RFC 8836: Congestion Control Requirements for Interactive Real-Time Media

The congestion control algorithm MUST attempt to provide as-low-as-possible-delay transit for interactive real-time traffic while still providing a useful amount of bandwidth. There may be lower limits on the amount of bandwidth that is useful, but this is largely application specific, and the application may be able to modify or remove flows in order to allow some useful flows to get enough bandwidth. For example, although there might not be enough bandwidth for low-latency video+audio, there could be enough for audio only.¶

1. Jitter (variation in the bitrate over short timescales) is also relevant, though moderate amounts of jitter will be absorbed by jitter buffers. Transit delay should be considered to track the short-term maximums of delay, including jitter.¶
2. The algorithm should provide this as-low-as-possible-delay transit and minimize self-induced latency even when faced with intermediate bottlenecks and competing flows. Competing flows may limit what's possible to achieve.¶
3. The algorithm should be resilient to the effects of events, such as routing changes, which may alter or remove bottlenecks or change the bandwidth available, especially if there is a reduction in available bandwidth or increase in observed delay. It is expected that the mechanism reacts quickly to such events to avoid delay buildup. In the context of this memo, a "quick" reaction is on the order of a few RTTs, subject to the constraints of the media codec, but is likely within a second. Reaction on the next RTT is explicitly not required, since many codecs cannot adapt their sending rate that quickly, but at the same time a response cannot be arbitrarily delayed.¶
4. The algorithm should react quickly to handle both local and remote interface changes (e.g., WLAN to 3G data) that may radically change the bandwidth available or bottlenecks, especially if there is a reduction in available bandwidth or an increase in bottleneck delay. It is assumed that an interface change can generate a notification to the algorithm.¶
5. The real-time interactive media applications can be rate limited. This means the offered loads can be less than the available bandwidth at any given moment and may vary dramatically over time, including dropping to no load and then resuming a high load, such as in a mute/unmute operation. Hence, the algorithm must be designed to handle such behavior from a media source or application. Note that the reaction time between a change in the bandwidth available from the algorithm and a change in the offered load is variable, and it may be different when increasing versus decreasing.¶
6. The algorithm is required to avoid building up queues when competing with short-term bursts of traffic (for example, traffic generated by web browsing), which can quickly saturate a local-bottleneck router or link but clear quickly. The algorithm should also react quickly to regain its previous share of the bandwidth when the local bottleneck or link is cleared.¶
7. Similarly, periodic bursty flows such as MPEG DASH [MPEG_DASH] or proprietary media streaming algorithms may compete in bursts with the algorithm and may not be adaptive within a burst. They are often layered on top of TCP but use TCP in a bursty manner that can interact poorly with competing flows during the bursts. The algorithm must not increase the already existing delay buildup during those bursts. Note that this competing traffic may be on a shared access link, or the traffic burst may cause a shift in the location of the bottleneck for the duration of the burst.¶

The algorithm MUST be fair to other flows, both real-time flows (such as other instances of itself) and TCP flows, both long-lived flows and bursts such as the traffic generated by a typical web-browsing session. Note that "fair" is a rather hard-to-define term. It SHOULD be fair with itself, giving a fair share of the bandwidth to multiple flows with similar RTTs, and if possible to multiple flows with different RTTs.¶

1. Existing flows at a bottleneck must also be fair to new flows to that bottleneck and must allow new flows to ramp up to a useful share of the bottleneck bandwidth as quickly as possible. A useful share will depend on the media types involved, total bandwidth available, and the user-experience requirements of a particular service. Note that relative RTTs may affect the rate at which new flows can ramp up to a reasonable share.¶

The algorithm SHOULD NOT starve competing TCP flows and SHOULD, as best as possible, avoid starvation by TCP flows.¶

1. The congestion control should prioritize achieving a useful share of the bandwidth depending on the media types and total available bandwidth over achieving as-low-as-possible transit delay, when these two requirements are in conflict.¶

The algorithm SHOULD adapt as quickly as possible to initial network conditions at the start of a flow. This SHOULD occur whether the initial bandwidth is above or below the bottleneck bandwidth.¶

1. The algorithm should allow different modes of adaptation; for example, the startup adaptation may be faster than adaptation later in a flow. It should allow for both slow-start operation (adapt up) and history-based startup (start at a point expected to be at or below channel bandwidth from historical information, which may need to adapt down quickly if the initial guess is wrong). Starting too low and/or adapting up too slowly can cause a critical point in a personal communication to be poor ("Hello!"). Starting too high above the available bandwidth causes other problems for user experience, so there's a tension here. Alternative methods to help startup, such as probing during setup with dummy data, may be useful in some applications; in some cases, there will be a considerable gap in time between flow creation and the initial flow of data. Again, a flow may need to change adaptation rates due to network conditions or changes in the provided flows (such as unmuting or sending data after a gap).¶

The algorithm SHOULD be stable if the RTP streams are halted or discontinuous (for example, when using Voice Activity Detection).¶

1. After stream resumption, the algorithm should attempt to rapidly regain its previous share of the bandwidth; the aggressiveness with which this is done will decay with the length of the pause.¶

Where possible, the algorithm SHOULD merge information across multiple RTP streams sent between two endpoints when those RTP streams share a common bottleneck, whether or not those streams are multiplexed onto the same ports. This will allow congestion control of the set of streams together instead of as multiple independent streams. It will also allow better overall bandwidth management, faster response to changing conditions, and fairer sharing of bandwidth with other network users.¶

1. The algorithm should also share information and adaptation with other non-RTP flows between the same endpoints, such as a WebRTC data channel [RFC8831], when possible.¶
2. When there are multiple streams across the same 5-tuple coordinating their bandwidth use and congestion control, the algorithm should allow the application to control the relative split of available bandwidth. The most correlated bandwidth usage would be with other flows on the same 5-tuple, but there may be use in coordinating measurement and control of the local link(s). Use of information about previous flows, especially on the same 5-tuple, may be useful input to the algorithm, especially regarding startup performance of a new flow.¶

The algorithm SHOULD NOT require any special support from network elements to be able to convey congestion-related information. As much as possible, it SHOULD leverage available information about the incoming flow to provide feedback to the sender. Examples of this information are the packet arrival times, acknowledgements and feedback, packet timestamps, packet losses, and Explicit Congestion Notification (ECN) [RFC3168]; all of these can provide information about the state of the path and any bottlenecks. However, the use of available information is algorithm dependent.¶

1. Extra information could be added to the packets to provide more detailed information on actual send times (as opposed to sampling times), but such information should not be required.¶

Since the assumption here is a set of RTP streams, the backchannel typically SHOULD be done via the RTP Control Protocol (RTCP) [RFC3550]; instead, one alternative would be to include it in a reverse-RTP channel using header extensions.¶

1. In order to react sufficiently quickly when using RTCP for a backchannel, an RTP profile such as RTP/AVPF [RFC4585] or RTP/SAVPF [RFC5124] that allows sufficiently frequent feedback must be used. Note that in some cases, backchannel messages may be delayed until the RTCP channel can be allocated enough bandwidth, even under AVPF rules. This may also imply negotiating a higher maximum percentage for RTCP data or allowing solutions to violate or modify the rules specified for AVPF.¶
2. Bandwidth for the feedback messages should be minimized using techniques such as those in [RFC5506], to allow RTCP without Sender/Receiver Reports.¶
3. Backchannel data should be minimized to avoid taking too much reverse-channel bandwidth (since this will often be used in a bidirectional set of flows). In areas of stability, backchannel data may be sent more infrequently so long as algorithm stability and fairness are maintained. When the channel is unstable or has not yet reached equilibrium after a change, backchannel feedback may be more frequent and use more reverse-channel bandwidth. This is an area with considerable flexibility of design, and different approaches to backchannel messages and frequency are expected to be evaluated.¶

Flows managed by this algorithm and flows competing against each other at a bottleneck may have different Differentiated Services Code Point (DSCP) [RFC5865] markings depending on the type of traffic or may be subject to flow-based QoS. A particular bottleneck or section of the network path may or may not honor DSCP markings. The algorithm SHOULD attempt to leverage DSCP markings when they're available.¶