended to be presented ("real time data"). Examples of data carried over RTP include: G.711 Audio - 64 Kbits/second, completely fixed bitrate GSM AMR Audio - 4.75 to 12.2 Kbits/second, variable bitrate OPUS audio compressed into near-incomprehensibility - 6 kbits/second, variable bitrate OPUS audio carrying high fidelity music - 500 kbits/second, variable bitrate QQVGA (160x120) video at 15 FPS in H.264 compression - 50 Kbits/second, variable bitrate, lots of schemes for error concealment and correction HD video at 1920x1080@60 in H.264 compression - 1.4 Mbits/second Real-time text (T.140) - very few bits/second DTMF tone signalling - very few bits/second Schemes designed to increase the reliability of data carried across RTP include: Forward error correction (FEC) Duplicated streams, codec-independent Duplicate sending of important information within the codec NAK-based resends signalled over RTCP Stream reset requests signalled over RTCP Some of these are only applicable to media types (in particular, "send me a new I-frame" doesn't make sense if you don't have I-frames). Others can be used with any type of data.

The network can apply various things to help the session data arrive according to policy: Capacity reservation for specific flows Priority queueing, sending certain types of data faster than others Filtering or blocking certain types of communication that the managers deem inappropriate The network can do these things in multiple ways, including so-called "deep packet inspection", but the most common techniques require being able to identify either the requested handling of the packets (DiffServ using DSCP codepoints) or recognizing the flow based on its 5-tuple (source and destination address and port + protocol), possibly correlating the 5-tuple with information carried to the router through some kind of management interface (either connected to the session setup protocol or managed via some other interface such as RSVP/IntServ), and behaving accordingly. All techniques have limitations; DSCP requires a certain trust in the endpoints using the codepoints for "deserving traffic"; deep packet inspection requires that packets be unencrypted, and stream control requires that 5-tuples be related back to their putative purpose either by heuristics or by being connected to management protocols.

An RTP session is defined in RFC 3550 section 3: "RTP Session: An association among a set of participants communicating with RTP. A participant may be involved in multiple RTP sessions at the same time. In a multimedia session, each medium is typically carried in a separate RTP session with its own RTCP packets unless the encoding itself multiplexes multiple media into a single data stream. A participant distinguishes multiple RTP sessions by reception of different sessions using different pairs of destination transport addresses, where a pair of transport addresses comprises one network address plus a pair of ports for RTP and RTCP. All participants in an RTP session may share a common destination transport address pair, as in the case of IP multicast, or the pairs may be different for each participant, as in the case of individual unicast network addresses and port pairs. In the unicast case, a participant may receive from all other participants in the session using the same pair of ports, or may use a distinct pair of ports for each. The distinguishing feature of an RTP session is that each maintains a full, separate space of SSRC identifiers (defined next). The set of participants included in one RTP session consists of those that can receive an SSRC identifier transmitted by any one of the participants either in RTP as the SSRC or a CSRC (also defined below) or in RTCP. For example, consider a three- party conference implemented using unicast UDP with each participant receiving from the other two on separate port pairs. If each participant sends RTCP feedback about data received from one other participant only back to that participant, then the conference is composed of three separate point-to-point RTP sessions. If each participant provides RTCP feedback about its reception of one other participant to both of the other participants, then the conference is composed of one multi-party RTP session. The latter case simulates the behavior that would occur with IP multicast communication among the three participants. The RTP framework allows the variations defined here, but a particular control protocol or application design will usually impose constraints on these variations." An RTP session is thus characterized by: A single SSRC space A single reporting space - all participants see all RTCP messages Non overlapping transport addresses As we can see here, it is not possible to tell from a single packet whether it belongs to the same session as another packet or not; if we observe two packets with the same source and destination addresses, it seems safe to assume that they belong to the same session, but for all other cases, deciding whether or not two packets or packet streams are in the same session requires knowledge of the configuration of the session.

Section 5.2 of RFC 3550 gives the canonical statement of RTP session (mis)use: "In RTP, multiplexing is provided by the destination transport address (network address and port number) which is different for each RTP session. For example, in a teleconference composed of audio and video media encoded separately, each medium SHOULD be carried in a separate RTP session with its own destination transport address." This sentence makes two very important leaps of faith: That distinguishing sessions by destination transport address is necessary and sufficient That it is appropriate to give strong guidance about the distribution of media streams across RTP sessions Both of these are shaky. As the cost of connecting ports has increased due to NATs, firewalls and IPv4 exhaustion, there has been a strong push towards using fewer ports, and indeed fewer 5-tuples, so that it is not uncommon to see flows that can be distinguished only by source address; there have also been proposals floated for putting multiple RTP sessions across one 5-tuple [draft-westerlund-avtcore-transport-multiplexing]. The cost of ports is also one factor pushing towards multiple media types in one RTP session; however, the more important underlying challenge is that this distinction is neither necessary nor sufficient to distinguish the cases in which RTP media streams want to have differential treatment from the network, and thus need to assign streams either to the same session (to guarantee the same treatment) or to different sessions (to allow for differential treatment). Consider the list of scenarios above, and imagine RTP being used for: A videoconference between 3 people who know each other well, using low end equipment and barely-sufficient bandwidth pipes A Berlin Philharmonic concert broadcast featuring Brahms' "Tragic Overture" A point-to-point transmission of a Manchester United vs Liverpool football match A professor's lecture, with "talking head" presentation simultaneous with slides, and opportunity for students to ask questions In each of these contexts, the tradeoff between audio and video is different; in the Brahms case, the audio (which is the point of the transmission) is likely to be transmitted at higher bandwidth than the video, and if one of them has to have his bandwidth reduced, the video should be reduced in quality before the audio is. In contrast, in the football match, spectators care about seeing the action; as long as they can understand the commentator's voice, the audio quality is "good enough". In the lecture case, quality of the lecturer's slides and voice is critical; video from students is almost irrelevant to the larger purpose. A logical arrangement of media streams in RTP sessions would be to group them by importance, and send them with appropriate traffic engineering structuring; in the lecturer case, the slides and the professor's voice would be carried in a high priority media stream, while the professor's picture would have second priority, and voice and video from students would be made available on an "if it works, it works" basis. Someone may easily decide that the student feedback track is not worth listening to, or remove the talking head of the professor; it would be strange indeed to try to listen to the lecture without viewing the supporting material. This illustrates two points: The RTP session mechanism, using the 5-tuple as the unit of differentiation, is a simple, effective and readily deployed mechanism for separating streams that require different treatment from the network in easily distinguished partitions. The assignment of media to such partitions is application dependent, and the decision on how to group and how to prioritize needs to be taken by the application developer.

In the list of reasons to argue against the inappropriate advice quoted above from RFC 3550, its pernicious influence on the MIME type system bears mentioning. The MIME type system, as described in , consists of a two-level hierarchy: A top level media type (text, audio, video, application and so on), and a media subtype that identifies (to some level of precision) the format of the data being carried. The system has mostly been respected, with some types (for instance PDF) forever being borderline between the various categories, but over the years, a few types have been entered into the system with their top level types being decided, not by the nature of their content, but by *the type with which their proponents wished to have them multiplexed in an RTP session*. This includes the types that designate repair mechanism (audio/parityfec, audio/red), timed data transfer (audio/clearmode) and that ultimate triumph of expediency over cleanliness: audio/t140c, audio/3gpp-tt and video/3gpp-tt: Text types registered as audio and video. For each of these, there is a fairly natural fit in the normal MIME hierarchy (application/ for the mechanism types and text/ for the text types); the assignment of them to the "media" top level types has been done as an expediency in order to get around the stultifying results of the advice given in RFC 3550.

One of the arguments in favour of the RFC 3550 separation has been that a mixer can be deployed that knows nothing of the semantics of the media streams; it can "just mix them". This applies partly to exactly one type of application: The audio conference bridge. For a video mixer, it does not apply; external logic (such as listening to the audio voume of the corresponding audio channel, or explicit flow control) is needed to select the right video stream to send out. And even for larger audio bridges, it is common to have functions like floor control, remote mute and other participant management tools in order to control the bridge - as soon as such tools are introduced, they are as relevant for a multi-media-type RTP session as they are to a single-media-type RTP session.

There are not many protocol changes that really need to be taken to solve this problem. The basic mechanism of RTP is media type independent. There are some RTCP issues with dealing with RTP flows of wildly varying bandwidth, but as can be seen from the table of media types in the introduction, this issue isn't solved by separating them; the bandwidth ranges of the types overlap. The thing that binds most in the current protocol suite is the conservation of the inappropriate binding in the SDP media description/negotiation format, where the MIME type is represented in two pieces, one of which is tied to the RTP session rather than to the payload type it is associated with, and there are fairly well-understood ways to get around that, such as the BUNDLE grouping extension . Better designed negotiation protocols would not have this problem at all. In order to get out of the bind that SDP places us in, a change such as BUNDLE should be adopted, and the IETF should record that the advice from RFC 3550 is to be considered *advice*, not command: It is sometimes appropriate to separate media streams according to top level type, and sometimes not appropriate to do so. The application is the one that needs to make this decision.

This document makes no request of IANA. Note to RFC Editor: this section may be removed on publication as an RFC.

This note does not discuss any change that the author thinks would have any significant influence on the security of RTP traffic.

This note has benefited greatly from exchanges with Colin Perkins, whose unwavering support of a sharply differing viewpoint has served to inform the arguments presented in this document. Magnus Westerlund and Christer Holmberg also deserve special mention for engaging constructively in the discussion.

Version number bump, since the debate is ongoing. A few nits fixed. Added the "Mixer Fallacy" section. Updated reference to "bundle" to new draft name. This should be the last version, since the author is in the process of working with the authors of to achieve a jointly agreeable text. Hopefully this will take lesss than 6 months.