Simula Research Lab

ietf@bobbriscoe.net http://bobbriscoe.net/

Nokia Bell Labs

Antwerp Belgium koen.de_schepper@nokia.com https://www.bell-labs.com/usr/koen.de_schepper

Universidad Carlos III de Madrid

Av. Universidad 30 Leganes, Madrid 28911 Spain 34 91 6249500 marcelo@it.uc3m.es http://www.it.uc3m.es

Transport Transport Area Working Group Internet-Draft I-D This document describes the L4S architecture for the provision of a new service that the Internet could provide to eventually replace best efforts for all traffic: Low Latency, Low Loss, Scalable throughput (L4S). It is becoming common for all (or most) applications being run by a user at any one time to require low latency. However, the only solution the IETF can offer for ultra-low queuing delay is Diffserv, which only favours a minority of packets at the expense of others. In extensive testing the new L4S service keeps average queuing delay under a millisecond for all applications even under very heavy load, without sacrificing utilization; and it keeps congestion loss to zero. It is becoming widely recognized that adding more access capacity gives diminishing returns, because latency is becoming the critical problem. Even with a high capacity broadband access, the reduced latency of L4S remarkably and consistently improves performance under load for applications such as interactive video, conversational video, voice, Web, gaming, instant messaging, remote desktop and cloud-based apps (even when all being used at once over the same access link). The insight is that the root cause of queuing delay is in TCP, not in the queue. By fixing the sending TCP (and other transports) queuing latency becomes so much better than today that operators will want to deploy the network part of L4S to enable new products and services. Further, the network part is simple to deploy - incrementally with zero-config. Both parts, sender and network, ensure coexistence with other legacy traffic. At the same time L4S solves the long-recognized problem with the future scalability of TCP throughput. This document describes the L4S architecture, briefly describing the different components and how the work together to provide the aforementioned enhanced Internet service.

It is increasingly common for all of a user's applications at any one time to require low delay: interactive Web, Web services, voice, conversational video, interactive video, instant messaging, online gaming, remote desktop and cloud-based applications. 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. When present it typically doubles the path delay from that due to the base speed-of-light. 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 capacity offers diminishing returns if latency (delay) is not addressed. Differentiated services (Diffserv) offers Expedited Forwarding for some packets at the expense of others, but this is not applicable when all (or most) of a user's applications require low latency. Therefore, the goal is an Internet service with ultra-Low queueing Latency, ultra-Low Loss and Scalable throughput (L4S) - for all traffic. A service for all traffic will need none of the configuration or management baggage (traffic policing, traffic contracts) associated with favouring some packets over others. This document describes the L4S architecture for achieving that goal. It must be said that queuing delay only degrades performance infrequently . It only occurs 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 when the low latency application is itself a large capacity-seeking flow (e.g. interactive video). At these times, the performance improvement must be so remarkable that network operators will 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 TCP congestion control (Reno ) or compatible variants (e.g. TCP Cubic ). We shall call this family of congestion controls 'Classic' TCP. It has been demonstrated that if the sending host replaces Classic TCP with a 'Scalable' alternative, when a suitable AQM is deployed in the network the performance under load of all the above interactive applications can be stunningly improved. For instance, queuing delay under heavy load with the example DCTCP/DualQ solution cited below is roughly 1 millisecond (1 ms) at the 99th percentile without losing link utilization. This compares with 5 to 20 ms on average with a Classic TCP and current state-of-the-art AQMs such as fq_CoDel  or PIE . Also, with a Classic TCP, 5 ms of queuing is usually only possible by losing some utilization. It has been convincingly demonstrated that it is possible to deploy such an L4S service alongside the existing best efforts service so that all of a user's applications can shift to it when their 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 a single node should give nearly all the benefit. The L4S approach requires a number of mechanisms in different parts of the Internet to fulfill its goal. This document presents the L4S architecture, by describing the different components and how they interact to provide the scalable low-latency, low-loss, Internet service.

There are three main components to the L4S architecture (illustrated in ): The L4S service traffic needs to be isolated from the queuing latency of the Classic service traffic. However, the two should be able to freely share a common pool of capacity. This is because there is no way to predict how many flows at any one time might use each service and capacity in access networks is too scarce to partition into two. So a 'semi-permeable' membrane is needed that partitions latency but not bandwidth. The Dual Queue Coupled AQM is an example of such a semi-permeable membrane.Per-flow queuing such as in could be used, but it partitions both latency and bandwidth between every end-to-end flow. So it is rather overkill, which brings disadvantages (see ), not least that thousands of queues are needed when two are sufficient. A host needs to distinguish L4S and Classic packets with an identifier so that the network can classify them into their separate treatments. considers various alternative identifiers, and concludes that all alternatives involve compromises, but the ECT(1) codepoint of the ECN field is a workable solution. Scalable congestion controls already exist. They solve the scaling problem with TCP first pointed out in . The one used most widely (in controlled environments) is Data Centre TCP (DCTCP ), which has been implemented and deployed in Windows Server Editions (since 2012), in Linux and in FreeBSD. Although DCTCP as-is 'works' well over the public Internet, most implementations lack certain safety features that will be necessary once it is used outside controlled environments like data centres (see later). A similar scalable congestion control will also need to be transplanted into protocols other than TCP (SCTP, RTP/RTCP, RMCAT, etc.)

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in . In this document, these words will appear with that interpretation only when in ALL CAPS. Lower case uses of these words are not to be interpreted as carrying RFC-2119 significance. COMMENT: Since this will be an information document, This should be removed. The 'Classic' service is intended for all the congestion control behaviours that currently co-exist with TCP Reno (e.g. TCP Cubic, Compound, SCTP, etc). The 'L4S' service is intended for traffic from scalable TCP algorithms such as Data Centre TCP. But it is also more general—it will allow a set of congestion controls with similar scaling properties to DCTCP (e.g. Relentless ) to evolve.Both Classic and L4S services can cope with a proportion of unresponsive or less-responsive traffic as well (e.g. DNS, VoIP, etc). A congestion control where flow rate is inversely proportional to the level of congestion signals. Then, as flow rate scales, the number of congestion signals per round trip remains invariant, maintaining the same degree of control. For instance, DCTCP averages 2 congestion signals per round-trip whatever the flow rate. A congestion control with a flow rate compatible with standard TCP Reno . With Classic congestion controls, as capacity increases enabling higher flow rates, the number of round trips between congestion signals (losses or ECN marks) rises in proportion to the flow rate. So control of queuing and/or utilization becomes very slack. For instance, with 1500 B packets and an RTT of 18 ms, as TCP Reno flow rate increases from 2 to 100 Mb/s the number of round trips between congestion signals rises proportionately, from 2 to 100. The default congestion control in Linux (TCP Cubic) is Reno-compatible for most scenarios expected for some years. For instance, with a typical domestic round-trip time (RTT) of 18ms, TCP Cubic only switches out of Reno-compatibility mode once the flow rate approaches 1 Gb/s. For a typical data centre RTT of 1 ms, the switch-over point is theoretically 1.3 Tb/s. However, with a less common transcontinental RTT of 100 ms, it only remains Reno-compatible up to 13 Mb/s. All examples assume 1,500 B packets. The original proposed standard Explicit Congestion Notification (ECN) protocol , which requires ECN signals to be treated the same as drops, both when generated in the network and when responded to by the sender. 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 generalisation.

The L4S architecture is composed by the following elements. Protocols:The L4S architecture encompass the two protocol changes that we describe next: recommends ECT(1) is used as the identifier to classify L4S and Classic packets into their separate treatments, as required by . An essential aspect of a scalable congestion control is the use of explicit congestion signals rather than losses, because the signals need to be sent immediately and frequently—too often to use drops. 'Classic' ECN requires an ECN signal to be treated the same as a drop, both when it is generated in the network and when it is responded to by hosts. L4S allows networks and hosts to support two separate meanings for ECN. So the standards track will need to be updated to allow ECT(1) packets to depart from the 'same as drop' constraint. has been prepared as a standards track update to relax specific requirements in RFC 3168 (and certain other standards track RFCs), which clears the way for the above experimental changes proposed for L4S. also obsoletes the original experimental assignment of the ECT(1) codepoint as an ECN nonce (it was never deployed, and it offers no security benefit now that deployment is optional). Network components:The Dual Queue Coupled AQM has been specified as generically as possible as a 'semi-permeable' membrane without specifying the particular AQMs to use in the two queues. An informational appendix of the draft is provided for pseudocode examples of different possible AQM approaches. Initially a zero-config variant of RED called Curvy RED was implemented, tested and documented. The aim is for designers to be free to implement diverse ideas. So the brief normative body of the draft only specifies the minimum constraints an AQM needs to comply with to ensure that the L4S and Classic services will coexist. For instance, a variant of PIE called Dual PI Squared has been implemented and found to perform better over a wide range of conditions, so it has been documented in a second appendix of . Host mechanisms: The L4S architecture includes a number of mechanisms in the end host that we enumerate next: Data Centre TCP is the most widely used example of a scalable congestion control. It is being documented in the TCPM WG as an informational record of the protocol currently in use . It will be necessary to define a number of safety features for a variant usable on the public Internet. A draft list of these, known as the TCP Prague requirements, has been drawn up (see ). The list also includes some optional performance improvements. Transport protocols other than TCP use various congestion controls designed to be friendly with Classic TCP. Before they can use the L4S service, it will be necessary to implement scalable variants of each of these transport behaviours. The following standards track RFCs currently define these protocols: ECN in TCP , in SCTP , in RTP , and in DCCP . Not all are in widespread use, but those that are will eventually need to be updated to allow a different congestion response, which they will have to indicate by using the ECT(1) codepoint. Scalable variants are under consideration for some new transport protocols that are themselves under development, e.g. QUIC and certain real-time media congestion avoidandance techniques (RMCAT) protocols. ECN feedback is sufficient for L4S in some transport protocols (RTCP, DCCP) but not others: For the case of TCP, the feedback protocol for ECN embeds the assumption from Classic ECN that it is the same as drop, making it unusable for a scalable TCP. Therefore, the implementation of TCP receivers will have to be upgraded . Work to standardize more accurate ECN feedback for TCP (AccECN ) is already in progress. ECN feedback is only roughly sketched in an appendix of the SCTP specification. A fuller specification has been proposed , which would need to be implemented and deployed before SCTCP could support L4S.

Explicit congestion signalling is a key part of the L4S approach. In contrast, use of drop as a congestion signal creates a tension because drop is both a useful signal (more would reduce delay) and an impairment (less would reduce delay). 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 the queue can be kept short whatever the load. Whereas state-of-the-art AQMs have to introduce very high packet drop at high load to keep the queue short. Further, TCP's sawtooth reduction can be smaller, and therefore return to the operating point more often, without worrying that this causes more signals (one at the top of each smaller sawtooth). The consequent smaller amplitude sawteeth fit between a very shallow marking threshold and an empty queue, so delay variation can be very low, without risk of under-utilization. All the above makes it clear that explicit congestion signalling is only advantageous for latency if it does not have to be considered 'the same as' drop (as required with Classic ECN ). Therefore, in a DualQ AQM, the L4S queue uses a new L4S variant of ECN that is not equivalent to drop , while the Classic queue uses either classic ECN or drop, which are equivalent.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 'the same as drop' was the only compromise that could be reached. RFC 3168 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." Using just two queues is not essential to L4S (more would be possible), but it is the simplest way to isolate all the L4S traffic that keeps latency low from all the legacy Classic traffic that does not.Similarly, coupling the congestion notification between the queues is not necessarily essential, but it is a clever and simple way to allow senders to determine their rate, packet-by-packet, rather than be overridden by a network scheduler. Because otherwise a network scheduler would have to inspect at least transport layer headers, and it would have to continually assign a rate to each flow without any easy way to understand application intent. Once there are at least two separate treatments in the network, hosts need an identifier at the IP layer to distinguish which treatment they intend to use. A scalable congestion control keeps the signalling frequency high so that rate variations can be small when signalling is stable, and rate can track variations in available capacity as rapidly as possible otherwise.

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 addresses the problem of bandwidth apportionment for important traffic as well as queuing latency for delay-sensitive traffic. L4S solely addresses the problem of queuing latency (as well as loss and throughput scaling). Diffserv will still be necessary where important traffic requires priority (e.g. for commercial reasons, or for protection of critical infrastructure traffic). Nonetheless, if there are Diffserv classes for important traffic, the L4S approach can provide low latency for all traffic within each Diffserv class (including the case where there is only one Diffserv class).Also, as already explained, Diffserv only works for a small subset of the traffic on a link. It is not applicable when all the applications in use at one time at a single site (home, small business or mobile device) require low latency. Also, because L4S is for all traffic, it needs none of the management baggage (traffic policing, traffic contracts) associated with favouring some packets over others. This baggage has held Diffserv back from widespread end-to-end deployment. AQMs such as PIE and fq_CoDel give a significant reduction in queuing delay relative to no AQM at all. The L4S work is intended to complement these AQMs, and we definitely do not want to distract from the need to deploy them as widely as possible. Nonetheless, without addressing the large saw-toothing rate variations of Classic congestion controls, AQMs alone cannot reduce queuing delay too far without significantly reducing link utilization. The L4S approach resolves this tension by ensuring hosts can minimize the size of their sawteeth without appearing so aggressive to legacy flows that they starve. Similarly per-flow queuing is not incompatible with the L4S approach. However, one queue for every flow can be thought of as overkill compared to the minimum of two queues for all traffic needed for the L4S approach. The overkill of per-flow queuing has side-effects: fq makes high performance networking equipment costly (processing and memory) - in contrast dual queue code can be very simple; fq requires packet inspection into the end-to-end transport layer, which doesn't sit well alongside encryption for privacy - in contrast a dual queue only operates at the IP layer; fq isolates the queuing of each flow from the others and it prevents any one flow from consuming more than 1/N of the capacity. In contrast, all L4S flows are expected to keep the queue shallow, and policing of individual flows to enforce this may be applied separately, as a policy choice. An fq scheduler has to decide packet-by-packet which flow to schedule without knowing application intent. Whereas a separate policing function can be configured less strictly, so that senders can still control the instantaneous rate of each flow dependent on the needs of each application (e.g. variable rate video), giving more wriggle-room before a flow is deemed non-compliant. Also policing of queuing and of flow-rates can be applied independently. Yet again, L4S is not an alternative to ABE but a complement that introduces much lower queuing delay. ABE alters the host behaviour in response to ECN marking to utilize a link better and give ECN flows a faster throughput, but it 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 far without losing link utilization (for other non-ABE flows).

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 will immediately experience significantly better quality of experience under load in the best effort class: Gaming VoIP Video conferencing Web browsing (Adaptive) video streaming 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: Cloud based interactive video 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. A panoramic video of a football stadium can be swiped and pinched so that on the fly a proxy in the cloud generates a sub-window of the match video under the finger-gesture control of each user. At the same time, a virtual reality headset fed from a 360 degree camera in a racing car has been demonstrated, where the user's head movements control the scene generated in the cloud. In both cases, with 7 ms end-to-end base delay, the additional queuing delay of roughly 1 ms is so low that it seems the video is generated locally. See https://riteproject.eu/dctth/ for videos of these demonstrations. Using a swiping finger gesture or head movement to pan a video are extremely demanding applications—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). If low network delay is not available, all fine interaction has to be done locally and therefore much more redundant data has to be downloaded. When all interactive processing can be done in the cloud, only the data to be rendered for the end user needs to be sent. Whereas, once applications can rely on minimal queues in the network, they can focus on reducing their own latency by only minimizing the application send queue.

The following use-cases for L4S are being considered by various interested parties: Where the bottleneck is one of various types of access network: DSL, cable, mobile, satellite Radio links (cellular, WiFi) that are distant from the source are particularly challenging. The radio link capacity can vary rapidly by orders of magnitude, so it is often desirable to hold a buffer to utilise sudden increases of capacity; cellular networks are further complicated by a perceived need to buffer in order to make hand-overs imperceptible; Satellite networks generally have a very large base RTT, so even with minimal queuing, overall delay can never be extremely low; Nonetheless, it is certainly desirable not to hold a buffer purely because of the sawteeth of Classic TCP, when it is more than is needed for all the above reasons. Private networks of heterogeneous data centres, where there is no single administrator that can arrange for all the simultaneous changes to senders, receivers and network 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); query (DNS/LDAP). Where low delay quality of service is required, but without inspecting or intervening above the IP layer : 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 gives favourable queuing to all traffic. If queuing delay is minimized, applications with a fixed delay budget can communicate over longer distances, or via a longer chain of service functions or onion routers.

The DualQ is, in itself, an incremental deployment framework for L4S AQMs so that L4S traffic can coexist with existing Classic "TCP-friendly" traffic. explains why only deploying AQM in one node at each end of the access link will realize nearly all the benefit. L4S involves both end systems and the network, so suggests some typical sequences to deploy each part, and why there will be an immediate and significant benefit after deploying just one part. If an ECN-enabled DualQ AQM has not been deployed at a bottleneck, an L4S flow is required to include a fall-back strategy to Classic behaviour. describes how an L4S flow detects this, and how to minimize the effect of false negative detection.

Nonetheless, DualQ AQMs will not have to be deployed throughout the Internet before L4S will work for 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. But 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 (). Also, this known-bottleneck case tends to be true whatever the access link technology; whether xDSL, cable, cellular, line-of-sight wireless or satellite. Therefore, the full benefit of the L4S service should be available in the downstream direction when the DualQ AQM is deployed at the ingress to this bottleneck link (or links for multihomed sites). And similarly, the full upstream service will be available once the DualQ is deployed at the upstream ingress.

Deployment in mesh topologies depends on how over-booked 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 the DualQ AQM at the edge bottlenecks. For example, some datacentre networks are designed with the bottleneck in the hypervisor or host NICs, while others bottleneck at the top-of-rack switch (both the output ports facing hosts and those facing the core). The DualQ would eventually also 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 datacentres.

For any one L4S flow to work, it requires 3 parts to have been deployed. This was the same deployment problem that ECN faced so we have learned from this. FIrstly, L4S deployment exploits the fact that DCTCP already exists on many Internet hosts (Windows, FreeBSD and Linux); both servers and clients. Therefore, just deploying DualQ AQM at a network bottleneck immediately gives a working deployment of all the L4S parts. DCTCP needs some safety concerns to be fixed for general use over the public Internet (see ), 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. If, in contrast, 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 intially network operators can enable L4S exclusively for certain customers or certain applications. But 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. Operators could do this anyway, even if it were not blessed by the IETF. However, it is best for the IETF to specify that they must use their own local identifier in combination with the IETF's identifier. Then, if an operator enables the optional local-use approach, they only have to remove this extra rule to make the service work Internet-wide - it will already traverse middleboxes, peerings, etc.

illustrates some example sequences in which the parts of L4S might be deployed. It consists of the following stages: Here, the immediate benefit of a single AQM deployment can be seen, but limited to a controlled trial or controlled deployment. In this example downstream deployment is first, but in other scenarios the upstream might be go first. The DualQ AQM also greatly improves the downstream Classic service, assuming no other AQM has already been deployed. In this stage, the name 'TCP Prague' is used to represent a variant of DCTCP that is safe to use in a production environment. If the application is primarily unidirectional, 'TCP Prague' is only needed at one end. Accurate ECN feedback (AccECN) 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, BBR . The two ends can be deployed in either order, because TCP Prague only enables itself if it has negotiated the use of AccECN feedback with the other end during the connection handshake. Thus, deployment on both ends (and in some cases only one) enables L4S trials to move to a production service, in one direction. This stage might be further motivated by performance improvements between DCTCP and TCP Prague . This is a two-move stage to enable L4S upstream. The DualQ 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 services provides this motivation. The DualQ AQM also greatly improves the upstream Classic service, assuming 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, RMCAT; a body such as the 3GPP might require L4S to be implemented in 5G user equipment, or other random acts of kindness.

If L4S is enabled between two hosts but there is no L4S AQM at the bottleneck, any drop from the bottleneck will trigger the L4S sender to fall back to a 'TCP-Friendly' behaviour (Requirement #4.1 in ). Unfortunately, as well as protecting legacy traffic, this rule degrades the L4S service whenever there is a loss, even if the loss was not from a non-DualQ bottleneck (false negative). And unfortunately, prevalent drop can be due to other causes, e.g.: congestion loss at other transient bottlenecks, e.g. due to bursts in shallower queues; transmission errors, e.g. due to electrical interference; rate policing. Three complementary approaches are in progress, but they are all currently research: In TCP Prague, use a similar approach to BBR to ignore selected losses. This could mask any of the above types of loss (requires consensus on how to safely interoperate with drop-based congestion controls). A combination of RACK, reconfigured link retransmission and L4S could address transmission errors (no reference yet); Hybrid ECN/drop policers (see ). 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 will continually widen L4S applicability. In recent studies there has been no evidence of Classic ECN support in AQMs on the Internet. If Classic ECN support does materialize, a way to satisfy Requirement #4.2 in will have to be added to TCP Prague.

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 correct propagation of the ECN field up the layers .

This specification contains no IANA considerations.

Because the L4S service can serve all traffic that is using the capacity of a link, it should not be necessary to police access to the L4S service. In contrast, Diffserv only works if some packets get less favourable treatement than others. So it has to use traffic policers to limit how much traffic can be favoured, In turn, traffic policers require traffic contracts between users and networks as well as pairwise between networks. Because L4S will lack all this management complexity, it is more likely to work end-to-end. During early deployment (and perhaps always), some networks will not offer the L4S service. These networks do not need to police or re-mark L4S traffic - they just forward it unchanged as best efforts traffic, as they would already forward traffic with ECT(1) today. At a bottleneck, such networks will introduce some queuing and dropping. When a scalable congestion control detects a drop it will have to respond as if it is a Classic congestion control (see item 3-1 in ). This will ensure safe interworking with other traffic at the 'legacy' bottleneck, but it will degrade the L4S service to no better (but never worse) than classic best efforts, whenever a legacy (non-L4S) bottleneck is encountered on a path. Certain network operators might choose to restict access to the L4S class, perhaps only to customers who have paid a premium. Their packet classifer (item 2 in ) could identify such customers against some other field (e.g. source address range) as well as ECN. If only the ECN L4S identifier matched, but not the source address (say), the classifier could direct these packets (from non-paying customers) into the Classic queue. Allowing operators to use an additional local classifier is intended to remove any incentive to bleach 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 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 requires.

The L4S service does rely on self-constraint - not in terms of limiting capacity usage, but in terms of limiting burstiness. It is hoped that standardisation of dynamic behaviour (cf. TCP slow-start) and self-interest 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. Whether burst policing becomes necessary remains to be seen. Without it, there will be potential for attacks on the low latency of the L4S service. However it may only be necessary to apply such policing reactively, e.g. punitively targeted at any deployments of new bursty malware.

As mentioned in , 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. Then, within such Diffserv classes, L4S would often be applicable to give traffic low latency and low loss. WIthin such a 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 used to partition shared capacity. A classic rate policer drops any packets exceeding a set rate, usually also giving a burst allowance (variant exist where the policer re-marks non-compliant traffic to a discard-eligible Diffserv codepoint, so they may be dropped elsewhere during contention). In networks that deploy L4S and use rate policers, it will be preferable to deploy a policer designed to be more friendly to the L4S service, This is currently a research area. 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. This could be applied to various types of policer, e.g. , or the local (non-ConEx) variant of the ConEx congestion policer . Otherwise, whenever L4S traffic encounters a rate policer, it will experience drops and the source will fall back to a Classic congestion control, thus losing all the benefits of L4S. Further discussion of the applicability of L4S to the various Diffserv classes, and the design of suitable L4S rate policers.

Receiving hosts can fool a sender into downloading faster by suppressing feedback of ECN marks (or of losses if retransmissions are not necessary or available otherwise). proposes that a TCP sender could pseudorandomly set either of ECT(0) or ECT(1) in each packet of a flow and remember the sequence it had set, termed the ECN nonce. If the receiver supports the nonce, it can prove that it is not suppressing feedback by reflecting its knowledge of the sequence back to the sender. The nonce was proposed on the assumption that receivers might be more likely to cheat congestion control than senders (although senders also have a motive to cheat). If L4S uses the ECT(1) codepoint of ECN for packet classification, it will have to obsolete the experimental nonce. As far as is known, the ECN Nonce has never been deployed, and it was only implemented for a couple of testbed evaluations. It would be nearly impossible to deploy now, because any misbehaving receiver can simply opt-out, which would be unremarkable given all receivers currently opt-out. Other ways to protect TCP feedback integrity have since been developed. For instance: the sender can test the integrity of the receiver's feedback by occasionally setting the IP-ECN field to a value normally only set by the network. Then it can test whether the receiver's feedback faithfully reports what it expects . This method consumes no extra codepoints. It works for loss and it will work for ECN feedback in any transport protocol suitable for L4S. However, it shares the same assumption as the nonce; that the sender is not cheating and it is motivated to prevent the receiver cheating; A network can enforce a congestion response to its ECN markings (or packet losses) by auditing congestion exposure (ConEx) . Whether the receiver or a downstream network is suppressing congestion feedback or the sender is unresponsive to the feedback, or both, ConEx audit can neutralise any advantage that any of these three parties would otherwise gain. ConEx is only currently defined for IPv6 and consumes a destination option header. It has been implemented, but not deployed as far as is known.

Thanks to Wes Eddy, Karen Nielsen and David Black for their useful review comments.

PSC Bell Labs Simula Research Lab Bell Labs BT (Under submission) Bell Labs Simula Research Lab Bell Labs BT BT Bell Labs Simula Nokia Research Centre Google Google Google Google Google

This list contains a list of features, mechanisms and modifications from currently defined behaviour for scalable Transport protocols so that they can be safely deployed over the public Internet. This list of requirements was produced at an ad hoc meeting during IETF-94 in Prague . One of such scalable transport protocols is DCTCP, currently specified in . In its current form, DCTCP is specified to be deployable in controlled environments and deploying it in the public Internet would lead to a number of issues, both from the safety and the performance perspective. In this section, we describe the modifications and additional mechanisms that are required for its deployment over the global Internet. We use DCTCP as a base, but it is likely that most of these requirements equally apply to other scalable transport protocols. We next provide a brief description of each required feature. Requirement #4.1: Fall back to Reno/Cubic congestion control on packet loss. Description: In case of packet loss, the scalable transport MUST react as classic TCP (whatever the classic version of TCP is running in the host, e.g. Reno, Cubic). Motivation: As part of the safety conditions for deploying a scalable transport over the public Internet is to make sure that it behaves properly when some or all the network devices connecting the two endpoints that implement the scalable transport have not been upgraded. In particular, it may be the case that some of the switches along the path between the two endpoints may only react to congestion by dropping packets (i.e. no ECN marking). It is important that in these cases, the scalable transport react to the congestion signal in the form of a packet drop similarly to classic TCP. In the particular case of DCTCP, the current DCTCP specification states that "It is RECOMMENDED that an implementation deal with loss episodes in the same way as conventional TCP." For safe deployment in the public Internet of a scalable transport, the above requirement needs to be defined as a MUST. Packet loss, while rare, may also occur in the case that the bottleneck is L4S capable. In this case, the sender may receive a high number of packets marked with the CE bit set and also experience a loss. Current DCTCP implementations react differently to this situation. At least one implementation reacts only to the drop signal (e.g. by halving the CWND) and at least another DCTCP implementation reacts to both signals (e.g. by halving the CWND due to the drop and also further reducing the CWND based on the proportion of marked packet). We believe that further experimentation is needed to understand what is the best behaviour for the public Internet, which may or not be one of the existent implementations. Requirement #4.2: Fall back to Reno/Cubic congestion control on classic ECN bottlenecks. Description: The scalable transport protocol SHOULD/MAY? behave as classic TCP with classic ECN if the path contains a legacy bottleneck which marks both ect(0) and ect(1) in the same way as drop (non L4S, but ECN capable bottleneck). Motivation: Similarly to Requirement #3.1, this requirement is a safety condition in case L4S-capable endpoints are communicating over a path that contains one or more non-L4S but ECN capable switches and one of them happens to be the bottleneck. In this case, the scalable transport will attempt to fill in the buffer of the bottleneck switch up to the marking threshold and produce a small sawtooth around that operation point. The result is that the switch will set its operation point with the buffer full and all other non-scalable transports will be starved (as they will react reducing their CWND more aggressively than the scalable transport). Scalable transports then MUST be able to detect the presence of a classic ECN bottleneck and fall back to classic TCP/classic ECN behaviour in this case. Discussion: It is not clear at this point if it is possible to design a mechanism that always detect the aforementioned cases. One possibility is to base the detection on an increase on top of a minimum RTT, but it is not yet clear which value should trigger this. Having a delay based fall back response on L4S may as well be beneficial for preserving low latency without legacy network nodes. Even if it possible to design such a mechanism, it may well be that it would encompass additional complexity that implementers may consider unnecessary. The need for this mechanism depends on the extent of classic ECN deployment. Requirement #4.3: Reduce RTT dependence Description: Scalable transport congestion control algorithms MUST reduce or eliminate the RTT bias within the range of RTTs available. Motivation: Classic TCP's throughput is known to be inversely proportional to RTT. One would expect flows over very low RTT paths to nearly starve flows over larger RTTs. However, because Classic TCP induces a large queue, it has never allowed a very low RTT path to exist, so far. For instance, consider two paths with base RTT 1ms and 100ms. If Classic TCP induces a 20ms queue, it turns these RTTs into 21ms and 120ms leading to a throughput ratio of about 1:6. Whereas if a Scalable TCP induces only a 1ms queue, the ratio is 2:101. Therefore, with small queues, long RTT flows will essentially starve. Scalable transport protocol MUST then accommodate flows across the range of RTTs enabled by the deployment of L4S service over the public Internet. Requirement #4.4: Scaling down the congestion window. Description: Scalable transports MUST be responsive to congestion when RTTs are significantly smaller than in the current public Internet. Motivation: As currently specified, the minimum CWND of TCP (and the scalable extensions such as DCTCP), is set to 2 MSS. Once this minimum CWND is reached, the transport protocol ceases to react to congestion signals (the CWND is not further reduced beyond this minimum size). L4S mechanisms reduce significantly the queueing delay, achieving smaller RTTs over the Internet. For the same CWND, smaller RTTs imply higher transmission rates. The result is that when scalable transport are used and small RTTs are achieved, the minimum value of the CWND currently defined in 2 MSS may still result in a high transmission rate for a large number of common scenarios. For example, as described in , consider a residential setting with an broadband Internet access of 40Mbps. Suppose now a number of equal TCP flows running in parallel with the Internet access link being the bottleneck. Suppose that for these flows, the RTT is 6ms and the MSS is 1500B. The minimum transmission rate supported by TCP in this scenario is when CWND is set to 2 MSS, which results in 4Mbps for each flow. This means that in this scenario, if the number of flows is higher than 10, the congestion control ceases to be responsive and starts to build up a queue in the network. In order to address this issue, the congestion control mechanism for scalable transports MUST be responsive for the new range of RTT resulting from the decrease of the queueing delay. There are several ways how this can be achieved. One possible sub-MSS window mechanism is described in . In addition to the safety requirements described before, there are some optimizations that while not required for the safe deployment of scalable transports over the public Internet, would results in an optimized performance. We describe them next. Optimization #5.1: Setting ECT in SYN, SYN/ACK and pure ACK packets. Description: Scalable transport SHOULD set the ECT bit in SYN, SYN/ACK and pure ACK packets. Motivation: Failing to set the ECT bit in SYN, SYN/ACK or ACK packets results in these packets being more likely dropped during congestion events. Dropping SYN and SYN/ACK packets is particularly bad for performance as the retransmission timers for these packets are large. prevents from marking these packets due to security reasons. The arguments provided should be revisited in the the context of L4S and evaluate if avoiding marking these packets is still the best approach. Optimization #5.2: Faster than additive increase. Description: Scalable transport MAY support faster than additive increase in the congestion avoidance phase. Motivation: As currently defined, DCTCP supports additive increase in congestion avoidance phase. It would be beneficial for performance to update the congestion control algorithm to increase the CWND more than 1 MSS per RTT during the congestion avoidance phase. In the context of L4S such mechanism, must also provide fairness with other classes of traffic, including classic TCP and possibly scalable TCP that uses additive increase. Optimization #5.3: Faster convergence to fairness. Description: Scalable transport SHOULD converge to a fair share allocation of the available capacity as fast as classic TCP or faster. Motivation: The time required for a new flow to obtain its fair share of the capacity of the bottleneck when the there are already ongoing flows using up all the bottleneck capacity is higher in the case of DCTCP than in the case of classic TCP (about a factor of 1,5 and 2 larger according to ). This is detrimental in general, but it is very harmful for short flows, which performance can be worse than the one obtained with classic TCP. for this reason it is desirable that scalable transport provide convergence times no larger than classic TCP.

The following table includes all the itmes that should be standardized to provide a full L4S architecture. The table is too wide for the ASCII draft format, so it has been split into two, with a common column of row index numbers on the left. The columns in the second part of the table have the following meanings: The IETF WG most relevant to this requirement. The "tcpm/iccrg" combination refers to the procedure typically used for congestion control changes, where tcpm owns the approval decision, but uses the iccrg for expert review ; Applicable to all forms of TCP congestion control; Applicable to Data Centre TCP as currently used (in controlled environments); Applicable to an future Data Centre TCP congestion control intended for controlled environments; Applicable to a Scalable variant of XXX (TCP/SCTP/RMCAT) congestion control. Req # Requirement Reference 0 ARCHITECTURE 1 L4S IDENTIFIER 2 DUAL QUEUE AQM 3 Suitable ECN Feedback , . SCALABLE TRANSPORT - SAFETY ADDITIONS 4-1 Fall back to Reno/Cubic on loss 4-2 Fall back to Reno/Cubic if classic ECN bottleneck detected 4-3 Reduce RTT-dependence 4-4 Scaling TCP's Congestion Window for Small Round Trip Times SCALABLE TRANSPORT - PERFORMANCE ENHANCEMENTS 5-1 Setting ECT in SYN, SYN/ACK and pure ACK packets draft-bagnulo-tsvwg-generalized-ECN 5-2 Faster-than-additive increase 5-3 Less drastic exit from slow-start # WG TCP DCTCP DCTCP-bis TCP Prague SCTP Prague RMCAT Prague 0 tsvwg? Y Y Y Y Y Y 1 tsvwg? Y Y Y Y 2 aqm? n/a n/a n/a n/a n/a n/a 3 tcpm Y Y Y Y n/a n/a 4-1 tcpm Y Y Y Y Y 4-2 tcpm/ iccrg? Y Y ? 4-3 tcpm/ iccrg? Y Y Y ? 4-4 tcpm Y Y Y Y Y ? 5-1 tsvwg Y Y Y Y n/a n/a 5-2 tcpm/ iccrg? Y Y Y ? 5-3 tcpm/ iccrg? Y Y Y ?