RFC 8836




Internet Engineering Task Force (IETF)                          R. Jesup
Request for Comments: 8836                                       Mozilla
Category: Informational                                   Z. Sarker, Ed.
ISSN: 2070-1721                                              Ericsson AB
                                                            January 2021


    Congestion Control Requirements for Interactive Real-Time Media

Abstract



   Congestion control is needed for all data transported across the
   Internet, in order to promote fair usage and prevent congestion
   collapse.  The requirements for interactive, point-to-point real-time
   multimedia, which needs low-delay, semi-reliable data delivery, are
   different from the requirements for bulk transfer like FTP or bursty
   transfers like web pages.  Due to an increasing amount of RTP-based
   real-time media traffic on the Internet (e.g., with the introduction
   of the Web Real-Time Communication (WebRTC)), it is especially
   important to ensure that this kind of traffic is congestion
   controlled.

   This document describes a set of requirements that can be used to
   evaluate other congestion control mechanisms in order to figure out
   their fitness for this purpose, and in particular to provide a set of
   possible requirements for a real-time media congestion avoidance
   technique.

Status of This Memo



   This document is not an Internet Standards Track specification; it is
   published for informational purposes.

   This document is a product of the Internet Engineering Task Force
   (IETF).  It represents the consensus of the IETF community.  It has
   received public review and has been approved for publication by the
   Internet Engineering Steering Group (IESG).  Not all documents
   approved by the IESG are candidates for any level of Internet
   Standard; see Section 2 of RFC 7841.

   Information about the current status of this document, any errata,
   and how to provide feedback on it may be obtained at
   https://www.rfc-editor.org/info/rfc8836.

Copyright Notice



   Copyright (c) 2021 IETF Trust and the persons identified as the
   document authors.  All rights reserved.

   This document is subject to BCP 78 and the IETF Trust's Legal
   Provisions Relating to IETF Documents
   (https://trustee.ietf.org/license-info) in effect on the date of
   publication of this document.  Please review these documents
   carefully, as they describe your rights and restrictions with respect
   to this document.  Code Components extracted from this document must
   include Simplified BSD License text as described in Section 4.e of
   the Trust Legal Provisions and are provided without warranty as
   described in the Simplified BSD License.

Table of Contents



   1.  Introduction
     1.1.  Requirements Language
   2.  Requirements
   3.  Deficiencies of Existing Mechanisms
   4.  IANA Considerations
   5.  Security Considerations
   6.  References
     6.1.  Normative References
     6.2.  Informative References
   Acknowledgements

   Authors' Addresses



1.  Introduction



   Most of today's TCP congestion control schemes were developed with a
   focus on a use of the Internet for reliable bulk transfer of non-
   time-critical data, such as transfer of large files.  They have also
   been used successfully to govern the reliable transfer of smaller
   chunks of data in as short a time as possible, such as when fetching
   web pages.

   These algorithms have also been used for transfer of media streams
   that are viewed in a non-interactive manner, such as "streaming"
   video, where having the data ready when the viewer wants it is
   important, but the exact timing of the delivery is not.

   When handling real-time interactive media, the requirements are
   different.  One needs to provide the data continuously, within a very
   limited time window (no more delay than hundreds of milliseconds end-
   to-end).  In addition, the sources of data may be able to adapt the
   amount of data that needs sending within fairly wide margins, but
   they can be rate limited by the application -- even not always having
   data to send.  They may tolerate some amount of packet loss, but
   since the data is generated in real time, sending "future" data is
   impossible, and since it's consumed in real time, data delivered late
   is commonly useless.

   While the requirements for real-time interactive media differ from
   the requirements for the other flow types, these other flow types
   will be present in the network.  The congestion control algorithm for
   real-time interactive media must work properly when these other flow
   types are present as cross traffic on the network.

   One particular protocol portfolio being developed for this use case
   is WebRTC [RFC8825], where one envisions sending multiple flows using
   the Real-time Transport Protocol (RTP) [RFC3550] between two peers,
   in conjunction with data flows, all at the same time, without having
   special arrangements with the intervening service providers.  As RTP
   does not provide any congestion control mechanism, a set of circuit
   breakers, such as those described in [RFC8083], are required to
   protect the network from excessive congestion caused by non-
   congestion-controlled flows.  When the real-time interactive media is
   congestion controlled, it is recommended that the congestion control
   mechanism operate within the constraints defined by these circuit
   breakers when a circuit breaker is present and that it should not
   cause congestion collapse when a circuit breaker is not implemented.

   Given that this use case is the focus of this document, use cases
   involving non-interactive media such as video streaming and those
   using multicast/broadcast-type technologies, are out of scope.

   The terminology defined in [RFC8825] is used in this memo.

1.1.  Requirements Language



   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 BCP 14 [RFC2119].

2.  Requirements



   1.   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.

        a.  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.

        b.  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.

        c.  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.

        d.  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.

        e.  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.

        f.  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.

        g.  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.

   2.   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.

        a.  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.

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

        a.  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.

   4.   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.

        a.  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).

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

        a.  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.

   6.   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.

        a.  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.

        b.  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.

   7.   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.

        a.  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.

   8.   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.

        a.  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.

        b.  Bandwidth for the feedback messages should be minimized
            using techniques such as those in [RFC5506], to allow RTCP
            without Sender/Receiver Reports.

        c.  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.

   9.   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.

   10.  The algorithm SHOULD sense the unexpected lack of backchannel
        information as a possible indication of a channel-overuse
        problem and react accordingly to avoid burst events causing a
        congestion collapse.

   11.  The algorithm SHOULD be stable and maintain low delay when faced
        with Active Queue Management (AQM) algorithms.  Also note that
        these algorithms may apply across multiple queues in the
        bottleneck or to a single queue.

3.  Deficiencies of Existing Mechanisms



   Among the existing congestion control mechanisms, TCP Friendly Rate
   Control (TFRC) [RFC5348] is the one that claims to be suitable for
   real-time interactive media.  TFRC is an equation-based congestion
   control mechanism that provides a reasonably fair share of bandwidth
   when competing with TCP flows and offers much lower throughput
   variations than TCP.  This is achieved by a slower response to the
   available bandwidth change than TCP.  TFRC is designed to perform
   best with applications that have a fixed packet size and do not have
   a fixed period between sending packets.

   TFRC detects loss events and reacts to congestion-caused loss by
   reducing its sending rate.  It allows applications to increase the
   sending rate until loss is observed in the flows.  As noted in IAB/
   IRTF report [RFC7295], large buffers are available in the network
   elements, which introduce additional delay in the communication.  It
   becomes important to take all possible congestion indications into
   consideration.  Looking at the current Internet deployment, TFRC's
   biggest deficiency is that it only considers loss events as a
   congestion indication.

   A typical real-time interactive communication includes live-encoded
   audio and video flow(s).  In such a communication scenario, an audio
   source typically needs a fixed interval between packets and needs to
   vary the segment size of the packets instead of the packet rate in
   response to congestion; therefore, it sends smaller packets.  A
   variant of TFRC, Small-Packet TFRC (TFRC-SP) [RFC4828], addresses the
   issues related to such kind of sources.  A video source generally
   varies video frame sizes, can produce large frames that need to be
   further fragmented to fit into path Maximum Transmission Unit (MTU)
   size, and has an almost fixed interval between producing frames under
   a certain frame rate.  TFRC is known to be less optimal when using
   such video sources.

   There are also some mismatches between TFRC's design assumptions and
   how the media sources in a typical real-time interactive application
   work.  TFRC is designed to maintain a smooth sending rate; however,
   media sources can change rates in steps for both rate increase and
   rate decrease.  TFRC can operate in two modes: i) bytes per second
   and ii) packets per second, where typical real-time interactive media
   sources operate on bit per second.  There are also limitations on how
   quickly the media sources can adapt to specific sending rates.
   Modern video encoders can operate in a mode in which they can vary
   the output bitrate a lot depending on the way they are configured,
   the current scene they are encoding, and more.  Therefore, it is
   possible that the video source will not always output at an allowable
   bitrate.  TFRC tries to increase its sending rate when transmitting
   at the maximum allowed rate, and it increases only twice the current
   transmission rate; hence, it may create issues when the video sources
   vary their bitrates.

   Moreover, there are a number of studies on TFRC that show its
   limitations, including TFRC's unfairness to low statistically
   multiplexed links, oscillatory behavior, performance issues in highly
   dynamic loss-rate conditions, and more [CH09].

   Looking at all these deficiencies, it can be concluded that the
   requirements for a congestion control mechanism for real-time
   interactive media cannot be met by TFRC as defined in the standard.

4.  IANA Considerations



   This document has no IANA actions.

5.  Security Considerations



   An attacker with the ability to delete, delay, or insert messages
   into the flow can fake congestion signals, unless they are passed on
   a tamper-proof path.  Since some possible algorithms depend on the
   timing of packet arrival, even a traditional, protected channel does
   not fully mitigate such attacks.

   An attack that reduces bandwidth is not necessarily significant,
   since an on-path attacker could break the connection by discarding
   all packets.  Attacks that increase the perceived available bandwidth
   are conceivable and need to be evaluated.  Such attacks could result
   in starvation of competing flows and permit amplification attacks.

   Algorithm designers should consider the possibility of malicious on-
   path attackers.

6.  References



6.1.  Normative References



   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119,
              DOI 10.17487/RFC2119, March 1997,
              <https://www.rfc-editor.org/info/rfc2119>.

   [RFC3550]  Schulzrinne, H., Casner, S., Frederick, R., and V.
              Jacobson, "RTP: A Transport Protocol for Real-Time
              Applications", STD 64, RFC 3550, DOI 10.17487/RFC3550,
              July 2003, <https://www.rfc-editor.org/info/rfc3550>.

   [RFC4585]  Ott, J., Wenger, S., Sato, N., Burmeister, C., and J. Rey,
              "Extended RTP Profile for Real-time Transport Control
              Protocol (RTCP)-Based Feedback (RTP/AVPF)", RFC 4585,
              DOI 10.17487/RFC4585, July 2006,
              <https://www.rfc-editor.org/info/rfc4585>.

   [RFC5124]  Ott, J. and E. Carrara, "Extended Secure RTP Profile for
              Real-time Transport Control Protocol (RTCP)-Based Feedback
              (RTP/SAVPF)", RFC 5124, DOI 10.17487/RFC5124, February
              2008, <https://www.rfc-editor.org/info/rfc5124>.

   [RFC8825]  Alvestrand, H., "Overview: Real-Time Protocols for
              Browser-Based Applications", RFC 8825,
              DOI 10.17487/RFC8825, January 2021,
              <https://www.rfc-editor.org/info/rfc8825>.

6.2.  Informative References



   [CH09]     Choi, S. and M. Handley, "Designing TCP-Friendly Window-
              based Congestion Control for Real-time Multimedia
              Applications", Proceedings of PFLDNeT, May 2009.

   [MPEG_DASH]
              ISO, "Information Technology -- Dynamic adaptive streaming
              over HTTP (DASH) -- Part 1: Media presentation description
              and segment formats", ISO/IEC 23009-1:2019, December 2019,
              <https://www.iso.org/standard/79329.html>.

   [RFC3168]  Ramakrishnan, K., Floyd, S., and D. Black, "The Addition
              of Explicit Congestion Notification (ECN) to IP",
              RFC 3168, DOI 10.17487/RFC3168, September 2001,
              <https://www.rfc-editor.org/info/rfc3168>.

   [RFC4828]  Floyd, S. and E. Kohler, "TCP Friendly Rate Control
              (TFRC): The Small-Packet (SP) Variant", RFC 4828,
              DOI 10.17487/RFC4828, April 2007,
              <https://www.rfc-editor.org/info/rfc4828>.

   [RFC5348]  Floyd, S., Handley, M., Padhye, J., and J. Widmer, "TCP
              Friendly Rate Control (TFRC): Protocol Specification",
              RFC 5348, DOI 10.17487/RFC5348, September 2008,
              <https://www.rfc-editor.org/info/rfc5348>.

   [RFC5506]  Johansson, I. and M. Westerlund, "Support for Reduced-Size
              Real-Time Transport Control Protocol (RTCP): Opportunities
              and Consequences", RFC 5506, DOI 10.17487/RFC5506, April
              2009, <https://www.rfc-editor.org/info/rfc5506>.

   [RFC5865]  Baker, F., Polk, J., and M. Dolly, "A Differentiated
              Services Code Point (DSCP) for Capacity-Admitted Traffic",
              RFC 5865, DOI 10.17487/RFC5865, May 2010,
              <https://www.rfc-editor.org/info/rfc5865>.

   [RFC7295]  Tschofenig, H., Eggert, L., and Z. Sarker, "Report from
              the IAB/IRTF Workshop on Congestion Control for
              Interactive Real-Time Communication", RFC 7295,
              DOI 10.17487/RFC7295, July 2014,
              <https://www.rfc-editor.org/info/rfc7295>.

   [RFC8083]  Perkins, C. and V. Singh, "Multimedia Congestion Control:
              Circuit Breakers for Unicast RTP Sessions", RFC 8083,
              DOI 10.17487/RFC8083, March 2017,
              <https://www.rfc-editor.org/info/rfc8083>.

   [RFC8831]  Jesup, R., Loreto, S., and M. Tüxen, "WebRTC Data
              Channels", RFC 8831, DOI 10.17487/RFC8831, January 2021,
              <https://www.rfc-editor.org/info/rfc8831>.

Acknowledgements



   This document is the result of discussions in various fora of the
   WebRTC effort, in particular on the <rtp-congestion@alvestrand.no>
   mailing list.  Many people contributed their thoughts to this.

Authors' Addresses



   Randell Jesup
   Mozilla
   United States of America

   Email: randell-ietf@jesup.org


   Zaheduzzaman Sarker (editor)
   Ericsson AB
   Torshamnsgatan 23
   SE-164 83 Stockholm
   Sweden

   Phone: +46 10 717 37 43
   Email: zaheduzzaman.sarker@ericsson.com