RFC 9330




Internet Engineering Task Force (IETF)                   B. Briscoe, Ed.
Request for Comments: 9330                                   Independent
Category: Informational                                   K. De Schepper
ISSN: 2070-1721                                          Nokia Bell Labs
                                                              M. Bagnulo
                                        Universidad Carlos III de Madrid
                                                                G. White
                                                               CableLabs
                                                            January 2023


Low Latency, Low Loss, and Scalable Throughput (L4S) Internet Service:
                              Architecture

Abstract



   This document describes the L4S architecture, which enables Internet
   applications to achieve low queuing latency, low congestion loss, and
   scalable throughput control.  L4S is based on the insight that the
   root cause of queuing delay is in the capacity-seeking congestion
   controllers of senders, not in the queue itself.  With the L4S
   architecture, all Internet applications could (but do not have to)
   transition away from congestion control algorithms that cause
   substantial queuing delay and instead adopt a new class of congestion
   controls that can seek capacity with very little queuing.  These are
   aided by a modified form of Explicit Congestion Notification (ECN)
   from the network.  With this new architecture, applications can have
   both low latency and high throughput.

   The architecture primarily concerns incremental deployment.  It
   defines mechanisms that allow the new class of L4S congestion
   controls to coexist with 'Classic' congestion controls in a shared
   network.  The aim is for L4S latency and throughput to be usually
   much better (and rarely worse) while typically not impacting Classic
   performance.

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/rfc9330.

Copyright Notice



   Copyright (c) 2023 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 Revised BSD License text as described in Section 4.e of the
   Trust Legal Provisions and are provided without warranty as described
   in the Revised BSD License.

Table of Contents



   1.  Introduction
     1.1.  Document Roadmap
   2.  L4S Architecture Overview
   3.  Terminology
   4.  L4S Architecture Components
     4.1.  Protocol Mechanisms
     4.2.  Network Components
     4.3.  Host Mechanisms
   5.  Rationale
     5.1.  Why These Primary Components?
     5.2.  What L4S Adds to Existing Approaches
   6.  Applicability
     6.1.  Applications
     6.2.  Use Cases
     6.3.  Applicability with Specific Link Technologies
     6.4.  Deployment Considerations
       6.4.1.  Deployment Topology
       6.4.2.  Deployment Sequences
       6.4.3.  L4S Flow but Non-ECN Bottleneck
       6.4.4.  L4S Flow but Classic ECN Bottleneck
       6.4.5.  L4S AQM Deployment within Tunnels
   7.  IANA Considerations
   8.  Security Considerations
     8.1.  Traffic Rate (Non-)Policing
       8.1.1.  (Non-)Policing Rate per Flow
       8.1.2.  (Non-)Policing L4S Service Rate
     8.2.  'Latency Friendliness'
     8.3.  Interaction between Rate Policing and L4S
     8.4.  ECN Integrity
     8.5.  Privacy Considerations
   9.  Informative References
   Acknowledgements

   Authors' Addresses



1.  Introduction



   At any one time, it is increasingly common for all of the traffic in
   a bottleneck link (e.g., a household's Internet access or Wi-Fi) to
   come from applications that prefer low delay: interactive web, web
   services, voice, conversational video, interactive video, interactive
   remote presence, instant messaging, online and cloud-rendered gaming,
   remote desktop, cloud-based applications, cloud-rendered virtual
   reality or augmented reality, and video-assisted remote control of
   machinery and industrial processes.  In the last decade or so, much
   has been done to reduce propagation delay by placing caches or
   servers closer to users.  However, queuing remains a major, albeit
   intermittent, component of latency.  For instance, spikes of hundreds
   of milliseconds are not uncommon, even with state-of-the-art Active
   Queue Management (AQM) [COBALT] [DOCSIS3AQM].  A Classic AQM in an
   access network bottleneck is typically configured to buffer the
   sawteeth of lone flows, which can cause peak overall network delay to
   roughly double during a long-running flow, relative to expected base
   (unloaded) path delay [BufferSize].  Low loss is also important
   because, for interactive applications, losses translate into even
   longer retransmission delays.

   It has been demonstrated that, once access network bit rates reach
   levels now common in the developed world, increasing link capacity
   offers diminishing returns if latency (delay) is not addressed
   [Dukkipati06] [Rajiullah15].  Therefore, the goal is an Internet
   service with very low queuing latency, very low loss, and scalable
   throughput.  Very low queuing latency means less than 1 millisecond
   (ms) on average and less than about 2 ms at the 99th percentile.
   End-to-end delay above 50 ms [Raaen14], or even above 20 ms [NASA04],
   starts to feel unnatural for more demanding interactive applications.
   Therefore, removing unnecessary delay variability increases the reach
   of these applications (the distance over which they are comfortable
   to use) and/or provides additional latency budget that can be used
   for enhanced processing.  This document describes the L4S
   architecture for achieving these goals.

   Differentiated services (Diffserv) offers Expedited Forwarding (EF)
   [RFC3246] for some packets at the expense of others, but this makes
   no difference when all (or most) of the traffic at a bottleneck at
   any one time requires low latency.  In contrast, L4S still works well
   when all traffic is L4S -- a service that gives without taking needs
   none of the configuration or management baggage (traffic policing or
   traffic contracts) associated with favouring some traffic flows over
   others.

   Queuing delay degrades performance intermittently [Hohlfeld14].  It
   occurs i) when a large enough capacity-seeking (e.g., TCP) flow is
   running alongside the user's traffic in the bottleneck link, which is
   typically in the access network, or ii) when the low latency
   application is itself a large capacity-seeking or adaptive rate flow
   (e.g., interactive video).  At these times, the performance
   improvement from L4S must be sufficient for network operators to be
   motivated to deploy it.

   Active Queue Management (AQM) is part of the solution to queuing
   under load.  AQM improves performance for all traffic, but there is a
   limit to how much queuing delay can be reduced by solely changing the
   network without addressing the root of the problem.

   The root of the problem is the presence of standard congestion
   control (Reno [RFC5681]) or compatible variants (e.g., CUBIC
   [RFC8312]) that are used in TCP and in other transports, such as QUIC
   [RFC9000].  We shall use the term 'Classic' for these Reno-friendly
   congestion controls.  Classic congestion controls induce relatively
   large sawtooth-shaped excursions of queue occupancy.  So if a network
   operator naively attempts to reduce queuing delay by configuring an
   AQM to operate at a shallower queue, a Classic congestion control
   will significantly underutilize the link at the bottom of every
   sawtooth.  These sawteeth have also been growing in duration as flow
   rate scales (see Section 5.1 and [RFC3649]).

   It has been demonstrated that, if the sending host replaces a Classic
   congestion control with a 'Scalable' alternative, the performance
   under load of all the above interactive applications can be
   significantly improved once a suitable AQM is deployed in the
   network.  Taking the example solution cited below that uses Data
   Center TCP (DCTCP) [RFC8257] and a Dual-Queue Coupled AQM [RFC9332]
   on a DSL or Ethernet link, queuing delay under heavy load is roughly
   1-2 ms at the 99th percentile without losing link utilization
   [L4Seval22] [DualPI2Linux] (for other link types, see Section 6.3).
   This compares with 5-20 ms on _average_ with a Classic congestion
   control and current state-of-the-art AQMs, such as Flow Queue CoDel
   [RFC8290], Proportional Integral controller Enhanced (PIE) [RFC8033],
   or DOCSIS PIE [RFC8034] and about 20-30 ms at the 99th percentile
   [DualPI2Linux].

   L4S is designed for incremental deployment.  It is possible to deploy
   the L4S service at a bottleneck link alongside the existing best
   efforts service [DualPI2Linux] so that unmodified applications can
   start using it as soon as the sender's stack is updated.  Access
   networks are typically designed with one link as the bottleneck for
   each site (which might be a home, small enterprise, or mobile
   device), so deployment at either or both ends of this link should
   give nearly all the benefit in the respective direction.  With some
   transport protocols, namely TCP [ACCECN], the sender has to check
   that the receiver has been suitably updated to give more accurate
   feedback, whereas with more recent transport protocols, such as QUIC
   [RFC9000] and Datagram Congestion Control Protocol (DCCP) [RFC4340],
   all receivers have always been suitable.

   This document presents the L4S architecture.  It consists of three
   components: network support to isolate L4S traffic from Classic
   traffic; protocol features that allow network elements to identify
   L4S traffic; and host support for L4S congestion controls.  The
   protocol is defined separately in [RFC9331] as an experimental change
   to Explicit Congestion Notification (ECN).  This document describes
   and justifies the component parts and how they interact to provide
   the low latency, low loss, and scalable Internet service.  It also
   details the approach to incremental deployment, as briefly summarized
   above.

1.1.  Document Roadmap



   This document describes the L4S architecture in three passes.  First,
   the brief overview in Section 2 gives the very high-level idea and
   states the main components with minimal rationale.  This is only
   intended to give some context for the terminology definitions that
   follow in Section 3 and to explain the structure of the rest of the
   document.  Then, Section 4 goes into more detail on each component
   with some rationale but still mostly stating what the architecture
   is, rather than why.  Finally, Section 5 justifies why each element
   of the solution was chosen (Section 5.1) and why these choices were
   different from other solutions (Section 5.2).

   After the architecture has been described, Section 6 clarifies its
   applicability by describing the applications and use cases that
   motivated the design, the challenges applying the architecture to
   various link technologies, and various incremental deployment models
   (including the two main deployment topologies, different sequences
   for incremental deployment, and various interactions with preexisting
   approaches).  The document ends with the usual tailpieces, including
   extensive discussion of traffic policing and other security
   considerations in Section 8.

2.  L4S Architecture Overview



   Below, we outline the three main components to the L4S architecture:
   1) the Scalable congestion control on the sending host; 2) the AQM at
   the network bottleneck; and 3) the protocol between them.

   But first, the main point to grasp is that low latency is not
   provided by the network; low latency results from the careful
   behaviour of the Scalable congestion controllers used by L4S senders.
   The network does have a role, primarily to isolate the low latency of
   the carefully behaving L4S traffic from the higher queuing delay
   needed by traffic with preexisting Classic behaviour.  The network
   also alters the way it signals queue growth to the transport.  It
   uses the Explicit Congestion Notification (ECN) protocol, but it
   signals the very start of queue growth immediately, without the
   smoothing delay typical of Classic AQMs.  Because ECN support is
   essential for L4S, senders use the ECN field as the protocol that
   allows the network to identify which packets are L4S and which are
   Classic.

   1)  Host:

       Scalable congestion controls already exist.  They solve the
       scaling problem with Classic congestion controls, such as Reno or
       CUBIC.  Because flow rate has scaled since TCP congestion control
       was first designed in 1988, assuming the flow lasts long enough,
       it now takes hundreds of round trips (and growing) to recover
       after a congestion signal (whether a loss or an ECN mark), as
       shown in the examples in Section 5.1 and [RFC3649].  Therefore,
       control of queuing and utilization becomes very slack, and the
       slightest disturbances (e.g., from new flows starting) prevent a
       high rate from being attained.

       With a Scalable congestion control, the average time from one
       congestion signal to the next (the recovery time) remains
       invariant as flow rate scales, all other factors being equal.
       This maintains the same degree of control over queuing and
       utilization, whatever the flow rate, as well as ensuring that
       high throughput is more robust to disturbances.  The Scalable
       control used most widely (in controlled environments) is DCTCP
       [RFC8257], which has been implemented and deployed in Windows
       Server Editions (since 2012), in Linux, and in FreeBSD.  Although
       DCTCP as-is functions well over wide-area round-trip times
       (RTTs), most implementations lack certain safety features that
       would be necessary for use outside controlled environments, like
       data centres (see Section 6.4.3).  Therefore, Scalable congestion
       control needs to be implemented in TCP and other transport
       protocols (QUIC, Stream Control Transmission Protocol (SCTP),
       RTP/RTCP, RTP Media Congestion Avoidance Techniques (RMCAT),
       etc.).  Indeed, between the present document being drafted and
       published, the following Scalable congestion controls were
       implemented: Prague over TCP and QUIC [PRAGUE-CC] [PragueLinux],
       an L4S variant of the RMCAT SCReAM controller [SCReAM-L4S], and
       the L4S ECN part of Bottleneck Bandwidth and Round-trip
       propagation time (BBRv2) [BBRv2] intended for TCP and QUIC
       transports.

   2)  Network:

       L4S traffic needs to be isolated from the queuing latency of
       Classic traffic.  One queue per application flow (FQ) is one way
       to achieve this, e.g., FQ-CoDel [RFC8290].  However, using just
       two queues is sufficient and does not require inspection of
       transport layer headers in the network, which is not always
       possible (see Section 5.2).  With just two queues, it might seem
       impossible to know how much capacity to schedule for each queue
       without inspecting how many flows at any one time are using each.
       And it would be undesirable to arbitrarily divide access network
       capacity into two partitions.  The Dual-Queue Coupled AQM was
       developed as a minimal complexity solution to this problem.  It
       acts like a 'semi-permeable' membrane that partitions latency but
       not bandwidth.  As such, the two queues are for transitioning
       from Classic to L4S behaviour, not bandwidth prioritization.

       Section 4 gives a high-level explanation of how the per-flow
       queue (FQ) and DualQ variants of L4S work, and [RFC9332] gives a
       full explanation of the DualQ Coupled AQM framework.  A specific
       marking algorithm is not mandated for L4S AQMs.  Appendices of
       [RFC9332] give non-normative examples that have been implemented
       and evaluated and give recommended default parameter settings.
       It is expected that L4S experiments will improve knowledge of
       parameter settings and whether the set of marking algorithms
       needs to be limited.

   3)  Protocol:

       A sending host needs to distinguish L4S and Classic packets with
       an identifier so that the network can classify them into their
       separate treatments.  The L4S identifier spec [RFC9331] concludes
       that all alternatives involve compromises, but the ECT(1) and
       Congestion Experienced (CE) codepoints of the ECN field represent
       a workable solution.  As already explained, the network also uses
       ECN to immediately signal the very start of queue growth to the
       transport.

3.  Terminology



   Classic Congestion Control:  A congestion control behaviour that can
      coexist with standard Reno [RFC5681] without causing significantly
      negative impact on its flow rate [RFC5033].  The scaling problem
      with Classic congestion control is explained, with examples, in
      Section 5.1 and in [RFC3649].

   Scalable Congestion Control:  A congestion control where the average
      time from one congestion signal to the next (the recovery time)
      remains invariant as flow rate scales, all other factors being
      equal.  For instance, DCTCP averages 2 congestion signals per
      round trip, whatever the flow rate, as do other recently developed
      Scalable congestion controls, e.g., Relentless TCP [RELENTLESS],
      Prague for TCP and QUIC [PRAGUE-CC] [PragueLinux], BBRv2 [BBRv2]
      [BBR-CC], and the L4S variant of SCReAM for real-time media
      [SCReAM-L4S] [RFC8298].  See Section 4.3 of [RFC9331] for more
      explanation.

   Classic Service:  The Classic service is intended for all the
      congestion control behaviours that coexist with Reno [RFC5681]
      (e.g., Reno itself, CUBIC [RFC8312], Compound [CTCP], and TFRC
      [RFC5348]).  The term 'Classic queue' means a queue providing the
      Classic service.

   Low Latency, Low Loss, and Scalable throughput (L4S) service:  The
      'L4S' service is intended for traffic from Scalable congestion
      control algorithms, such as the Prague congestion control
      [PRAGUE-CC], which was derived from DCTCP [RFC8257].  The L4S
      service is for more general traffic than just Prague -- it allows
      the set of congestion controls with similar scaling properties to
      Prague to evolve, such as the examples listed above (Relentless,
      SCReAM, etc.).  The term 'L4S queue' means a queue providing the
      L4S service.

      The terms Classic or L4S can also qualify other nouns, such as
      'queue', 'codepoint', 'identifier', 'classification', 'packet',
      and 'flow'.  For example, an L4S packet means a packet with an L4S
      identifier sent from an L4S congestion control.

      Both Classic and L4S services can cope with a proportion of
      unresponsive or less-responsive traffic as well but, in the L4S
      case, its rate has to be smooth enough or low enough to not build
      a queue (e.g., DNS, Voice over IP (VoIP), game sync datagrams,
      etc.).

   Reno-friendly:  The subset of Classic traffic that is friendly to the
      standard Reno congestion control defined for TCP in [RFC5681].
      The TFRC spec [RFC5348] indirectly implies that 'friendly' is
      defined as "generally within a factor of two of the sending rate
      of a TCP flow under the same conditions".  Reno-friendly is used
      here in place of 'TCP-friendly', given the latter has become
      imprecise, because the TCP protocol is now used with so many
      different congestion control behaviours, and Reno is used in non-
      TCP transports, such as QUIC [RFC9000].

   Classic ECN:  The original Explicit Congestion Notification (ECN)
      protocol [RFC3168] that requires ECN signals to be treated as
      equivalent to drops, both when generated in the network and when
      responded to by the sender.

      For L4S, the names used for the four codepoints of the 2-bit IP-
      ECN field are unchanged from those defined in the ECN spec
      [RFC3168], i.e., Not-ECT, ECT(0), ECT(1), and CE, where ECT stands
      for ECN-Capable Transport and CE stands for Congestion
      Experienced.  A packet marked with the CE codepoint is termed
      'ECN-marked' or sometimes just 'marked' where the context makes
      ECN obvious.

   Site:  A home, mobile device, small enterprise, or campus where the
      network bottleneck is typically the access link to the site.  Not
      all network arrangements fit this model, but it is a useful,
      widely applicable generalization.

   Traffic Policing:  Limiting traffic by dropping packets or shifting
      them to a lower service class (as opposed to introducing delay,
      which is termed 'traffic shaping').  Policing can involve limiting
      the average rate and/or burst size.  Policing focused on limiting
      queuing but not the average flow rate is termed 'congestion
      policing', 'latency policing', 'burst policing', or 'queue
      protection' in this document.  Otherwise, the term rate policing
      is used.

4.  L4S Architecture Components



   The L4S architecture is composed of the elements in the following
   three subsections.

4.1.  Protocol Mechanisms



   The L4S architecture involves: a) unassignment of the previous use of
   the identifier; b) reassignment of the same identifier; and c)
   optional further identifiers:

   a.  An essential aspect of a Scalable congestion control is the use
       of explicit congestion signals.  Classic ECN [RFC3168] requires
       an ECN signal to be treated as equivalent to drop, both when it
       is generated in the network and when it is responded to by hosts.
       L4S needs networks and hosts to support a more fine-grained
       meaning for each ECN signal that is less severe than a drop, so
       that the L4S signals:

       *  can be much more frequent and

       *  can be signalled immediately, without the significant delay
          required to smooth out fluctuations in the queue.

       To enable L4S, the Standards Track Classic ECN spec [RFC3168] has
       had to be updated to allow L4S packets to depart from the
       'equivalent-to-drop' constraint.  [RFC8311] is a Standards Track
       update to relax specific requirements in [RFC3168] (and certain
       other Standards Track RFCs), which clears the way for the
       experimental changes proposed for L4S.  Also, the ECT(1)
       codepoint was previously assigned as the experimental ECN nonce
       [RFC3540], which [RFC8311] recategorizes as historic to make the
       codepoint available again.

   b.  [RFC9331] specifies that ECT(1) is used as the identifier to
       classify L4S packets into a separate treatment from Classic
       packets.  This satisfies the requirement for identifying an
       alternative ECN treatment in [RFC4774].

       The CE codepoint is used to indicate Congestion Experienced by
       both L4S and Classic treatments.  This raises the concern that a
       Classic AQM earlier on the path might have marked some ECT(0)
       packets as CE.  Then, these packets will be erroneously
       classified into the L4S queue.  Appendix B of [RFC9331] explains
       why five unlikely eventualities all have to coincide for this to
       have any detrimental effect, which even then would only involve a
       vanishingly small likelihood of a spurious retransmission.

   c.  A network operator might wish to include certain unresponsive,
       non-L4S traffic in the L4S queue if it is deemed to be paced
       smoothly enough and at a low enough rate not to build a queue,
       for instance, VoIP, low rate datagrams to sync online games,
       relatively low rate application-limited traffic, DNS, Lightweight
       Directory Access Protocol (LDAP), etc.  This traffic would need
       to be tagged with specific identifiers, e.g., a low-latency
       Diffserv codepoint such as Expedited Forwarding (EF) [RFC3246],
       Non-Queue-Building (NQB) [NQB-PHB], or operator-specific
       identifiers.

4.2.  Network Components



   The L4S architecture aims to provide low latency without the _need_
   for per-flow operations in network components.  Nonetheless, the
   architecture does not preclude per-flow solutions.  The following
   bullets describe the known arrangements: a) the DualQ Coupled AQM
   with an L4S AQM in one queue coupled from a Classic AQM in the other;
   b) per-flow queues with an instance of a Classic and an L4S AQM in
   each queue; and c) Dual queues with per-flow AQMs but no per-flow
   queues:

   a.  The Dual-Queue Coupled AQM (illustrated in Figure 1) achieves the
       'semi-permeable' membrane property mentioned earlier as follows:

       *  Latency isolation: Two separate queues are used to isolate L4S
          queuing delay from the larger queue that Classic traffic needs
          to maintain full utilization.

       *  Bandwidth pooling: The two queues act as if they are a single
          pool of bandwidth in which flows of either type get roughly
          equal throughput without the scheduler needing to identify any
          flows.  This is achieved by having an AQM in each queue, but
          the Classic AQM provides a congestion signal to both queues in
          a manner that ensures a consistent response from the two
          classes of congestion control.  Specifically, the Classic AQM
          generates a drop/mark probability based on congestion in its
          own queue, which it uses both to drop/mark packets in its own
          queue and to affect the marking probability in the L4S queue.
          The strength of the coupling of the congestion signalling
          between the two queues is enough to make the L4S flows slow
          down to leave the right amount of capacity for the Classic
          flows (as they would if they were the same type of traffic
          sharing the same queue).

       Then, the scheduler can serve the L4S queue with priority
       (denoted by the '1' on the higher priority input), because the
       L4S traffic isn't offering up enough traffic to use all the
       priority that it is given.  Therefore:

       *  for latency isolation on short timescales (sub-round-trip),
          the prioritization of the L4S queue protects its low latency
          by allowing bursts to dissipate quickly;

       *  but for bandwidth pooling on longer timescales (round-trip and
          longer), the Classic queue creates an equal and opposite
          pressure against the L4S traffic to ensure that neither has
          priority when it comes to bandwidth -- the tension between
          prioritizing L4S and coupling the marking from the Classic AQM
          results in approximate per-flow fairness.

       To protect against the prioritization of persistent L4S traffic
       deadlocking the Classic queue for a while in some
       implementations, it is advisable for the priority to be
       conditional, not strict (see Appendix A of the DualQ spec
       [RFC9332]).

       When there is no Classic traffic, the L4S queue's own AQM comes
       into play.  It starts congestion marking with a very shallow
       queue, so L4S traffic maintains very low queuing delay.

       If either queue becomes persistently overloaded, drop of some
       ECN-capable packets is introduced, as recommended in Section 7 of
       the ECN spec [RFC3168] and Section 4.2.1 of the AQM
       recommendations [RFC7567].  The trade-offs with different
       approaches are discussed in Section 4.2.3 of the DualQ spec
       [RFC9332] (not shown in the figure here).

       The Dual-Queue Coupled AQM has been specified as generically as
       possible [RFC9332] without specifying the particular AQMs to use
       in the two queues so that designers are free to implement diverse
       ideas.  Informational appendices in that document give pseudocode
       examples of two different specific AQM approaches: one called
       DualPI2 (pronounced Dual PI Squared) [DualPI2Linux] that uses the
       PI2 variant of PIE and a zero-config variant of Random Early
       Detection (RED) called Curvy RED.  A DualQ Coupled AQM based on
       PIE has also been specified and implemented for Low Latency
       DOCSIS [DOCSIS3.1].

                     (3)                  (2)
                     .-------^------..------------^------------------.
        ,-(1)-----.                               _____
       ; ________  :            L4S  -------.    |     |
       :|Scalable| :               _\      ||__\_|mark |
       :| sender | :  __________  / /      ||  / |_____|\   _________
       :|________|\; |          |/   -------'       ^    \1|condit'nl|
        `---------'\_|  IP-ECN  |          Coupling :     \|priority |_\
         ________  / |Classifier|                   :     /|scheduler| /
        |Classic |/  |__________|\   -------.     __:__  / |_________|
        | sender |                \_\ || | ||__\_|mark/|/
        |________|                  / || | ||  / |drop |
                             Classic -------'    |_____|


       (1) Scalable sending host
       (2) Isolation in separate network queues
       (3) Packet identification protocol

           Figure 1: Components of an L4S DualQ Coupled AQM Solution

   b.  Per-Flow Queues and AQMs: A scheduler with per-flow queues, such
       as FQ-CoDel or FQ-PIE, can be used for L4S.  For instance, within
       each queue of an FQ-CoDel system, as well as a CoDel AQM, there
       is typically also the option of ECN marking at an immediate
       (unsmoothed) shallow threshold to support use in data centres
       (see Section 5.2.7 of the FQ-CoDel spec [RFC8290]).  In Linux,
       this has been modified so that the shallow threshold can be
       solely applied to ECT(1) packets [FQ_CoDel_Thresh].  Then, if
       there is a flow of Not-ECT or ECT(0) packets in the per-flow
       queue, the Classic AQM (e.g., CoDel) is applied; whereas, if
       there is a flow of ECT(1) packets in the queue, the shallower
       (typically sub-millisecond) threshold is applied.  In addition,
       ECT(0) and Not-ECT packets could potentially be classified into a
       separate flow queue from ECT(1) and CE packets to avoid them
       mixing if they share a common flow identifier (e.g., in a VPN).

   c.  Dual queues but per-flow AQMs: It should also be possible to use
       dual queues for isolation but with per-flow marking to control
       flow rates (instead of the coupled per-queue marking of the Dual-
       Queue Coupled AQM).  One of the two queues would be for isolating
       L4S packets, which would be classified by the ECN codepoint.
       Flow rates could be controlled by flow-specific marking.  The
       policy goal of the marking could be to differentiate flow rates
       (e.g., [Nadas20], which requires additional signalling of a per-
       flow 'value') or to equalize flow rates (perhaps in a similar way
       to Approx Fair CoDel [AFCD] [CODEL-APPROX-FAIR] but with two
       queues not one).

       Note that, whenever the term 'DualQ' is used loosely without
       saying whether marking is per queue or per flow, it means a dual-
       queue AQM with per-queue marking.

4.3.  Host Mechanisms



   The L4S architecture includes two main mechanisms in the end host
   that we enumerate next:

   a.  Scalable congestion control at the sender: Section 2 defines a
       Scalable congestion control as one where the average time from
       one congestion signal to the next (the recovery time) remains
       invariant as flow rate scales, all other factors being equal.
       DCTCP is the most widely used example.  It has been documented as
       an informational record of the protocol currently in use in
       controlled environments [RFC8257].  A list of safety and
       performance improvements for a Scalable congestion control to be
       usable on the public Internet has been drawn up (see the so-
       called 'Prague L4S requirements' in Appendix A of [RFC9331]).
       The subset that involve risk of harm to others have been captured
       as normative requirements in Section 4 of [RFC9331].  TCP Prague
       [PRAGUE-CC] has been implemented in Linux as a reference
       implementation to address these requirements [PragueLinux].

       Transport protocols other than TCP use various congestion
       controls that are designed to be friendly with Reno.  Before they
       can use the L4S service, they will need to be updated to
       implement a Scalable congestion response, which they will have to
       indicate by using the ECT(1) codepoint.  Scalable variants are
       under consideration for more recent transport protocols (e.g.,
       QUIC), and the L4S ECN part of BBRv2 [BBRv2] [BBR-CC] is a
       Scalable congestion control intended for the TCP and QUIC
       transports, amongst others.  Also, an L4S variant of the RMCAT
       SCReAM controller [RFC8298] has been implemented [SCReAM-L4S] for
       media transported over RTP.

       Section 4.3 of the L4S ECN spec [RFC9331] defines Scalable
       congestion control in more detail and specifies the requirements
       that an L4S Scalable congestion control has to comply with.

   b.  The ECN feedback in some transport protocols is already
       sufficiently fine-grained for L4S (specifically DCCP [RFC4340]
       and QUIC [RFC9000]).  But others either require updates or are in
       the process of being updated:

       *  For the case of TCP, the feedback protocol for ECN embeds the
          assumption from Classic ECN [RFC3168] that an ECN mark is
          equivalent to a drop, making it unusable for a Scalable TCP.
          Therefore, the implementation of TCP receivers will have to be
          upgraded [RFC7560].  Work to standardize and implement more
          accurate ECN feedback for TCP (AccECN) is in progress [ACCECN]
          [PragueLinux].

       *  ECN feedback was only roughly sketched in the appendix of the
          now obsoleted second specification of SCTP [RFC4960], while a
          fuller specification was proposed in a long-expired document
          [ECN-SCTP].  A new design would need to be implemented and
          deployed before SCTP could support L4S.

       *  For RTP, sufficient ECN feedback was defined in [RFC6679], but
          [RFC8888] defines the latest Standards Track improvements.

5.  Rationale



5.1.  Why These Primary Components?



   Explicit congestion signalling (protocol):  Explicit congestion
      signalling is a key part of the L4S approach.  In contrast, use of
      drop as a congestion signal creates tension because drop is both
      an impairment (less would be better) and a useful signal (more
      would be better):

      *  Explicit congestion signals can be used many times per round
         trip to keep tight control without any impairment.  Under heavy
         load, even more explicit signals can be applied so that the
         queue can be kept short whatever the load.  In contrast,
         Classic AQMs have to introduce very high packet drop at high
         load to keep the queue short.  By using ECN, an L4S congestion
         control's sawtooth reduction can be smaller and therefore
         return to the operating point more often, without worrying that
         more sawteeth will cause more signals.  The consequent smaller
         amplitude sawteeth fit between an empty queue and a very
         shallow marking threshold (~1 ms in the public Internet), so
         queue delay variation can be very low, without risk of
         underutilization.

      *  Explicit congestion signals can be emitted immediately to track
         fluctuations of the queue.  L4S shifts smoothing from the
         network to the host.  The network doesn't know the round-trip
         times (RTTs) of any of the flows.  So if the network is
         responsible for smoothing (as in the Classic approach), it has
         to assume a worst case RTT, otherwise long RTT flows would
         become unstable.  This delays Classic congestion signals by
         100-200 ms.  In contrast, each host knows its own RTT.  So, in
         the L4S approach, the host can smooth each flow over its own
         RTT, introducing no more smoothing delay than strictly
         necessary (usually only a few milliseconds).  A host can also
         choose not to introduce any smoothing delay if appropriate,
         e.g., during flow start-up.

      Neither of the above are feasible if explicit congestion
      signalling has to be considered 'equivalent to drop' (as was
      required with Classic ECN [RFC3168]), because drop is an
      impairment as well as a signal.  So drop cannot be excessively
      frequent, and drop cannot be immediate; otherwise, too many drops
      would turn out to have been due to only a transient fluctuation in
      the queue that would not have warranted dropping a packet in
      hindsight.  Therefore, in an L4S AQM, the L4S queue uses a new L4S
      variant of ECN that is not equivalent to drop (see Section 5.2 of
      the L4S ECN spec [RFC9331]), while the Classic queue uses either
      Classic ECN [RFC3168] or drop, which are still equivalent to each
      other.

      Before Classic ECN was standardized, there were various proposals
      to give an ECN mark a different meaning from drop.  However, there
      was no particular reason to agree on any one of the alternative
      meanings, so 'equivalent to drop' was the only compromise that
      could be reached.  [RFC3168] contains a statement that:

         An environment where all end nodes were ECN-Capable could
          allow new criteria to be developed for setting the CE
          codepoint, and new congestion control mechanisms for end-node
          reaction to CE packets.  However, this is a research issue,
          and as such is not addressed in this document.

   Latency isolation (network):  L4S congestion controls keep queue
      delay low, whereas Classic congestion controls need a queue of the
      order of the RTT to avoid underutilization.  One queue cannot have
      two lengths; therefore, L4S traffic needs to be isolated in a
      separate queue (e.g., DualQ) or queues (e.g., FQ).

   Coupled congestion notification:  Coupling the congestion
      notification between two queues as in the DualQ Coupled AQM is not
      necessarily essential, but it is a simple way to allow senders to
      determine their rate packet by packet, rather than be overridden
      by a network scheduler.  An alternative is for a network scheduler
      to control the rate of each application flow (see the discussion
      in Section 5.2).

   L4S packet identifier (protocol):  Once there are at least two
      treatments in the network, hosts need an identifier at the IP
      layer to distinguish which treatment they intend to use.

   Scalable congestion notification:  A Scalable congestion control in
      the host keeps the signalling frequency from the network high,
      whatever the flow rate, so that queue delay variations can be
      small when conditions are stable, and rate can track variations in
      available capacity as rapidly as possible otherwise.

   Low loss:  Latency is not the only concern of L4S.  The 'Low Loss'
      part of the name denotes that L4S generally achieves zero
      congestion loss due to its use of ECN.  Otherwise, loss would
      itself cause delay, particularly for short flows, due to
      retransmission delay [RFC2884].

   Scalable throughput:  The 'Scalable throughput' part of the name
      denotes that the per-flow throughput of Scalable congestion
      controls should scale indefinitely, avoiding the imminent scaling
      problems with Reno-friendly congestion control algorithms
      [RFC3649].  It was known when TCP congestion avoidance was first
      developed in 1988 that it would not scale to high bandwidth-delay
      products (see footnote 6 in [TCP-CA]).  Today, regular broadband
      flow rates over WAN distances are already beyond the scaling range
      of Classic Reno congestion control.  So 'less unscalable' CUBIC
      [RFC8312] and Compound [CTCP] variants of TCP have been
      successfully deployed.  However, these are now approaching their
      scaling limits.

      For instance, we will consider a scenario with a maximum RTT of 30
      ms at the peak of each sawtooth.  As Reno packet rate scales 8
      times from 1,250 to 10,000 packet/s (from 15 to 120 Mb/s with 1500
      B packets), the time to recover from a congestion event rises
      proportionately by 8 times as well, from 422 ms to 3.38 s.  It is
      clearly problematic for a congestion control to take multiple
      seconds to recover from each congestion event.  CUBIC [RFC8312]
      was developed to be less unscalable, but it is approaching its
      scaling limit; with the same max RTT of 30 ms, at 120 Mb/s, CUBIC
      is still fully in its Reno-friendly mode, so it takes about 4.3 s
      to recover.  However, once flow rate scales by 8 times again to
      960 Mb/s it enters true CUBIC mode, with a recovery time of 12.2
      s.  From then on, each further scaling by 8 times doubles CUBIC's
      recovery time (because the cube root of 8 is 2), e.g., at 7.68 Gb/
      s, the recovery time is 24.3 s.  In contrast, a Scalable
      congestion control like DCTCP or Prague induces 2 congestion
      signals per round trip on average, which remains invariant for any
      flow rate, keeping dynamic control very tight.

      For a feel of where the global average lone-flow download sits on
      this scale at the time of writing (2021), according to [BDPdata],
      the global average fixed access capacity was 103 Mb/s in 2020 and
      the average base RTT to a CDN was 25 to 34 ms in 2019.  Averaging
      of per-country data was weighted by Internet user population (data
      collected globally is necessarily of variable quality, but the
      paper does double-check that the outcome compares well against a
      second source).  So a lone CUBIC flow would at best take about 200
      round trips (5 s) to recover from each of its sawtooth reductions,
      if the flow even lasted that long.  This is described as 'at best'
      because it assumes everyone uses an AQM, whereas in reality, most
      users still have a (probably bloated) tail-drop buffer.  In the
      tail-drop case, the likely average recovery time would be at least
      4 times 5 s, if not more, because RTT under load would be at least
      double that of an AQM, and the recovery time of Reno-friendly
      flows depends on the square of RTT.

      Although work on scaling congestion controls tends to start with
      TCP as the transport, the above is not intended to exclude other
      transports (e.g., SCTP and QUIC) or less elastic algorithms (e.g.,
      RMCAT), which all tend to adopt the same or similar developments.

5.2.  What L4S Adds to Existing Approaches



   All the following approaches address some part of the same problem
   space as L4S.  In each case, it is shown that L4S complements them or
   improves on them, rather than being a mutually exclusive alternative:

   Diffserv:  Diffserv addresses the problem of bandwidth apportionment
      for important traffic as well as queuing latency for delay-
      sensitive traffic.  Of these, L4S solely addresses the problem of
      queuing latency.  Diffserv will still be necessary where important
      traffic requires priority (e.g., for commercial reasons or for
      protection of critical infrastructure traffic) -- see
      [L4S-DIFFSERV].  Nonetheless, the L4S approach can provide low
      latency for all traffic within each Diffserv class (including the
      case where there is only the one default Diffserv class).

      Also, Diffserv can only provide a latency benefit if a small
      subset of the traffic on a bottleneck link requests low latency.
      As already explained, it has no effect when all the applications
      in use at one time at a single site (e.g., a home, small business,
      or mobile device) require low latency.  In contrast, because L4S
      works for all traffic, it needs none of the management baggage
      (traffic policing or traffic contracts) associated with favouring
      some packets over others.  This lack of management baggage ought
      to give L4S a better chance of end-to-end deployment.

      In particular, if networks do not trust end systems to identify
      which packets should be favoured, they assign packets to Diffserv
      classes themselves.  However, the techniques available to such
      networks, like inspection of flow identifiers or deeper inspection
      of application signatures, do not always sit well with encryption
      of the layers above IP [RFC8404].  In these cases, users can have
      either privacy or Quality of Service (QoS), but not both.

      As with Diffserv, the L4S identifier is in the IP header.  But, in
      contrast to Diffserv, the L4S identifier does not convey a want or
      a need for a certain level of quality.  Rather, it promises a
      certain behaviour (Scalable congestion response), which networks
      can objectively verify if they need to.  This is because low delay
      depends on collective host behaviour, whereas bandwidth priority
      depends on network behaviour.

   State-of-the-art AQMs:  AQMs for Classic traffic, such as PIE and FQ-
      CoDel, give a significant reduction in queuing delay relative to
      no AQM at all.  L4S is intended to complement these AQMs and
      should not distract from the need to deploy them as widely as
      possible.  Nonetheless, AQMs alone cannot reduce queuing delay too
      far without significantly reducing link utilization, because the
      root cause of the problem is on the host -- where Classic
      congestion controls use large sawtoothing rate variations.  The
      L4S approach resolves this tension between delay and utilization
      by enabling hosts to minimize the amplitude of their sawteeth.  A
      single-queue Classic AQM is not sufficient to allow hosts to use
      small sawteeth for two reasons: i) smaller sawteeth would not get
      lower delay in an AQM designed for larger amplitude Classic
      sawteeth, because a queue can only have one length at a time and
      ii) much smaller sawteeth implies much more frequent sawteeth, so
      L4S flows would drive a Classic AQM into a high level of ECN-
      marking, which would appear as heavy congestion to Classic flows,
      which in turn would greatly reduce their rate as a result (see
      Section 6.4.4).

   Per-flow queuing or marking:  Similarly, per-flow approaches, such as
      FQ-CoDel or Approx Fair CoDel [AFCD], are not incompatible with
      the L4S approach.  However, per-flow queuing alone is not enough
      -- it only isolates the queuing of one flow from others, not from
      itself.  Per-flow implementations need to have support for
      Scalable congestion control added, which has already been done for
      FQ-CoDel in Linux (see Section 5.2.7 of [RFC8290] and
      [FQ_CoDel_Thresh]).  Without this simple modification, per-flow
      AQMs, like FQ-CoDel, would still not be able to support
      applications that need both very low delay and high bandwidth,
      e.g., video-based control of remote procedures or interactive
      cloud-based video (see Note 1 below).

      Although per-flow techniques are not incompatible with L4S, it is
      important to have the DualQ alternative.  This is because handling
      end-to-end (layer 4) flows in the network (layer 3 or 2) precludes
      some important end-to-end functions.  For instance:

      a.  Per-flow forms of L4S, like FQ-CoDel, are incompatible with
          full end-to-end encryption of transport layer identifiers for
          privacy and confidentiality (e.g., IPsec or encrypted VPN
          tunnels, as opposed to DTLS over UDP), because they require
          packet inspection to access the end-to-end transport flow
          identifiers.

          In contrast, the DualQ form of L4S requires no deeper
          inspection than the IP layer.  So as long as operators take
          the DualQ approach, their users can have both very low queuing
          delay and full end-to-end encryption [RFC8404].

      b.  With per-flow forms of L4S, the network takes over control of
          the relative rates of each application flow.  Some see it as
          an advantage that the network will prevent some flows running
          faster than others.  Others consider it an inherent part of
          the Internet's appeal that applications can control their rate
          while taking account of the needs of others via congestion
          signals.  They maintain that this has allowed applications
          with interesting rate behaviours to evolve, for instance: i) a
          variable bit-rate video that varies around an equal share,
          rather than being forced to remain equal at every instant or
          ii) end-to-end scavenger behaviours [RFC6817] that use less
          than an equal share of capacity [LEDBAT_AQM].

          The L4S architecture does not require the IETF to commit to
          one approach over the other, because it supports both so that
          the 'market' can decide.  Nonetheless, in the spirit of 'Do
          one thing and do it well' [McIlroy78], the DualQ option
          provides low delay without prejudging the issue of flow-rate
          control.  Then, flow rate policing can be added separately if
          desired.  In contrast to scheduling, a policer would allow
          application control up to a point, but the network would still
          be able to set the point at which it intervened to prevent one
          flow completely starving another.

      Note:

      1.  It might seem that self-inflicted queuing delay within a per-
          flow queue should not be counted, because if the delay wasn't
          in the network, it would just shift to the sender.  However,
          modern adaptive applications, e.g., HTTP/2 [RFC9113] or some
          interactive media applications (see Section 6.1), can keep low
          latency objects at the front of their local send queue by
          shuffling priorities of other objects dependent on the
          progress of other transfers (for example, see [lowat]).  They
          cannot shuffle objects once they have released them into the
          network.

   Alternative Back-off ECN (ABE):  Here again, L4S is not an
      alternative to ABE but a complement that introduces much lower
      queuing delay.  ABE [RFC8511] alters the host behaviour in
      response to ECN marking to utilize a link better and give ECN
      flows faster throughput.  It uses ECT(0) and assumes the network
      still treats ECN and drop the same.  Therefore, ABE exploits any
      lower queuing delay that AQMs can provide.  But, as explained
      above, AQMs still cannot reduce queuing delay too much without
      losing link utilization (to allow for other, non-ABE, flows).

   BBR:  Bottleneck Bandwidth and Round-trip propagation time (BBR)
      [BBR-CC] controls queuing delay end-to-end without needing any
      special logic in the network, such as an AQM.  So it works pretty
      much on any path.  BBR keeps queuing delay reasonably low, but
      perhaps not quite as low as with state-of-the-art AQMs, such as
      PIE or FQ-CoDel, and certainly nowhere near as low as with L4S.
      Queuing delay is also not consistently low, due to BBR's regular
      bandwidth probing spikes and its aggressive flow start-up phase.

      L4S complements BBR.  Indeed, BBRv2 can use L4S ECN where
      available and a Scalable L4S congestion control behaviour in
      response to any ECN signalling from the path [BBRv2].  The L4S ECN
      signal complements the delay-based congestion control aspects of
      BBR with an explicit indication that hosts can use, both to
      converge on a fair rate and to keep below a shallow queue target
      set by the network.  Without L4S ECN, both these aspects need to
      be assumed or estimated.

6.  Applicability



6.1.  Applications



   A transport layer that solves the current latency issues will provide
   new service, product, and application opportunities.

   With the L4S approach, the following existing applications also
   experience significantly better quality of experience under load:

   *  gaming, including cloud-based gaming;

   *  VoIP;

   *  video conferencing;

   *  web browsing;

   *  (adaptive) video streaming; and

   *  instant messaging.

   The significantly lower queuing latency also enables some interactive
   application functions to be offloaded to the cloud that would hardly
   even be usable today, including:

   *  cloud-based interactive video and

   *  cloud-based virtual and augmented reality.

   The above two applications have been successfully demonstrated with
   L4S, both running together over a 40 Mb/s broadband access link
   loaded up with the numerous other latency-sensitive applications in
   the previous list, as well as numerous downloads, with all sharing
   the same bottleneck queue simultaneously [L4Sdemo16]
   [L4Sdemo16-Video].  For the former, a panoramic video of a football
   stadium could be swiped and pinched so that, on the fly, a proxy in
   the cloud could generate a sub-window of the match video under the
   finger-gesture control of each user.  For the latter, a virtual
   reality headset displayed a viewport taken from a 360-degree camera
   in a racing car.  The user's head movements controlled the viewport
   extracted by a cloud-based proxy.  In both cases, with a 7 ms end-to-
   end base delay, the additional queuing delay of roughly 1 ms was so
   low that it seemed the video was generated locally.

   Using a swiping finger gesture or head movement to pan a video are
   extremely latency-demanding actions -- far more demanding than VoIP
   -- because human vision can detect extremely low delays of the order
   of single milliseconds when delay is translated into a visual lag
   between a video and a reference point (the finger or the orientation
   of the head sensed by the balance system in the inner ear, i.e., the
   vestibular system).  With an alternative AQM, the video noticeably
   lagged behind the finger gestures and head movements.

   Without the low queuing delay of L4S, cloud-based applications like
   these would not be credible without significantly more access-network
   bandwidth (to deliver all possible areas of the video that might be
   viewed) and more local processing, which would increase the weight
   and power consumption of head-mounted displays.  When all interactive
   processing can be done in the cloud, only the data to be rendered for
   the end user needs to be sent.

   Other low latency high bandwidth applications, such as:

   *  interactive remote presence and

   *  video-assisted remote control of machinery or industrial processes

   are not credible at all without very low queuing delay.  No amount of
   extra access bandwidth or local processing can make up for lost time.

6.2.  Use Cases



   The following use cases for L4S are being considered by various
   interested parties:

   *  where the bottleneck is one of various types of access network,
      e.g., DSL, Passive Optical Networks (PONs), DOCSIS cable, mobile,
      satellite; or where it's a Wi-Fi link (see Section 6.3 for some
      technology-specific details)

   *  private networks of heterogeneous data centres, where there is no
      single administrator that can arrange for all the simultaneous
      changes to senders, receivers, and networks needed to deploy
      DCTCP:

      -  a set of private data centres interconnected over a wide area
         with separate administrations but within the same company

      -  a set of data centres operated by separate companies
         interconnected by a community of interest network (e.g., for
         the finance sector)

      -  multi-tenant (cloud) data centres where tenants choose their
         operating system stack (Infrastructure as a Service (IaaS))

   *  different types of transport (or application) congestion control:

      -  elastic (TCP/SCTP);

      -  real-time (RTP, RMCAT); and

      -  query-response (DNS/LDAP).

   *  where low delay QoS is required but without inspecting or
      intervening above the IP layer [RFC8404]:

      -  Mobile and other networks have tended to inspect higher layers
         in order to guess application QoS requirements.  However, with
         growing demand for support of privacy and encryption, L4S
         offers an alternative.  There is no need to select which
         traffic to favour for queuing when L4S can give favourable
         queuing to all traffic.

   *  If queuing delay is minimized, applications with a fixed delay
      budget can communicate over longer distances or via more
      circuitous paths, e.g., longer chains of service functions
      [RFC7665] or of onion routers.

   *  If delay jitter is minimized, it is possible to reduce the
      dejitter buffers on the receiving end of video streaming, which
      should improve the interactive experience.

6.3.  Applicability with Specific Link Technologies



   Certain link technologies aggregate data from multiple packets into
   bursts and buffer incoming packets while building each burst.  Wi-Fi,
   PON, and cable all involve such packet aggregation, whereas fixed
   Ethernet and DSL do not.  No sender, whether L4S or not, can do
   anything to reduce the buffering needed for packet aggregation.  So
   an AQM should not count this buffering as part of the queue that it
   controls, given no amount of congestion signals will reduce it.

   Certain link technologies also add buffering for other reasons,
   specifically:

   *  Radio links (cellular, Wi-Fi, or satellite) that are distant from
      the source are particularly challenging.  The radio link capacity
      can vary rapidly by orders of magnitude, so it is considered
      desirable to hold a standing queue that can utilize sudden
      increases of capacity.

   *  Cellular networks are further complicated by a perceived need to
      buffer in order to make hand-overs imperceptible.

   L4S cannot remove the need for all these different forms of
   buffering.  However, by removing 'the longest pole in the tent'
   (buffering for the large sawteeth of Classic congestion controls),
   L4S exposes all these 'shorter poles' to greater scrutiny.

   Until now, the buffering needed for these additional reasons tended
   to be over-specified -- with the excuse that none were 'the longest
   pole in the tent'.  But having removed the 'longest pole', it becomes
   worthwhile to minimize them, for instance, reducing packet
   aggregation burst sizes and MAC scheduling intervals.

   Also, certain link types, particularly radio-based links, are far
   more prone to transmission losses.  Section 6.4.3 explains how an L4S
   response to loss has to be as drastic as a Classic response.
   Nonetheless, research referred to in the same section has
   demonstrated potential for considerably more effective loss repair at
   the link layer, due to the relaxed ordering constraints of L4S
   packets.

6.4.  Deployment Considerations



   L4S AQMs, whether DualQ [RFC9332] or FQ [RFC8290], are in themselves
   an incremental deployment mechanism for L4S -- so that L4S traffic
   can coexist with existing Classic (Reno-friendly) traffic.
   Section 6.4.1 explains why only deploying an L4S AQM in one node at
   each end of the access link will realize nearly all the benefit of
   L4S.

   L4S involves both the network and end systems, so Section 6.4.2
   suggests some typical sequences to deploy each part and why there
   will be an immediate and significant benefit after deploying just one
   part.

   Sections 6.4.3 and 6.4.4 describe the converse incremental deployment
   case where there is no L4S AQM at the network bottleneck, so any L4S
   flow traversing this bottleneck has to take care in case it is
   competing with Classic traffic.

6.4.1.  Deployment Topology



   L4S AQMs will not have to be deployed throughout the Internet before
   L4S can benefit anyone.  Operators of public Internet access networks
   typically design their networks so that the bottleneck will nearly
   always occur at one known (logical) link.  This confines the cost of
   queue management technology to one place.

   The case of mesh networks is different and will be discussed later in
   this section.  However, the known-bottleneck case is generally true
   for Internet access to all sorts of different 'sites', where the word
   'site' includes home networks, small- to medium-sized campus or
   enterprise networks and even cellular devices (Figure 2).  Also, this
   known-bottleneck case tends to be applicable whatever the access link
   technology, whether xDSL, cable, PON, cellular, line of sight
   wireless, or satellite.

   Therefore, the full benefit of the L4S service should be available in
   the downstream direction when an L4S AQM is deployed at the ingress
   to this bottleneck link.  And similarly, the full upstream service
   will typically be available once an L4S AQM is deployed at the
   ingress into the upstream link.  (Of course, multihomed sites would
   only see the full benefit once all their access links were covered.)

                                            ______
                                           (      )
                         __          __  (          )
                        |DQ\________/DQ|( enterprise )
                    ___ |__/        \__| ( /campus  )
                   (   )                   (______)
                 (      )                           ___||_
   +----+      (          )  __                 __ /      \
   | DC |-----(    Core    )|DQ\_______________/DQ|| home |
   +----+      (          ) |__/               \__||______|
                  (_____) __
                         |DQ\__/\        __ ,===.
                         |__/    \  ____/DQ||| ||mobile
                                  \/    \__|||_||device
                                            | o |
                                            `---'

       Figure 2: Likely Location of DualQ (DQ) Deployments in Common
                             Access Topologies

   Deployment in mesh topologies depends on how overbooked the core is.
   If the core is non-blocking, or at least generously provisioned so
   that the edges are nearly always the bottlenecks, it would only be
   necessary to deploy an L4S AQM at the edge bottlenecks.  For example,
   some data-centre networks are designed with the bottleneck in the
   hypervisor or host Network Interface Controllers (NICs), while others
   bottleneck at the top-of-rack switch (both the output ports facing
   hosts and those facing the core).

   An L4S AQM would often next be needed where the Wi-Fi links in a home
   sometimes become the bottleneck.  Also an L4S AQM would eventually
   need to be deployed at any other persistent bottlenecks, such as
   network interconnections, e.g., some public Internet exchange points
   and the ingress and egress to WAN links interconnecting data centres.

6.4.2.  Deployment Sequences



   For any one L4S flow to provide benefit, it requires three (or
   sometimes two) parts to have been deployed: i) the congestion control
   at the sender; ii) the AQM at the bottleneck; and iii) older
   transports (namely TCP) need upgraded receiver feedback too.  This
   was the same deployment problem that ECN faced [RFC8170], so we have
   learned from that experience.

   Firstly, L4S deployment exploits the fact that DCTCP already exists
   on many Internet hosts (e.g., Windows, FreeBSD, and Linux), both
   servers and clients.  Therefore, an L4S AQM can be deployed at a
   network bottleneck to immediately give a working deployment of all
   the L4S parts for testing, as long as the ECT(0) codepoint is
   switched to ECT(1).  DCTCP needs some safety concerns to be fixed for
   general use over the public Internet (see Section 4.3 of the L4S ECN
   spec [RFC9331]), but DCTCP is not on by default, so these issues can
   be managed within controlled deployments or controlled trials.

   Secondly, the performance improvement with L4S is so significant that
   it enables new interactive services and products that were not
   previously possible.  It is much easier for companies to initiate new
   work on deployment if there is budget for a new product trial.  In
   contrast, if there were only an incremental performance improvement
   (as with Classic ECN), spending on deployment tends to be much harder
   to justify.

   Thirdly, the L4S identifier is defined so that network operators can
   initially enable L4S exclusively for certain customers or certain
   applications.  However, this is carefully defined so that it does not
   compromise future evolution towards L4S as an Internet-wide service.
   This is because the L4S identifier is defined not only as the end-to-
   end ECN field, but it can also optionally be combined with any other
   packet header or some status of a customer or their access link (see
   Section 5.4 of [RFC9331]).  Operators could do this anyway, even if
   it were not blessed by the IETF.  However, it is best for the IETF to
   specify that, if they use their own local identifier, it must be in
   combination with the IETF's identifier, ECT(1).  Then, if an operator
   has opted for an exclusive local-use approach, they only have to
   remove this extra rule later to make the service work across the
   Internet -- it will already traverse middleboxes, peerings, etc.

   +-+--------------------+----------------------+---------------------+
   | | Servers or proxies |      Access link     |             Clients |
   +-+--------------------+----------------------+---------------------+
   |0| DCTCP (existing)   |                      |    DCTCP (existing) |
   +-+--------------------+----------------------+---------------------+
   |1|                    |Add L4S AQM downstream|                     |
   | |       WORKS DOWNSTREAM FOR CONTROLLED DEPLOYMENTS/TRIALS        |
   +-+--------------------+----------------------+---------------------+
   |2| Upgrade DCTCP to   |                      |Replace DCTCP feedb'k|
   | | TCP Prague         |                      |         with AccECN |
   | |                 FULLY     WORKS     DOWNSTREAM                  |
   +-+--------------------+----------------------+---------------------+
   | |                    |                      |    Upgrade DCTCP to |
   |3|                    | Add L4S AQM upstream |          TCP Prague |
   | |                    |                      |                     |
   | |              FULLY WORKS UPSTREAM AND DOWNSTREAM                |
   +-+--------------------+----------------------+---------------------+

                 Figure 3: Example L4S Deployment Sequence

   Figure 3 illustrates some example sequences in which the parts of L4S
   might be deployed.  It consists of the following stages, preceded by
   a presumption that DCTCP is already installed at both ends:

   1.  DCTCP is not applicable for use over the public Internet, so it
       is emphasized here that any DCTCP flow has to be completely
       contained within a controlled trial environment.

       Within this trial environment, once an L4S AQM has been deployed,
       the trial DCTCP flow will experience immediate benefit, without
       any other deployment being needed.  In this example, downstream
       deployment is first, but in other scenarios, the upstream might
       be deployed first.  If no AQM at all was previously deployed for
       the downstream access, an L4S AQM greatly improves the Classic
       service (as well as adding the L4S service).  If an AQM was
       already deployed, the Classic service will be unchanged (and L4S
       will add an improvement on top).

   2.  In this stage, the name 'TCP Prague' [PRAGUE-CC] is used to
       represent a variant of DCTCP that is designed to be used in a
       production Internet environment (that is, it has to comply with
       all the requirements in Section 4 of the L4S ECN spec [RFC9331],
       which then means it can be used over the public Internet).  If
       the application is primarily unidirectional, 'TCP Prague' at the
       sending end will provide all the benefit needed, as long as the
       receiving end supports Accurate ECN (AccECN) feedback [ACCECN].

       For TCP transports, AccECN feedback is needed at the other end,
       but it is a generic ECN feedback facility that is already planned
       to be deployed for other purposes, e.g., DCTCP and BBR.  The two
       ends can be deployed in either order because, in TCP, an L4S
       congestion control only enables itself if it has negotiated the
       use of AccECN feedback with the other end during the connection
       handshake.  Thus, deployment of TCP Prague on a server enables
       L4S trials to move to a production service in one direction,
       wherever AccECN is deployed at the other end.  This stage might
       be further motivated by the performance improvements of TCP
       Prague relative to DCTCP (see Appendix A.2 of the L4S ECN spec
       [RFC9331]).

       Unlike TCP, from the outset, QUIC ECN feedback [RFC9000] has
       supported L4S.  Therefore, if the transport is QUIC, one-ended
       deployment of a Prague congestion control at this stage is simple
       and sufficient.

       For QUIC, if a proxy sits in the path between multiple origin
       servers and the access bottlenecks to multiple clients, then
       upgrading the proxy with a Scalable congestion control would
       provide the benefits of L4S over all the clients' downstream
       bottlenecks in one go -- whether or not all the origin servers
       were upgraded.  Conversely, where a proxy has not been upgraded,
       the clients served by it will not benefit from L4S at all in the
       downstream, even when any origin server behind the proxy has been
       upgraded to support L4S.

       For TCP, a proxy upgraded to support 'TCP Prague' would provide
       the benefits of L4S downstream to all clients that support AccECN
       (whether or not they support L4S as well).  And in the upstream,
       the proxy would also support AccECN as a receiver, so that any
       client deploying its own L4S support would benefit in the
       upstream direction, irrespective of whether any origin server
       beyond the proxy supported AccECN.

   3.  This is a two-move stage to enable L4S upstream.  An L4S AQM or
       TCP Prague can be deployed in either order as already explained.
       To motivate the first of two independent moves, the deferred
       benefit of enabling new services after the second move has to be
       worth it to cover the first mover's investment risk.  As
       explained already, the potential for new interactive services
       provides this motivation.  An L4S AQM also improves the upstream
       Classic service significantly if no other AQM has already been
       deployed.

   Note that other deployment sequences might occur.  For instance, the
   upstream might be deployed first; a non-TCP protocol might be used
   end to end, e.g., QUIC and RTP; a body, such as the 3GPP, might
   require L4S to be implemented in 5G user equipment; or other random
   acts of kindness might arise.

6.4.3.  L4S Flow but Non-ECN Bottleneck



   If L4S is enabled between two hosts, the L4S sender is required to
   coexist safely with Reno in response to any drop (see Section 4.3 of
   the L4S ECN spec [RFC9331]).

   Unfortunately, as well as protecting Classic traffic, this rule
   degrades the L4S service whenever there is any loss, even if the
   cause is not persistent congestion at a bottleneck, for example:

   *  congestion loss at other transient bottlenecks, e.g., due to
      bursts in shallower queues;

   *  transmission errors, e.g., due to electrical interference; and

   *  rate policing.

   Three complementary approaches are in progress to address this issue,
   but they are all currently research:

   *  In Prague congestion control, ignore certain losses deemed
      unlikely to be due to congestion (using some ideas from BBR
      [BBR-CC] regarding isolated losses).  This could mask any of the
      above types of loss while still coexisting with drop-based
      congestion controls.

   *  A combination of Recent Acknowledgement (RACK) [RFC8985], L4S, and
      link retransmission without resequencing could repair transmission
      errors without the head of line blocking delay usually associated
      with link-layer retransmission [UnorderedLTE] [RFC9331].

   *  Hybrid ECN/drop rate policers (see Section 8.3).

   L4S deployment scenarios that minimize these issues (e.g., over
   wireline networks) can proceed in parallel to this research, in the
   expectation that research success could continually widen L4S
   applicability.

6.4.4.  L4S Flow but Classic ECN Bottleneck



   Classic ECN support is starting to materialize on the Internet as an
   increased level of CE marking.  It is hard to detect whether this is
   all due to the addition of support for ECN in implementations of FQ-
   CoDel and/or FQ-COBALT, which is not generally problematic, because
   flow queue (FQ) scheduling inherently prevents a flow from exceeding
   the 'fair' rate irrespective of its aggressiveness.  However, some of
   this Classic ECN marking might be due to single-queue ECN deployment.
   This case is discussed in Section 4.3 of the L4S ECN spec [RFC9331].

6.4.5.  L4S AQM Deployment within Tunnels



   An L4S AQM uses the ECN field to signal congestion.  So in common
   with Classic ECN, if the AQM is within a tunnel or at a lower layer,
   correct functioning of ECN signalling requires standards-compliant
   propagation of the ECN field up the layers [RFC6040] [ECN-SHIM]
   [ECN-ENCAP].

7.  IANA Considerations



   This document has no IANA actions.

8.  Security Considerations



8.1.  Traffic Rate (Non-)Policing



8.1.1.  (Non-)Policing Rate per Flow



   In the current Internet, ISPs usually enforce separation between the
   capacity of shared links assigned to different 'sites' (e.g.,
   households, businesses, or mobile users -- see terminology in
   Section 3) using some form of scheduler [RFC0970].  And they use
   various techniques, like redirection to traffic scrubbing facilities,
   to deal with flooding attacks.  However, there has never been a
   universal need to police the rate of individual application flows --
   the Internet has generally always relied on self-restraint of
   congestion controls at senders for sharing intra-'site' capacity.

   L4S has been designed not to upset this status quo.  If a DualQ is
   used to provide L4S service, Section 4.2 of [RFC9332] explains how it
   is designed to give no more rate advantage to unresponsive flows than
   a single-queue AQM would, whether or not there is traffic overload.

   Also, in case per-flow rate policing is ever required, it can be
   added because it is orthogonal to the distinction between L4S and
   Classic.  As explained in Section 5.2, the DualQ variant of L4S
   provides low delay without prejudging the issue of flow-rate control.
   So if flow-rate control is needed, per-flow queuing (FQ) with L4S
   support can be used instead, or flow rate policing can be added as a
   modular addition to a DualQ.  However, per-flow rate control is not
   usually deployed as a security mechanism, because an active attacker
   can just shard its traffic over more flow identifiers if the rate of
   each is restricted.

8.1.2.  (Non-)Policing L4S Service Rate



   Section 5.2 explains how Diffserv only makes a difference if some
   packets get less favourable treatment than others, which typically
   requires traffic rate policing for a low latency class.  In contrast,
   it should not be necessary to rate-police access to the L4S service
   to protect the Classic service, because L4S is designed to reduce
   delay without harming the delay or rate of any Classic traffic.

   During early deployment (and perhaps always), some networks will not
   offer the L4S service.  In general, these networks should not need to
   police L4S traffic.  They are required (by both the ECN spec
   [RFC3168] and the L4S ECN spec [RFC9331]) not to change the L4S
   identifier, which would interfere with end-to-end congestion control.
   If they already treat ECN traffic as Not-ECT, they can merely treat
   L4S traffic as Not-ECT too.  At a bottleneck, such networks will
   introduce some queuing and dropping.  When a Scalable congestion
   control detects a drop, it will have to respond safely with respect
   to Classic congestion controls (as required in Section 4.3 of
   [RFC9331]).  This will degrade the L4S service to be no better (but
   never worse) than Classic best efforts whenever a non-ECN bottleneck
   is encountered on a path (see Section 6.4.3).

   In cases that are expected to be rare, networks that solely support
   Classic ECN [RFC3168] in a single queue bottleneck might opt to
   police L4S traffic so as to protect competing Classic ECN traffic
   (for instance, see Section 6.1.3 of the L4S operational guidance
   [L4SOPS]).  However, Section 4.3 of the L4S ECN spec [RFC9331]
   recommends that the sender adapts its congestion response to properly
   coexist with Classic ECN flows, i.e., reverting to the self-restraint
   approach.

   Certain network operators might choose to restrict access to the L4S
   service, perhaps only to selected premium customers as a value-added
   service.  Their packet classifier (item 2 in Figure 1) could identify
   such customers against some other field (e.g., source address range),
   as well as classifying on the ECN field.  If only the ECN L4S
   identifier matched, but not (say) the source address, the classifier
   could direct these packets (from non-premium customers) into the
   Classic queue.  Explaining clearly how operators can use additional
   local classifiers (see Section 5.4 of [RFC9331]) is intended to
   remove any motivation to clear the L4S identifier.  Then at least the
   L4S ECN identifier will be more likely to survive end to end, even
   though the service may not be supported at every hop.  Such local
   arrangements would only require simple registered/not-registered
   packet classification, rather than the managed, application-specific
   traffic policing against customer-specific traffic contracts that
   Diffserv uses.

8.2.  'Latency Friendliness'



   Like the Classic service, the L4S service relies on self-restraint to
   limit the rate in response to congestion.  In addition, the L4S
   service requires self-restraint in terms of limiting latency
   (burstiness).  It is hoped that self-interest and guidance on dynamic
   behaviour (especially flow start-up, which might need to be
   standardized) will be sufficient to prevent transports from sending
   excessive bursts of L4S traffic, given the application's own latency
   will suffer most from such behaviour.

   Because the L4S service can reduce delay without discernibly
   increasing the delay of any Classic traffic, it should not be
   necessary to police L4S traffic to protect the delay of Classic
   traffic.  However, whether burst policing becomes necessary to
   protect other L4S traffic remains to be seen.  Without it, there will
   be potential for attacks on the low latency of the L4S service.

   If needed, various arrangements could be used to address this
   concern:

   Local bottleneck queue protection:  A per-flow (5-tuple) queue
      protection function [DOCSIS-Q-PROT] has been developed for the low
      latency queue in DOCSIS, which has adopted the DualQ L4S
      architecture.  It protects the low latency service from any queue-
      building flows that accidentally or maliciously classify
      themselves into the low latency queue.  It is designed to score
      flows based solely on their contribution to queuing (not flow rate
      in itself).  Then, if the shared low latency queue is at risk of
      exceeding a threshold, the function redirects enough packets of
      the highest scoring flow(s) into the Classic queue to preserve low
      latency.

   Distributed traffic scrubbing:  Rather than policing locally at each
      bottleneck, it may only be necessary to address problems
      reactively, e.g., punitively target any deployments of new bursty
      malware, in a similar way to how traffic from flooding attack
      sources is rerouted via scrubbing facilities.

   Local bottleneck per-flow scheduling:  Per-flow scheduling should
      inherently isolate non-bursty flows from bursty flows (see
      Section 5.2 for discussion of the merits of per-flow scheduling
      relative to per-flow policing).

   Distributed access subnet queue protection:  Per-flow queue
      protection could be arranged for a queue structure distributed
      across a subnet intercommunicating using lower layer control
      messages (see Section 2.1.4 of [QDyn]).  For instance, in a radio
      access network, user equipment already sends regular buffer status
      reports to a radio network controller, which could use this
      information to remotely police individual flows.

   Distributed Congestion Exposure to ingress policers:  The Congestion
      Exposure (ConEx) architecture [RFC7713] uses an egress audit to
      motivate senders to truthfully signal path congestion in-band,
      where it can be used by ingress policers.  An edge-to-edge variant
      of this architecture is also possible.

   Distributed domain-edge traffic conditioning:  An architecture
      similar to Diffserv [RFC2475] may be preferred, where traffic is
      proactively conditioned on entry to a domain, rather than
      reactively policed only if it leads to queuing once combined with
      other traffic at a bottleneck.

   Distributed core network queue protection:  The policing function
      could be divided between per-flow mechanisms at the network
      ingress that characterize the burstiness of each flow into a
      signal carried with the traffic and per-class mechanisms at
      bottlenecks that act on these signals if queuing actually occurs
      once the traffic converges.  This would be somewhat similar to
      [Nadas20], which is in turn similar to the idea behind core
      stateless fair queuing.

   No single one of these possible queue protection capabilities is
   considered an essential part of the L4S architecture, which works
   without any of them under non-attack conditions (much as the Internet
   normally works without per-flow rate policing).  Indeed, even where
   latency policers are deployed, under normal circumstances, they would
   not intervene, and if operators found they were not necessary, they
   could disable them.  Part of the L4S experiment will be to see
   whether such a function is necessary and which arrangements are most
   appropriate to the size of the problem.

8.3.  Interaction between Rate Policing and L4S



   As mentioned in Section 5.2, L4S should remove the need for low
   latency Diffserv classes.  However, those Diffserv classes that give
   certain applications or users priority over capacity would still be
   applicable in certain scenarios (e.g., corporate networks).  Then,
   within such Diffserv classes, L4S would often be applicable to give
   traffic low latency and low loss as well.  Within such a Diffserv
   class, the bandwidth available to a user or application is often
   limited by a rate policer.  Similarly, in the default Diffserv class,
   rate policers are sometimes used to partition shared capacity.

   A Classic rate policer drops any packets exceeding a set rate,
   usually also giving a burst allowance (variants exist where the
   policer re-marks noncompliant traffic to a discard-eligible Diffserv
   codepoint, so they can be dropped elsewhere during contention).
   Whenever L4S traffic encounters one of these rate policers, it will
   experience drops and the source will have to fall back to a Classic
   congestion control, thus losing the benefits of L4S (Section 6.4.3).
   So in networks that already use rate policers and plan to deploy L4S,
   it will be preferable to redesign these rate policers to be more
   friendly to the L4S service.

   L4S-friendly rate policing is currently a research area (note that
   this is not the same as latency policing).  It might be achieved by
   setting a threshold where ECN marking is introduced, such that it is
   just under the policed rate or just under the burst allowance where
   drop is introduced.  For instance, the two-rate, three-colour marker
   [RFC2698] or a PCN threshold and excess-rate marker [RFC5670] could
   mark ECN at the lower rate and drop at the higher.  Or an existing
   rate policer could have congestion-rate policing added, e.g., using
   the 'local' (non-ConEx) variant of the ConEx aggregate congestion
   policer [CONG-POLICING].  It might also be possible to design
   Scalable congestion controls to respond less catastrophically to loss
   that has not been preceded by a period of increasing delay.

   The design of L4S-friendly rate policers will require a separate,
   dedicated document.  For further discussion of the interaction
   between L4S and Diffserv, see [L4S-DIFFSERV].

8.4.  ECN Integrity



   Various ways have been developed to protect the integrity of the
   congestion feedback loop (whether signalled by loss, Classic ECN, or
   L4S ECN) against misbehaviour by the receiver, sender, or network (or
   all three).  Brief details of each, including applicability, pros,
   and cons, are given in Appendix C.1 of the L4S ECN spec [RFC9331].

8.5.  Privacy Considerations



   As discussed in Section 5.2, the L4S architecture does not preclude
   approaches that inspect end-to-end transport layer identifiers.  For
   instance, L4S support has been added to FQ-CoDel, which classifies by
   application flow identifier in the network.  However, the main
   innovation of L4S is the DualQ AQM framework that does not need to
   inspect any deeper than the outermost IP header, because the L4S
   identifier is in the IP-ECN field.

   Thus, the L4S architecture enables very low queuing delay without
   _requiring_ inspection of information above the IP layer.  This means
   that users who want to encrypt application flow identifiers, e.g., in
   IPsec or other encrypted VPN tunnels, don't have to sacrifice low
   delay [RFC8404].

   Because L4S can provide low delay for a broad set of applications
   that choose to use it, there is no need for individual applications
   or classes within that broad set to be distinguishable in any way
   while traversing networks.  This removes much of the ability to
   correlate between the delay requirements of traffic and other
   identifying features [RFC6973].  There may be some types of traffic
   that prefer not to use L4S, but the coarse binary categorization of
   traffic reveals very little that could be exploited to compromise
   privacy.

9.  Informative References



   [ACCECN]   Briscoe, B., Kühlewind, M., and R. Scheffenegger, "More
              Accurate ECN Feedback in TCP", Work in Progress, Internet-
              Draft, draft-ietf-tcpm-accurate-ecn-22, 9 November 2022,
              <https://datatracker.ietf.org/doc/html/draft-ietf-tcpm-
              accurate-ecn-22>.

   [AFCD]     Xue, L., Kumar, S., Cui, C., Kondikoppa, P., Chiu, C-H.,
              and S-J. Park, "Towards fair and low latency next
              generation high speed networks: AFCD queuing", Journal of
              Network and Computer Applications, Volume 70, pp. 183-193,
              DOI 10.1016/j.jnca.2016.03.021, July 2016,
              <https://doi.org/10.1016/j.jnca.2016.03.021>.

   [BBR-CC]   Cardwell, N., Cheng, Y., Hassas Yeganeh, S., Swett, I.,
              and V. Jacobson, "BBR Congestion Control", Work in
              Progress, Internet-Draft, draft-cardwell-iccrg-bbr-
              congestion-control-02, 7 March 2022,
              <https://datatracker.ietf.org/doc/html/draft-cardwell-
              iccrg-bbr-congestion-control-02>.

   [BBRv2]    "TCP BBR v2 Alpha/Preview Release", commit 17700ca, June
              2022, <https://github.com/google/bbr>.

   [BDPdata]  Briscoe, B., "PI2 Parameters", TR-BB-2021-001,
              arXiv:2107.01003 [cs.NI], DOI 10.48550/arXiv.2107.01003,
              October 2021, <https://arxiv.org/abs/2107.01003>.

   [BufferSize]
              Appenzeller, G., Keslassy, I., and N. McKeown, "Sizing
              Router Buffers", SIGCOMM '04: Proceedings of the 2004
              conference on Applications, technologies, architectures,
              and protocols for computer communications, pp. 281-292,
              DOI 10.1145/1015467.1015499, October 2004,
              <https://doi.org/10.1145/1015467.1015499>.

   [COBALT]   Palmei, J., Gupta, S., Imputato, P., Morton, J.,
              Tahiliani, M. P., Avallone, S., and D. Täht, "Design and
              Evaluation of COBALT Queue Discipline", IEEE International
              Symposium on Local and Metropolitan Area Networks
              (LANMAN), DOI 10.1109/LANMAN.2019.8847054, July 2019,
              <https://ieeexplore.ieee.org/abstract/document/8847054>.

   [CODEL-APPROX-FAIR]
              Morton, J. and P. Heist, "Controlled Delay Approximate
              Fairness AQM", Work in Progress, Internet-Draft, draft-
              morton-tsvwg-codel-approx-fair-01, 9 March 2020,
              <https://datatracker.ietf.org/doc/html/draft-morton-tsvwg-
              codel-approx-fair-01>.

   [CONG-POLICING]
              Briscoe, B., "Network Performance Isolation using
              Congestion Policing", Work in Progress, Internet-Draft,
              draft-briscoe-conex-policing-01, 14 February 2014,
              <https://datatracker.ietf.org/doc/html/draft-briscoe-
              conex-policing-01>.

   [CTCP]     Sridharan, M., Tan, K., Bansal, D., and D. Thaler,
              "Compound TCP: A New TCP Congestion Control for High-Speed
              and Long Distance Networks", Work in Progress, Internet-
              Draft, draft-sridharan-tcpm-ctcp-02, 11 November 2008,
              <https://datatracker.ietf.org/doc/html/draft-sridharan-
              tcpm-ctcp-02>.

   [DOCSIS-Q-PROT]
              Briscoe, B., Ed. and G. White, "The DOCSIS® Queue
              Protection Algorithm to Preserve Low Latency", Work in
              Progress, Internet-Draft, draft-briscoe-docsis-q-
              protection-06, 13 May 2022,
              <https://datatracker.ietf.org/doc/html/draft-briscoe-
              docsis-q-protection-06>.

   [DOCSIS3.1]
              CableLabs, "MAC and Upper Layer Protocols Interface
              (MULPI) Specification, CM-SP-MULPIv3.1", Data-Over-Cable
              Service Interface Specifications DOCSIS 3.1 Version i17 or
              later, 21 January 2019, <https://specification-
              search.cablelabs.com/CM-SP-MULPIv3.1>.

   [DOCSIS3AQM]
              White, G., "Active Queue Management Algorithms for DOCSIS
              3.0: A Simulation Study of CoDel, SFQ-CoDel and PIE in
              DOCSIS 3.0 Networks", CableLabs Technical Report, April
              2013, <https://www.cablelabs.com/wp-
              content/uploads/2013/11/
              Active_Queue_Management_Algorithms_DOCSIS_3_0.pdf>.

   [DualPI2Linux]
              Albisser, O., De Schepper, K., Briscoe, B., Tilmans, O.,
              and H. Steen, "DUALPI2 - Low Latency, Low Loss and
              Scalable (L4S) AQM", Proceedings of Linux Netdev 0x13,
              March 2019, <https://www.netdevconf.org/0x13/
              session.html?talk-DUALPI2-AQM>.

   [Dukkipati06]
              Dukkipati, N. and N. McKeown, "Why Flow-Completion Time is
              the Right Metric for Congestion Control", ACM SIGCOMM
              Computer Communication Review, Volume 36, Issue 1, pp.
              59-62, DOI 10.1145/1111322.1111336, January 2006,
              <https://dl.acm.org/doi/10.1145/1111322.1111336>.

   [ECN-ENCAP]
              Briscoe, B. and J. Kaippallimalil, "Guidelines for Adding
              Congestion Notification to Protocols that Encapsulate IP",
              Work in Progress, Internet-Draft, draft-ietf-tsvwg-ecn-
              encap-guidelines-17, 11 July 2022,
              <https://datatracker.ietf.org/doc/html/draft-ietf-tsvwg-
              ecn-encap-guidelines-17>.

   [ECN-SCTP] Stewart, R., Tuexen, M., and X. Dong, "ECN for Stream
              Control Transmission Protocol (SCTP)", Work in Progress,
              Internet-Draft, draft-stewart-tsvwg-sctpecn-05, 15 January
              2014, <https://datatracker.ietf.org/doc/html/draft-
              stewart-tsvwg-sctpecn-05>.

   [ECN-SHIM] Briscoe, B., "Propagating Explicit Congestion Notification
              Across IP Tunnel Headers Separated by a Shim", Work in
              Progress, Internet-Draft, draft-ietf-tsvwg-rfc6040update-
              shim-15, 11 July 2022,
              <https://datatracker.ietf.org/doc/html/draft-ietf-tsvwg-
              rfc6040update-shim-15>.

   [FQ_CoDel_Thresh]
              "fq_codel: generalise ce_threshold marking for subset of
              traffic", commit dfcb63ce1de6b10b, October 2021,
              <https://git.kernel.org/pub/scm/linux/kernel/git/netdev/
              net-next.git/commit/?id=dfcb63ce1de6b10b>.

   [Hohlfeld14]
              Hohlfeld, O., Pujol, E., Ciucu, F., Feldmann, A., and P.
              Barford, "A QoE Perspective on Sizing Network Buffers",
              IMC '14: Proceedings of the 2014 Conference on Internet
              Measurement, pp. 333-346, DOI 10.1145/2663716.2663730,
              November 2014,
              <https://doi.acm.org/10.1145/2663716.2663730>.

   [L4S-DIFFSERV]
              Briscoe, B., "Interactions between Low Latency, Low Loss,
              Scalable Throughput (L4S) and Differentiated Services",
              Work in Progress, Internet-Draft, draft-briscoe-tsvwg-l4s-
              diffserv-02, 4 November 2018,
              <https://datatracker.ietf.org/doc/html/draft-briscoe-
              tsvwg-l4s-diffserv-02>.

   [L4Sdemo16]
              Bondarenko, O., De Schepper, K., Tsang, I., Briscoe, B.,
              Petlund, A., and C. Griwodz, "Ultra-Low Delay for All:
              Live Experience, Live Analysis", Proceedings of the 7th
              International Conference on Multimedia Systems, Article
              No. 33, pp. 1-4, DOI 10.1145/2910017.2910633, May 2016,
              <https://dl.acm.org/citation.cfm?doid=2910017.2910633>.

   [L4Sdemo16-Video]
              "Videos used in IETF dispatch WG 'Ultra-Low Queuing Delay
              for All Apps' slot",
              <https://riteproject.eu/dctth/#1511dispatchwg>.

   [L4Seval22]
              De Schepper, K., Albisser, O., Tilmans, O., and B.
              Briscoe, "Dual Queue Coupled AQM: Deployable Very Low
              Queuing Delay for All", TR-BB-2022-001, arXiv:2209.01078
              [cs.NI], DOI 10.48550/arXiv.2209.01078, September 2022,
              <https://arxiv.org/abs/2209.01078>.

   [L4SOPS]   White, G., Ed., "Operational Guidance for Deployment of
              L4S in the Internet", Work in Progress, Internet-Draft,
              draft-ietf-tsvwg-l4sops-03, 28 April 2022,
              <https://datatracker.ietf.org/doc/html/draft-ietf-tsvwg-
              l4sops-03>.

   [LEDBAT_AQM]
              Al-Saadi, R., Armitage, G., and J. But, "Characterising
              LEDBAT Performance Through Bottlenecks Using PIE, FQ-CoDel
              and FQ-PIE Active Queue Management", IEEE 42nd Conference
              on Local Computer Networks (LCN), DOI 10.1109/LCN.2017.22,
              October 2017,
              <https://ieeexplore.ieee.org/document/8109367>.

   [lowat]    Meenan, P., "Optimizing HTTP/2 prioritization with BBR and
              tcp_notsent_lowat", Cloudflare Blog, October 2018,
              <https://blog.cloudflare.com/http-2-prioritization-with-
              nginx/>.

   [McIlroy78]
              McIlroy, M.D., Pinson, E. N., and B. A. Tague, "UNIX Time-
              Sharing System: Foreword", The Bell System Technical
              Journal 57: 6, pp. 1899-1904,
              DOI 10.1002/j.1538-7305.1978.tb02135.x, July 1978,
              <https://archive.org/details/bstj57-6-1899>.

   [Nadas20]  Nádas, S., Gombos, G., Fejes, F., and S. Laki, "A
              Congestion Control Independent L4S Scheduler", ANRW '20:
              Proceedings of the Applied Networking Research Workshop,
              pp. 45-51, DOI 10.1145/3404868.3406669, July 2020,
              <https://doi.org/10.1145/3404868.3406669>.

   [NASA04]   Bailey, R., Trey Arthur III, J., and S. Williams, "Latency
              Requirements for Head-Worn Display S/EVS Applications",
              Proceedings of SPIE 5424, DOI 10.1117/12.554462, April
              2004, <https://ntrs.nasa.gov/api/citations/20120009198/
              downloads/20120009198.pdf?attachment=true>.

   [NQB-PHB]  White, G. and T. Fossati, "A Non-Queue-Building Per-Hop
              Behavior (NQB PHB) for Differentiated Services", Work in
              Progress, Internet-Draft, draft-ietf-tsvwg-nqb-15, 11
              January 2023, <https://datatracker.ietf.org/doc/html/
              draft-ietf-tsvwg-nqb-15>.

   [PRAGUE-CC]
              De Schepper, K., Tilmans, O., and B. Briscoe, Ed., "Prague
              Congestion Control", Work in Progress, Internet-Draft,
              draft-briscoe-iccrg-prague-congestion-control-01, 11 July
              2022, <https://datatracker.ietf.org/doc/html/draft-
              briscoe-iccrg-prague-congestion-control-01>.

   [PragueLinux]
              Briscoe, B., De Schepper, K., Albisser, O., Misund, J.,
              Tilmans, O., Kühlewind, M., and A.S. Ahmed, "Implementing
              the 'TCP Prague' Requirements for Low Latency Low Loss
              Scalable Throughput (L4S)", Proceedings Linux Netdev 0x13,
              March 2019, <https://www.netdevconf.org/0x13/
              session.html?talk-tcp-prague-l4s>.

   [QDyn]     Briscoe, B., "Rapid Signalling of Queue Dynamics", TR-BB-
              2017-001, arXiv:1904.07044 [cs.NI],
              DOI 10.48550/arXiv.1904.07044, April 2019,
              <https://arxiv.org/abs/1904.07044>.

   [Raaen14]  Raaen, K. and T-M. Grønli, "Latency Thresholds for
              Usability in Games: A Survey", Norsk IKT-konferanse for
              forskning og utdanning (Norwegian ICT conference for
              research and education), 2014,
              <http://ojs.bibsys.no/index.php/NIK/article/view/9/6>.

   [Rajiullah15]
              Rajiullah, M., "Towards a Low Latency Internet:
              Understanding and Solutions", Dissertation, Karlstad
              University, 2015, <https://www.diva-
              portal.org/smash/get/diva2:846109/FULLTEXT01.pdf>.

   [RELENTLESS]
              Mathis, M., "Relentless Congestion Control", Work in
              Progress, Internet-Draft, draft-mathis-iccrg-relentless-
              tcp-00, 4 March 2009,
              <https://datatracker.ietf.org/doc/html/draft-mathis-iccrg-
              relentless-tcp-00>.

   [RFC0970]  Nagle, J., "On Packet Switches With Infinite Storage",
              RFC 970, DOI 10.17487/RFC0970, December 1985,
              <https://www.rfc-editor.org/info/rfc970>.

   [RFC2475]  Blake, S., Black, D., Carlson, M., Davies, E., Wang, Z.,
              and W. Weiss, "An Architecture for Differentiated
              Services", RFC 2475, DOI 10.17487/RFC2475, December 1998,
              <https://www.rfc-editor.org/info/rfc2475>.

   [RFC2698]  Heinanen, J. and R. Guerin, "A Two Rate Three Color
              Marker", RFC 2698, DOI 10.17487/RFC2698, September 1999,
              <https://www.rfc-editor.org/info/rfc2698>.

   [RFC2884]  Hadi Salim, J. and U. Ahmed, "Performance Evaluation of
              Explicit Congestion Notification (ECN) in IP Networks",
              RFC 2884, DOI 10.17487/RFC2884, July 2000,
              <https://www.rfc-editor.org/info/rfc2884>.

   [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>.

   [RFC3246]  Davie, B., Charny, A., Bennet, J.C.R., Benson, K., Le
              Boudec, J.Y., Courtney, W., Davari, S., Firoiu, V., and D.
              Stiliadis, "An Expedited Forwarding PHB (Per-Hop
              Behavior)", RFC 3246, DOI 10.17487/RFC3246, March 2002,
              <https://www.rfc-editor.org/info/rfc3246>.

   [RFC3540]  Spring, N., Wetherall, D., and D. Ely, "Robust Explicit
              Congestion Notification (ECN) Signaling with Nonces",
              RFC 3540, DOI 10.17487/RFC3540, June 2003,
              <https://www.rfc-editor.org/info/rfc3540>.

   [RFC3649]  Floyd, S., "HighSpeed TCP for Large Congestion Windows",
              RFC 3649, DOI 10.17487/RFC3649, December 2003,
              <https://www.rfc-editor.org/info/rfc3649>.

   [RFC4340]  Kohler, E., Handley, M., and S. Floyd, "Datagram
              Congestion Control Protocol (DCCP)", RFC 4340,
              DOI 10.17487/RFC4340, March 2006,
              <https://www.rfc-editor.org/info/rfc4340>.

   [RFC4774]  Floyd, S., "Specifying Alternate Semantics for the
              Explicit Congestion Notification (ECN) Field", BCP 124,
              RFC 4774, DOI 10.17487/RFC4774, November 2006,
              <https://www.rfc-editor.org/info/rfc4774>.

   [RFC4960]  Stewart, R., Ed., "Stream Control Transmission Protocol",
              RFC 4960, DOI 10.17487/RFC4960, September 2007,
              <https://www.rfc-editor.org/info/rfc4960>.

   [RFC5033]  Floyd, S. and M. Allman, "Specifying New Congestion
              Control Algorithms", BCP 133, RFC 5033,
              DOI 10.17487/RFC5033, August 2007,
              <https://www.rfc-editor.org/info/rfc5033>.

   [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>.

   [RFC5670]  Eardley, P., Ed., "Metering and Marking Behaviour of PCN-
              Nodes", RFC 5670, DOI 10.17487/RFC5670, November 2009,
              <https://www.rfc-editor.org/info/rfc5670>.

   [RFC5681]  Allman, M., Paxson, V., and E. Blanton, "TCP Congestion
              Control", RFC 5681, DOI 10.17487/RFC5681, September 2009,
              <https://www.rfc-editor.org/info/rfc5681>.

   [RFC6040]  Briscoe, B., "Tunnelling of Explicit Congestion
              Notification", RFC 6040, DOI 10.17487/RFC6040, November
              2010, <https://www.rfc-editor.org/info/rfc6040>.

   [RFC6679]  Westerlund, M., Johansson, I., Perkins, C., O'Hanlon, P.,
              and K. Carlberg, "Explicit Congestion Notification (ECN)
              for RTP over UDP", RFC 6679, DOI 10.17487/RFC6679, August
              2012, <https://www.rfc-editor.org/info/rfc6679>.

   [RFC6817]  Shalunov, S., Hazel, G., Iyengar, J., and M. Kuehlewind,
              "Low Extra Delay Background Transport (LEDBAT)", RFC 6817,
              DOI 10.17487/RFC6817, December 2012,
              <https://www.rfc-editor.org/info/rfc6817>.

   [RFC6973]  Cooper, A., Tschofenig, H., Aboba, B., Peterson, J.,
              Morris, J., Hansen, M., and R. Smith, "Privacy
              Considerations for Internet Protocols", RFC 6973,
              DOI 10.17487/RFC6973, July 2013,
              <https://www.rfc-editor.org/info/rfc6973>.

   [RFC7560]  Kuehlewind, M., Ed., Scheffenegger, R., and B. Briscoe,
              "Problem Statement and Requirements for Increased Accuracy
              in Explicit Congestion Notification (ECN) Feedback",
              RFC 7560, DOI 10.17487/RFC7560, August 2015,
              <https://www.rfc-editor.org/info/rfc7560>.

   [RFC7567]  Baker, F., Ed. and G. Fairhurst, Ed., "IETF
              Recommendations Regarding Active Queue Management",
              BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015,
              <https://www.rfc-editor.org/info/rfc7567>.

   [RFC7665]  Halpern, J., Ed. and C. Pignataro, Ed., "Service Function
              Chaining (SFC) Architecture", RFC 7665,
              DOI 10.17487/RFC7665, October 2015,
              <https://www.rfc-editor.org/info/rfc7665>.

   [RFC7713]  Mathis, M. and B. Briscoe, "Congestion Exposure (ConEx)
              Concepts, Abstract Mechanism, and Requirements", RFC 7713,
              DOI 10.17487/RFC7713, December 2015,
              <https://www.rfc-editor.org/info/rfc7713>.

   [RFC8033]  Pan, R., Natarajan, P., Baker, F., and G. White,
              "Proportional Integral Controller Enhanced (PIE): A
              Lightweight Control Scheme to Address the Bufferbloat
              Problem", RFC 8033, DOI 10.17487/RFC8033, February 2017,
              <https://www.rfc-editor.org/info/rfc8033>.

   [RFC8034]  White, G. and R. Pan, "Active Queue Management (AQM) Based
              on Proportional Integral Controller Enhanced (PIE) for
              Data-Over-Cable Service Interface Specifications (DOCSIS)
              Cable Modems", RFC 8034, DOI 10.17487/RFC8034, February
              2017, <https://www.rfc-editor.org/info/rfc8034>.

   [RFC8170]  Thaler, D., Ed., "Planning for Protocol Adoption and
              Subsequent Transitions", RFC 8170, DOI 10.17487/RFC8170,
              May 2017, <https://www.rfc-editor.org/info/rfc8170>.

   [RFC8257]  Bensley, S., Thaler, D., Balasubramanian, P., Eggert, L.,
              and G. Judd, "Data Center TCP (DCTCP): TCP Congestion
              Control for Data Centers", RFC 8257, DOI 10.17487/RFC8257,
              October 2017, <https://www.rfc-editor.org/info/rfc8257>.

   [RFC8290]  Hoeiland-Joergensen, T., McKenney, P., Taht, D., Gettys,
              J., and E. Dumazet, "The Flow Queue CoDel Packet Scheduler
              and Active Queue Management Algorithm", RFC 8290,
              DOI 10.17487/RFC8290, January 2018,
              <https://www.rfc-editor.org/info/rfc8290>.

   [RFC8298]  Johansson, I. and Z. Sarker, "Self-Clocked Rate Adaptation
              for Multimedia", RFC 8298, DOI 10.17487/RFC8298, December
              2017, <https://www.rfc-editor.org/info/rfc8298>.

   [RFC8311]  Black, D., "Relaxing Restrictions on Explicit Congestion
              Notification (ECN) Experimentation", RFC 8311,
              DOI 10.17487/RFC8311, January 2018,
              <https://www.rfc-editor.org/info/rfc8311>.

   [RFC8312]  Rhee, I., Xu, L., Ha, S., Zimmermann, A., Eggert, L., and
              R. Scheffenegger, "CUBIC for Fast Long-Distance Networks",
              RFC 8312, DOI 10.17487/RFC8312, February 2018,
              <https://www.rfc-editor.org/info/rfc8312>.

   [RFC8404]  Moriarty, K., Ed. and A. Morton, Ed., "Effects of
              Pervasive Encryption on Operators", RFC 8404,
              DOI 10.17487/RFC8404, July 2018,
              <https://www.rfc-editor.org/info/rfc8404>.

   [RFC8511]  Khademi, N., Welzl, M., Armitage, G., and G. Fairhurst,
              "TCP Alternative Backoff with ECN (ABE)", RFC 8511,
              DOI 10.17487/RFC8511, December 2018,
              <https://www.rfc-editor.org/info/rfc8511>.

   [RFC8888]  Sarker, Z., Perkins, C., Singh, V., and M. Ramalho, "RTP
              Control Protocol (RTCP) Feedback for Congestion Control",
              RFC 8888, DOI 10.17487/RFC8888, January 2021,
              <https://www.rfc-editor.org/info/rfc8888>.

   [RFC8985]  Cheng, Y., Cardwell, N., Dukkipati, N., and P. Jha, "The
              RACK-TLP Loss Detection Algorithm for TCP", RFC 8985,
              DOI 10.17487/RFC8985, February 2021,
              <https://www.rfc-editor.org/info/rfc8985>.

   [RFC9000]  Iyengar, J., Ed. and M. Thomson, Ed., "QUIC: A UDP-Based
              Multiplexed and Secure Transport", RFC 9000,
              DOI 10.17487/RFC9000, May 2021,
              <https://www.rfc-editor.org/info/rfc9000>.

   [RFC9113]  Thomson, M., Ed. and C. Benfield, Ed., "HTTP/2", RFC 9113,
              DOI 10.17487/RFC9113, June 2022,
              <https://www.rfc-editor.org/info/rfc9113>.

   [RFC9331]  De Schepper, K. and B. Briscoe, Ed., "The Explicit
              Congestion Notification (ECN) Protocol for Low Latency,
              Low Loss, and Scalable Throughput (L4S)", RFC 9331,
              DOI 10.17487/RFC9331, January 2023,
              <https://www.rfc-editor.org/info/rfc9331>.

   [RFC9332]  De Schepper, K., Briscoe, B., Ed., and G. White, "Dual-
              Queue Coupled Active Queue Management (AQM) for Low
              Latency, Low Loss, and Scalable Throughput (L4S)",
              RFC 9332, DOI 10.17487/RFC9332, January 2023,
              <https://www.rfc-editor.org/info/rfc9332>.

   [SCReAM-L4S]
              "SCReAM", commit fda6c53, June 2022,
              <https://github.com/EricssonResearch/scream>.

   [TCP-CA]   Jacobson, V. and M. Karels, "Congestion Avoidance and
              Control", Laurence Berkeley Labs Technical Report ,
              November 1988, <https://ee.lbl.gov/papers/congavoid.pdf>.

   [UnorderedLTE]
              Austrheim, M., "Implementing immediate forwarding for 4G
              in a network simulator", Master's Thesis, University of
              Oslo, 2018.

Acknowledgements



   Thanks to Richard Scheffenegger, Wes Eddy, Karen Nielsen, David
   Black, Jake Holland, Vidhi Goel, Ermin Sakic, Praveen
   Balasubramanian, Gorry Fairhurst, Mirja Kuehlewind, Philip Eardley,
   Neal Cardwell, Pete Heist, and Martin Duke for their useful review
   comments.  Thanks also to the area reviewers: Marco Tiloca, Lars
   Eggert, Roman Danyliw, and Éric Vyncke.

   Bob Briscoe and Koen De Schepper were partly funded by the European
   Community under its Seventh Framework Programme through the Reducing
   Internet Transport Latency (RITE) project (ICT-317700).  The
   contribution of Koen De Schepper was also partly funded by the
   5Growth and DAEMON EU H2020 projects.  Bob Briscoe was also partly
   funded by the Research Council of Norway through the TimeIn project,
   partly by CableLabs, and partly by the Comcast Innovation Fund.  The
   views expressed here are solely those of the authors.

Authors' Addresses



   Bob Briscoe (editor)
   Independent
   United Kingdom
   Email: ietf@bobbriscoe.net
   URI:   https://bobbriscoe.net/


   Koen De Schepper
   Nokia Bell Labs
   Antwerp
   Belgium
   Email: koen.de_schepper@nokia.com
   URI:   https://www.bell-labs.com/about/researcher-profiles/
   koende_schepper/


   Marcelo Bagnulo
   Universidad Carlos III de Madrid
   Av. Universidad 30
   28911 Madrid
   Spain
   Phone: 34 91 6249500
   Email: marcelo@it.uc3m.es
   URI:   https://www.it.uc3m.es


   Greg White
   CableLabs
   United States of America
   Email: G.White@CableLabs.com