| Transport Services (tsv) | K. De Schepper |
| Internet-Draft | Nokia Bell Labs |
| Intended status: Experimental | B. Briscoe, Ed. |
| Expires: May 7, 2020 | Independent |
| November 4, 2019 |
Identifying Modified Explicit Congestion Notification (ECN) Semantics for Ultra-Low Queuing Delay (L4S)
draft-ietf-tsvwg-ecn-l4s-id-08
This specification defines the identifier to be used on IP packets for a new network service called low latency, low loss and scalable throughput (L4S). It is similar to the original (or 'Classic') Explicit Congestion Notification (ECN). 'Classic' ECN marking was required to be equivalent to a drop, both when applied in the network and when responded to by a transport. Unlike 'Classic' ECN marking, for packets carrying the L4S identifier, the network applies marking more immediately and more aggressively than drop, and the transport response to each mark is reduced and smoothed relative to that for drop. The two changes counterbalance each other so that the throughput of an L4S flow will be roughly the same as a 'Classic' flow under the same conditions. However, the much more frequent control signals and the finer responses to them result in ultra-low queuing delay without compromising link utilization, and low delay is maintained during high traffic load. Examples of new active queue management (AQM) marking algorithms and examples of new transports (whether TCP-like or real-time) are specified separately. The new L4S identifier is the key piece that enables them to interwork and distinguishes them from 'Classic' traffic. It gives an incremental migration path so that existing 'Classic' TCP traffic will be no worse off, but it can be prevented from degrading the ultra-low delay and loss of the new scalable transports.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on May 7, 2020.
Copyright (c) 2019 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
This specification defines the identifier to be used on IP packets for a new network service called low latency, low loss and scalable throughput (L4S). It is similar to the original (or 'Classic') Explicit Congestion Notification (ECN [RFC3168]). 'Classic' ECN marking was required to be equivalent to a drop, both when applied in the network and when responded to by a transport. Unlike 'Classic' ECN marking, the network applies L4S marking more immediately and more aggressively than drop, and the transport response to each mark is reduced and smoothed relative to that for drop. The two changes counterbalance each other so that the throughput of an L4S flow will be roughly the same as a 'Classic' flow under the same conditions. Nonetheless, the much more frequent control signals and the finer responses to them result in ultra-low queuing delay without compromising link utilization, and low delay is maintained during high load.
An example of a scalable congestion control that would enable the L4S service is Data Center TCP (DCTCP), which until now has been applicable solely to controlled environments like data centres [RFC8257], because it is too aggressive to co-exist with existing TCP. The DualQ Coupled AQM, which is defined in a complementary experimental specification [I-D.ietf-tsvwg-aqm-dualq-coupled], is an AQM framework that enables scalable congestion controls like DCTCP to co-exist with existing traffic, each getting roughly the same flow rate when they compete under similar conditions. Note that a transport such as DCTCP is still not safe to deploy on the Internet unless it satisfies the requirements listed in Section 4. Also note that L4S is not only for elastic TCP-like traffic - there are scalable congestion controls for real-time media, such as the L4S variant of the SCReAM [RFC8298] real-time media congestion avoidance technique (RMCAT).
The new L4S identifier is the key piece that enables L4S hosts and L4S network nodes to interwork and distinguishes their traffic from 'Classic' traffic. It gives an incremental migration path so that existing 'Classic' TCP traffic will be no worse off, but it can be prevented from degrading the ultra-low delay and loss of the new scalable congestion controls. Initial implementation of the separate parts of the system has been motivated by the performance benefits.
Latency is becoming the critical performance factor for many (most?) applications on the public Internet, e.g. interactive Web, Web services, voice, conversational video, interactive video, interactive remote presence, instant messaging, online gaming, remote desktop, cloud-based applications, and video-assisted remote control of machinery and industrial processes. In the 'developed' world, further increases in access network bit-rate offer diminishing returns, whereas latency is still a multi-faceted problem. In the last decade or so, much has been done to reduce propagation time by placing caches or servers closer to users. However, queuing remains a major intermittent component of latency.
The Diffserv architecture provides Expedited Forwarding [RFC3246], so that low latency traffic can jump the queue of other traffic. However, on access links dedicated to individual sites (homes, small enterprises or mobile devices), often all traffic at any one time will be latency-sensitive. Then, given nothing to differentiate from, Diffserv makes no difference. Instead, we need to remove the causes of any unnecessary delay.
The bufferbloat project has shown that excessively-large buffering ('bufferbloat') has been introducing significantly more delay than the underlying propagation time. These delays appear only intermittently—only when a capacity-seeking (e.g. TCP) flow is long enough for the queue to fill the buffer, making every packet in other flows sharing the buffer sit through the queue.
Active queue management (AQM) was originally developed to solve this problem (and others). Unlike Diffserv, which gives low latency to some traffic at the expense of others, AQM controls latency for all traffic in a class. In general, AQM methods introduce an increasing level of discard from the buffer the longer the queue persists above a shallow threshold. This gives sufficient signals to capacity-seeking (aka. greedy) flows to keep the buffer empty for its intended purpose: absorbing bursts. However, RED [RFC2309] and other algorithms from the 1990s were sensitive to their configuration and hard to set correctly. So, this form of AQM was not widely deployed.
More recent state-of-the-art AQM methods, e.g. fq_CoDel [RFC8290], PIE [RFC8033], Adaptive RED [ARED01], are easier to configure, because they define the queuing threshold in time not bytes, so it is invariant for different link rates. However, no matter how good the AQM, the sawtoothing rate of TCP will either cause queuing delay to vary or cause the link to be under-utilized. Even with a perfectly tuned AQM, the additional queuing delay will be of the same order as the underlying speed-of-light delay across the network. Flow-queuing (e.g. [RFC8290]) can isolate one flow from another, but it cannot isolate a TCP flow from the delay variations it inflicts on itself, and it has other problems - it overrides the flow rate decisions of variable rate video applications, it does not recognise the flows within IPSec VPN tunnels and it is relatively expensive to implement.
Latency is not the only concern addressed by L4S: It was known when TCP was first developed that it would not scale to high bandwidth-delay products (footnote 6 of [TCP-CA]). Given regular broadband bit-rates over WAN distances are already [RFC3649] beyond the scaling range of 'Classic' TCP Reno, 'less unscalable' Cubic [RFC8312] and Compound [I-D.sridharan-tcpm-ctcp] variants of TCP have been successfully deployed. However, these are now approaching their scaling limits. Unfortunately, fully scalable congestion controls such as DCTCP [RFC8257] cause 'Classic' TCP to starve itself, which is why they have been confined to private data centres or research testbeds (until now).
It turns out that a TCP algorithm like DCTCP that solves the latency problem also solves TCP's scalability problem. The finer sawteeth in the congestion window have low amplitude, so they cause very little queuing delay variation and the number of sawteeth per round trip remains invariant, which maintains constant tight control as flow-rate scales. A background paper [DCttH15] gives the full explanation of why the design solves both the latency and the scaling problems, both in plain English and in more precise mathematical form. The explanation is summarised without the maths in the L4S architecture document [I-D.ietf-tsvwg-l4s-arch].
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in [RFC2119]. 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.
The new L4S identifier defined in this specification is applicable for IPv4 and IPv6 packets (as for classic ECN [RFC3168]). It is applicable for the unicast, multicast and anycast forwarding modes.
The L4S identifier is an orthogonal packet classification to the Differentiated Services Code Point (DSCP) [RFC2474]. Section 5.4 explains what this means in practice.
This document is intended for experimental status, so it does not update any standards track RFCs. Therefore it depends on [RFC8311], which is a standards track specification that:
This document is about identifiers that are used for interoperation between hosts and networks. So the audience is broad, covering developers of host transports and network AQMs, as well as covering how operators might wish to combine various identifiers, which would require flexibility from equipment developers.
This subsection briefly records the process that led to a consensus choice of L4S identifier, selected from all the alternatives in Appendix B.
The identifier for packets using the Low Latency, Low Loss, Scalable throughput (L4S) service needs to meet the following requirements:
Whether the identifier would be recoverable if the experiment failed is a factor that could be taken into account. However, this has not been made a requirement, because that would favour schemes that would be easier to fail, rather than those more likely to succeed.
It is recognised that the chosen identifier is unlikely to satisfy all these requirements, particularly given the limited space left in the IP header. Therefore a compromise will be necessary, which is why all the requirements are expressed with the word 'SHOULD' not 'MUST'. Appendix B discusses the pros and cons of the compromises made in various competing identification schemes against the above requirements.
On the basis of this analysis, "ECT(1) and CE codepoints" is the best compromise. Therefore this scheme is defined in detail in the following sections, while Appendix B records the rationale for this decision.
The L4S treatment is an experimental track alternative packet marking treatment [RFC4774] to the classic ECN treatment in [RFC3168], which has been updated by [RFC8311] to allow experiment such as the one defined in the present specification. Like classic ECN, L4S ECN identifies both network and host behaviour: it identifies the marking treatment that network nodes are expected to apply to L4S packets, and it identifies packets that have been sent from hosts that are expected to comply with a broad type of sending behaviour.
For a packet to receive L4S treatment as it is forwarded, the sender sets the ECN field in the IP header to the ECT(1) codepoint. See Section 4 for full transport layer behaviour requirements, including feedback and congestion response.
A network node that implements the L4S service normally classifies arriving ECT(1) and CE packets for L4S treatment. See Section 5 for full network element behaviour requirements, including classification, ECN-marking and interaction of the L4S identifier with other identifiers and per-hop behaviours.
A sender that wishes a packet to receive L4S treatment as it is forwarded, MUST set the ECN field in the IP header (v4 or v6) to the ECT(1) codepoint.
For a transport protocol to provide scalable congestion control it MUST provide feedback of the extent of CE marking on the forward path. When ECN was added to TCP [RFC3168], the feedback method reported no more than one CE mark per round trip. Some transport protocols derived from TCP mimic this behaviour while others report the accurate extent of TCP marking. This means that some transport protocols will need to be updated as a prerequisite for scalable congestion control. The position for a few well-known transport protocols is given below.
As a condition for a host to send packets with the L4S identifier (ECT(1)), it SHOULD implement a congestion control behaviour that ensures the flow rate is inversely proportional to the proportion of bytes in packets marked with the CE codepoint. This is termed a scalable congestion control, because the average number of control signals (ECN marks) per round trip remains roughly constant for any flow rate. As with all transport behaviours, a detailed specification (probably an experimental RFC) will need to be defined for each type of transport or application, including the timescale over which the proportionality is averaged, and control of burstiness. The inverse proportionality requirement above is worded as a 'SHOULD' rather than a 'MUST' to allow reasonable flexibility when defining these specifications.
Data Center TCP (DCTCP [RFC8257]) and the L4S variant of SCReAM [RFC8298] are examples of scalable congestion controls.
Each sender in a session can use a scalable congestion control independently of the congestion control used by the receiver(s) when they send data. Therefore there might be ECT(1) packets in one direction and ECT(0) or Not-ECT in the other.
In order to coexist safely with other Internet traffic, a scalable congestion control MUST NOT tag its packets with the ECT(1) codepoint unless it complies with the following bulleted requirements. The specification of a particular scalable congestion control MUST describe in detail how it satisfies each requirement:
As well as traffic controlled by a scalable congestion control, a reasonable level of smooth unresponsive traffic at a low rate relative to typical broadband capacities is likely to be acceptable (see "'Safe' Unresponsive Traffic" in Section 5.4.1.1.1).
A network node that implements the L4S service MUST classify arriving ECT(1) packets for L4S treatment and, other than in the exceptional case referred to next, it MUST classify arriving CE packets for L4S treatment as well. CE packets might have originated as ECT(1) or ECT(0), but the above rule to classify them as if they originated as ECT(1) is the safe choice (see Appendix B.1 for rationale). The exception is where some flow-aware in-network mechanism happens to be available for distinguishing CE packets that originated as ECT(0), as described in Section 5.3, but there is no implication that such a mechanism is necessary.
An L4S AQM treatment follows similar codepoint transition rules to those in RFC 3168. Specifically, the ECT(1) codepoint MUST NOT be changed to any other codepoint than CE, and CE MUST NOT be changed to any other codepoint. An ECT(1) packet is classified as ECN-capable and, if congestion increases, an L4S AQM algorithm will increasingly mark the ECN field as CE, otherwise forwarding packets unchanged as ECT(1). Necessary conditions for an L4S marking treatment are defined in Section 5.2. Under persistent overload an L4S marking treatment SHOULD turn off ECN marking, using drop as a congestion signal until the overload episode has subsided, as recommended for all AQM methods in [RFC7567] (Section 4.2.1), which follows the similar advice in RFC 3168 (Section 7).
For backward compatibility in uncontrolled environments, a network node that implements the L4S treatment MUST also implement a classic AQM treatment. It MUST classify arriving ECT(0) and Not-ECT packets for treatment by the Classic AQM (see the discussion of the classifier for the dual-queue coupled AQM in [I-D.ietf-tsvwg-aqm-dualq-coupled]). Classic treatment means that the AQM will mark ECT(0) packets under the same conditions as it would drop Not-ECT packets [RFC3168].
The likelihood that an AQM drops a Not-ECT Classic packet (p_C) MUST be roughly proportional to the square of the likelihood that it would have marked it if it had been an L4S packet (p_L). That is
The constant of proportionality (k) does not have to be standardised for interoperability, but a value of 2 is RECOMMENDED.
This formula ensures that Scalable and Classic flows will converge to roughly equal congestion windows. This is because the congestion windows of Scalable and Classic congestion controls are inversely proportional to p_L and sqrt(p_C) respectively. So squaring p_C in the above formula counterbalances the square root that characterizes all TCP-friendly flows.
[I-D.ietf-tsvwg-aqm-dualq-coupled] specifies the essential aspects of an L4S AQM, as well as recommending other aspects. It gives example implementations in appendices.
The term 'likelihood' is used above to allow for marking and dropping to be either probabilistic or deterministic. The example AQM methods in [I-D.ietf-tsvwg-aqm-dualq-coupled] drop and mark probabilistically, so the drop probability is arranged to be the square of the marking probability. Nonetheless, an alternative AQM that dropped and marked deterministically would be valid, as long as the dropping frequency was proportional to the square of the marking frequency.
Note that, contrary to RFC 3168, a Dual AQM implementing the L4S and Classic treatments does not mark an ECT(1) packet under the same conditions that it would have dropped a Not-ECT packet, as allowed by [RFC8311], which updates RFC 3168. However, it does mark an ECT(0) packet under the same conditions that it would have dropped a Not-ECT packet.
To implement the L4S treatment, a network node does not need to identify transport-layer flows. Nonetheless, if an implementer is willing to identify transport-layer flows at a network node, and if the most recent ECT packet in the same flow was ECT(0), the node MAY classify CE packets for classic ECN [RFC3168] treatment. In all other cases, a network node MUST classify all CE packets for L4S treatment. Examples of such other cases are: i) if no ECT packets have yet been identified in a flow; ii) if it is not desirable for a network node to identify transport-layer flows; or iii) if the most recent ECT packet in a flow was ECT(1).
If an implementer uses flow-awareness to classify CE packets, to determine whether the flow is using ECT(0) or ECT(1) it only uses the most recent ECT packet of a flow (this advice will need to be verified as part of L4S experiments). This is because a sender might switch from sending ECT(1) (L4S) packets to sending ECT(0) (Classic) packets, or back again, in the middle of a transport-layer flow (e.g. it might manually switch its congestion control module mid-connection, or it might be deliberately attempting to confuse the network).
The examples in this section concern how additional identifiers might complement the L4S identifier to classify packets between class-based queues. FIrstly two queues, L4S and Classic, as in the Coupled DualQ AQM [I-D.ietf-tsvwg-aqm-dualq-coupled], then more complex structures within a larger queuing hierarchy.
In a typical case for the public Internet a network element that implements L4S might want to classify some low-rate but unresponsive traffic (e.g. DNS, LDAP, NTP, voice, game sync packets) into the low latency queue to mix with L4S traffic. Such non-ECN-based packet types MUST be safe to mix with L4S traffic without harming the low latency service, where 'safe' is explained in Section 5.4.1.1.1 below.
In this case it would not be appropriate to call the queue an L4S queue, because it is shared by L4S and non-L4S traffic. Instead it will be called the low latency or L queue. The L queue then offers two different treatments:
To identify packets for just the scheduling treatment, it would be inappropriate to use the L4S ECT(1) identifier, because such traffic is unresponsive to ECN marking. Therefore, a network element that implements L4S MAY classify additional packets into the L queue if they carry certain non-ECN identifiers. For instance:
Of course, a packet that carried both the ECT(1) codepoint and a non-ECN identifier associated with the L queue would be classified into the L queue.
For clarity, non-ECN identifiers, such as the examples itemized above, might be used by some network operators who believe they identify non-L4S traffic that would be safe to mix with L4S traffic. They are not alternative ways for a host to indicate that it is sending L4S packets. Only the ECT(1) ECN codepoint indicates to a network element that a host is sending L4S packets (and CE indicates that it could have originated as ECT(1)). Specifically ECT(1) indicates that the host claims its behaviour satisfies the prerequisite transport requirements in Section 4.
To include additional traffic with L4S, a network element only reads identifiers such as those itemized above. It MUST NOT alter these non-ECN identifiers.
The above section requires unresponsive traffic to be 'safe' to mix with L4S traffic. Ideally this means that the sender never sends any sequence of packets at a rate that exceeds the available capacity of the bottleneck link. However, typically an unresponsive transport does not even know the bottleneck capacity of the path, let alone its available capacity. Nonetheless, an application can be considered safe enough if it paces packets out (not necessarily completely regularly) such that its maximum instantaneous rate from packet to packet stays well below a typical broadband access rate.
This is a vague but useful definition, because it encompasses many low latency applications of interest, such as DNS, voice, game sync packets, RPC, ACKs, keep-alives, etc.
To extend the above example, an operator might want to exclude some traffic from the L4S treatment for a policy reason, e.g. security (traffic from malicious sources) or commercial (e.g. initially the operator may wish to confine the benefits of L4S to business customers).
In this exclusion case, the operator MUST classify on the relevant locally-used identifiers (e.g. source addresses) before classifying the non-matching traffic on the end-to-end L4S ECN identifier.
The operator MUST NOT alter the end-to-end L4S ECN identifier from L4S to classic, because its decision to exclude certain traffic from L4S treatment is local-only. The end-to-end L4S identifier then survives for other operators to use, or indeed, they can apply their own policy, independently based on their own choice of locally-used identifiers. This approach also allows any operator to remove its locally-applied exclusions in future, e.g. if it wishes to widen the benefit of the L4S treatment to all its customers.
L4S concerns low latency, which it can provide for all traffic without differentiation and without affecting bandwidth allocation. Diffserv provides for differentiation of both bandwidth and low latency, but its control of latency depends on its control of bandwidth. The two can be combined if a network operator wants to control bandwidth allocation but it also wants to provide low latency - for any amount of traffic within one of these allocations of bandwidth (rather than only providing low latency by limiting bandwidth) [I-D.briscoe-tsvwg-l4s-diffserv].
The examples above were framed in the context of providing the default Best Efforts Per-Hop Behaviour (PHB) using two queues - a Low Latency (L) queue and a Classic (C) Queue. This single DualQ structure is expected to be by far the most common and useful arrangement. But, more generally, an operator might choose to control bandwidth allocation through a hierarchy of Diffserv PHBs at a node, and to offer one (or more) of these PHBs with a low latency and a classic variant.
In the first case, if we assume that there are no other PHBs except the DualQ, if a packet carries ECT(1) or CE, a network element would classify it for the L4S treatment irrespective of its DSCP. And, if a packet carried (say) the EF DSCP, the network element could classify it into the L queue irrespective of its ECN codepoint. However, where the DualQ is in a hierarchy of other PHBs, the classifier would classify some traffic into other PHBs based on DSCP before classifying between the latency and classic queues (based on ECT(1), CE and perhaps also the EF DSCP or other identifiers as in the above example). [I-D.briscoe-tsvwg-l4s-diffserv] gives a number of examples of such arrangements to address various requirements.
[I-D.briscoe-tsvwg-l4s-diffserv] describes how an operator might use L4S to offer low latency for all L4S traffic as well as using Diffserv for bandwidth differentiation. It identifies two main types of approach, which can be combined: the operator might split certain Diffserv PHBs between L4S and a corresponding Classic service. Or it might split the L4S and/or the Classic service into multiple Diffserv PHBs. In any of these cases, a packet would have to be classified on its Diffserv and ECN codepoints.
In summary, there are numerous ways in which the L4S ECN identifier (ECT(1) and CE) could be combined with other identifiers to achieve particular objectives. The following categorization articulates those that are valid, but it is not necessarily exhaustive. Those tagged 'Recommended-standard-use' could be set by the sending host or a network. Those tagged 'Local-use' would only be set by a network:
At a node with per-flow queueing (e.g. FQ-CoDel [RFC8290]), the L4S identifier could complement the Layer-4 flow ID as a further level of flow granularity (i.e. Not-ECT and ECT(0) queued separately from ECT(1) and CE packets). "Risk of reordering classic CE packets" in Appendix B.1 discusses the resulting ambiguity if packets originally marked ECT(0) are marked CE by an upstream AQM before they arrive at a node that classifies CE as L4S. It argues that the risk of re-ordering is vanishingly small.
Alternatively, it could be assumed that it is not in a flow's own interest to mix Classic and L4S identifiers. Then the ECN field could be used to switch between a Classic and an L4S AQM behaviour within one per-flow queue. For instance the AQM might consist of a simple marking threshold and an L4S identifier might simply select a shallower threshold than a Classic identifier would.
[I-D.ietf-tsvwg-aqm-dualq-coupled] sets operational and management requirements for experiments with DualQ Coupled AQMs. General operational and management requirements for experiments with L4S congestion controls are given in Section 4 and Section 5 above, e.g. co-existence and scaling requirements, incremental deployment arrangements. The specification of each scalable congestion control will need to include protocol-specific requirements for configuration and monitoring performance during experiments. Appendix A of [RFC5706] provides a helpful checklist.
This specification contains no IANA considerations.
Approaches to assure the integrity of signals using the new identifer are introduced in Appendix C.1. See the security considerations in the L4S architecture [I-D.ietf-tsvwg-l4s-arch] for further discussion of mis-use of the identifier.
The requirement to detect loss in time units prevents the ACK-splitting attacks described in [Savage-TCP].
Thanks to Richard Scheffenegger, John Leslie, David Täht, Jonathan Morton, Gorry Fairhurst, Michael Welzl, Mikael Abrahamsson and Andrew McGregor for the discussions that led to this specification. Ing-jyh (Inton) Tsang was a contributor to the early drafts of this document. And thanks to Mikael Abrahamsson, Lloyd Wood, Nicolas Kuhn, Greg White, Tom Henderson, David Black, Gorry Fairhurst and Brian Carpenter for providing help and reviewing this draft and to Ingemar Johansson for reviewing and providing substantial text. Appendix A listing the Prague L4S Requirements is based on text authored by Marcelo Bagnulo Braun that was originally an appendix to [I-D.ietf-tsvwg-l4s-arch]. That text was in turn based on the collective output of the attendees listed in the minutes of a 'bar BoF' on DCTCP Evolution during IETF-94 [TCPPrague].
The authors' contributions were part-funded by the European Community under its Seventh Framework Programme through the Reducing Internet Transport Latency (RITE) project (ICT-317700). Bob Briscoe was also funded partly 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.
| [RFC2119] | Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997. |
| [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. |
| [RFC4774] | Floyd, S., "Specifying Alternate Semantics for the Explicit Congestion Notification (ECN) Field", BCP 124, RFC 4774, DOI 10.17487/RFC4774, November 2006. |
| [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. |
This appendix is informative, not normative. It gives a list of modifications to current scalable congestion controls so that they can be deployed over the public Internet and coexist safely with existing traffic. The list complements the normative requirements in Section 4 that a sender has to comply with before it can set the L4S identifier in packets it sends into the Internet. As well as necessary safety improvements (requirements) this appendix also includes preferable performance improvements (optimizations).
These recommendations have become know as the Prague L4S Requirements, because they were originally identified at an ad hoc meeting during IETF-94 in Prague [TCPPrague]. The wording has been generalized to apply to all scalable congestion controls, not just TCP congestion control specifically. They were originally called the 'TCP Prague Requirements', but they are not solely applicable to TCP, so the name has been generalized, and TCP Prague is now used for a specific implementation of the requirements.
DCTCP [RFC8257] is currently the most widely used scalable transport protocol. In its current form, DCTCP is specified to be deployable only in controlled environments. Deploying it in the public Internet would lead to a number of issues, both from the safety and the performance perspective. The modifications and additional mechanisms listed in this section will be necessary for its deployment over the global Internet. Where an example is needed, DCTCP is used as a base, but it is likely that most of these requirements equally apply to other scalable congestion controls.
Description: A scalable congestion control needs to distinguish the packets it sends from those sent by classic congestion controls.
Motivation: It needs to be possible for a network node to classify L4S packets without flow state into a queue that applies an L4S ECN marking behaviour and isolates L4S packets from the queuing delay of classic packets.
Description: The transport protocol for a scalable congestion control needs to provide timely, accurate feedback about the extent of ECN marking experienced by all packets.
Motivation: Classic congestion controls only need feedback about the existence of a congestion episode within a round trip, not precisely how many packets were marked with ECN or dropped. Therefore, in 2001, when ECN feedback was added to TCP [RFC3168], it could not inform the sender of more than one ECN mark per RTT. Since then, requirements for more accurate ECN feedback in TCP have been defined in [RFC7560] and [I-D.ietf-tcpm-accurate-ecn] specifies an experimental change to the TCP wire protocol to satisfy these requirements. Most other transport protocols already satisfy this requirement.
Description: As well as responding to ECN markings in a scalable way, a scalable congestion control needs to react to packet loss in a way that will coexist safely with a TCP Reno congestion control [RFC5681].
Motivation: Part of the safety conditions for deploying a scalable congestion control on the public Internet is to make sure that it behaves properly when it builds a queue at a network bottleneck that has not been upgraded to support L4S. Packet loss can have many causes, but it usually has to be conservatively assumed that it is a sign of congestion. Therefore, on detecting packet loss, a scalable congestion control will need to fall back to classic congestion control behaviour. If it does not comply with this requirement it could starve classic traffic.
A scalable congestion control can be used for different types of transport, e.g. for real-time media or for reliable bulk transport like TCP. Therefore, the particular classic congestion control behaviour to fall back on will need to be part of the congestion control specification of the relevant transport. 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 of a scalable congestion control in the public Internet, the above requirement would need to be defined as a "MUST".
Even though a bottleneck is L4S capable, it might still become overloaded and have to drop packets. In this case, the sender may receive a high number of packets marked with the CE bit set and also experience 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). A third approach for the public Internet has been proposed that adjusts the loss response to result in a halving when combined with the ECN response. 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 these existing approaches.
Description: A scalable congestion control needs to react to ECN marking from a non-L4S, but ECN-capable, bottleneck in a way that will coexist with a TCP Reno congestion control [RFC5681].
Motivation: Similarly to the requirement in Appendix A.1.3, this requirement is a safety condition to ensure a scalable congestion control behaves properly when it builds a queue at a network bottleneck that has not been upgraded to support L4S. On detecting classic ECN marking (see below), a scalable congestion control will need to fall back to classic congestion control behaviour. If it does not comply with this requirement it could starve classic traffic.
It would take time for endpoints to distinguish classic and L4S ECN marking. An increase in queuing delay or in delay variation would be a tell-tale sign, but it is not yet clear where a line would be drawn between the two behaviours. It might be possible to cache what was learned about the path to help subsequent attempts to detect the type of marking.
Description: A scalable congestion control needs to reduce or eliminate RTT bias over as wide a range of RTTs as possible, or at least over the typical range of RTTs that will interact in the intended deployment scenario.
Motivation: Classic TCP's throughput is known to be inversely proportional to RTT, so one would expect flows over very low RTT paths to nearly starve flows over larger RTTs. However, Classic TCP has never allowed a very low RTT path to exist because it induces a large queue. For instance, consider two paths with base RTT 1ms and 100ms. If Classic TCP induces a 100ms queue, it turns these RTTs into 101ms and 200ms leading to a throughput ratio of about 2:1. Whereas if a Scalable TCP induces only a 1ms queue, the ratio is 2:101, leading to a throughput ratio of about 50:1.
Therefore, with very small queues, long RTT flows will essentially starve, unless scalable congestion controls comply with this requirement.
Description: A scalable congestion control needs to remain responsive to congestion when RTTs are significantly smaller than in the current public Internet.
Motivation: As currently specified, the minimum required congestion window of TCP (and its derivatives) is set to 2 maximum segment sizes (MSS) (see equation (4) in [RFC5681]). Once the congestion window reaches this minimum, all current TCP algorithms become unresponsive to congestion signals. No matter how much drop or ECN marking, the congestion window no longer reduces. Instead, TCP forces the queue to grow, overriding any AQM and increasing queuing delay.
L4S mechanisms significantly reduce queueing delay so, over the same path, the RTT becomes lower. Then this problem becomes surprisingly common [TCP-sub-mss-w]. This is because, for the same link capacity, smaller RTT implies a smaller window. For instance, consider a residential setting with an upstream broadband Internet access of 8 Mb/s, assuming a max segment size of 1500 B. Two upstream flows will each have the minimum window of 2 MSS if the RTT is 6ms or less, which is quite common when accessing a nearby data centre. So, any more than two such parallel TCP flows will become unresponsive and increase queuing delay.
Unless scalable congestion controls are required to comply with this requirement from the start, they will frequently become unresponsive, negating the low latency benefit of L4S, for themselves and for others. One possible sub-MSS window mechanism is described in [TCP-sub-mss-w], and other approaches are likely to be feasible.
Description: A scalable congestion control needs to detect loss by counting in time-based units, which is scalable, rather than counting in units of packets, which is not.
Motivation: A primary purpose of L4S is scalable throughput (it's in the name). Scalability in all dimensions is, of course, also a goal of all IETF technology. The inverse linear congestion response in Section 4.3 is necessary, but not sufficient, to solve the congestion control scalability problem first identified in [RFC3649]. As well as maintaining frequent ECN signals as rate scales, it is also necessary to ensure that a potentially false perception of loss does not limit throughput scaling.
End-systems cannot know whether a missing packet is due to loss or reordering, except in hindsight - if it appears later. So they can only deem that there has been a loss if a gap in the sequence space has not been filled, either after a certain number of subsequent packets has arrived (e.g. the 3 DupACK rule of standard TCP) or after a certain amount of time (e.g. the experimental RACK approach [I-D.ietf-tcpm-rack]).
As we attempt to scale packet rate over the years:
Therefore hosts have an incentive to detect loss in time units (so as not to fool themselves too often into detecting losses when there are none). And for hosts that are changing their congestion control implementation to L4S, there is no downside to including time-based loss detection code in the change (loss recovery implemented in hardware is an exception, covered later). Therefore requiring L4S hosts to detect loss in time-based units would not be a burden.
If this requirement is not placed on L4S hosts, even though it would be no burden on them to do so, all networks will face unnecessary uncertainty over whether some L4S hosts might be detecting loss by counting packets. Then all link technologies will have to unnecessarily keep reducing the time within which reordering occurs. That is not a problem for some link technologies, but it is becomes increasingly challenging for other link technologies to continue to scale, particularly those relying on channel bonding for scaling, such as LTE, 5G and DOCSIS.
Given Internet paths traverse many link technologies, any scaling limit for these more challenging access link technologies would become a scaling limit for the Internet as a whole.
It might be asked how it helps to place this loss detection requirement only on L4S hosts, because networks will still face uncertainty over whether non-L4S packets are detecting loss by counting DupACKs. The answer is that link technologies, for which it is challenging to keep squeezing the reordering time, will only need to do so for non-L4S traffic (which they can do because the L4S identifier is visible at the IP layer). Therefore, they can focus their processing and memory resources into scaling non-L4S (Classic) traffic. Then, the higher the proportion of L4S traffic, the less of a scaling challenge they will have.
To summarize, there is no reason for L4S hosts not to be part of the solution instead of part of the problem.
Requirement ("MUST") or recommendation ("SHOULD"? As explained above, this is a subtle interoperability issue between hosts and networks, which seems to need a "MUST". Unless networks can be certain that all L4S hosts follow the time-based approach, they still have to cater for the worst case - continually squeeze reordering into a smaller and smaller duration - just for hosts that might be using the counting approach. Admittedly, if this "MUST" requirement were relaxed to a "SHOULD" recommendation, networks would probably expect L4S hosts to still follow it, because following it offers only gain and no pain. But in that case, why not a "MUST"? Why make certainty appear uncertain, when it isn't?
Details:
The speed of loss recovery is much more significant for short flows than long, therefore a good compromise is to adapt the reordering window; from a small fraction of the RTT at the start of a flow, to a larger fraction of the RTT for flows that continue for many round trips.
This is broadly the approach adopted by TCP RACK (Recent ACKnowledgements) [I-D.ietf-tcpm-rack]. However, RACK starts with the 3 DupACK approach, because the RTT estimate is not necessarily stable. As long as the initial window is paced, such initial use of 3 DupACK counting would amount to time-based loss detection and therefore would satisfy the time-based loss detection requirement of L4S. This is because pacing of the initial window would ensure that 3 DupACKs early in the connection would be spread over a small fraction of the round trip.
As mentioned above, hardware implementations of loss recovery using DupACK counting exist (e.g. some implementations of RoCEv2 for RDMA). For low latency, these implementations can change their congestion control to implement L4S, because the congestion control (as distinct from loss recovery) is implemented in software. But they cannot easily satisfy this loss recovery requirement. However, it is believed they do not need to. It is believed that such implementations solely exist in controlled environments, where the network technology keeps reordering extremely low anyway. This is why the scope of the normative requirement in Section 4.3 is limited to 'reordering-prone' networks.
Detecting loss in time units also prevents the ACK-splitting attacks described in [Savage-TCP].
Description: This item only concerns TCP and its derivatives (e.g. SCTP), because the original specification of ECN for TCP precluded the use of ECN on control packets and retransmissions. To improve performance, scalable transport protocols ought to enable ECN at the IP layer in TCP control packets (SYN, SYN-ACK, pure ACKs, etc.) and in retransmitted packets. The same is true for derivatives of TCP, e.g. SCTP.
Motivation: RFC 3168 prohibits the use of ECN on these types of TCP packet, based on a number of arguments. This means these packets are not protected from congestion loss by ECN, which considerably harms performance, particularly for short flows. [I-D.ietf-tcpm-generalized-ecn] counters each argument in RFC 3168 in turn, showing it was over-cautious. Instead it proposes experimental use of ECN on all types of TCP packet as long as AccECN feedback [I-D.ietf-tcpm-accurate-ecn] is available (which is itself a prerequisite for using a scalable congestion control).
Description: It would improve performance if scalable congestion controls did not limit their congestion window increase to the traditional additive increase of 1 MSS per round trip [RFC5681] during congestion avoidance. The same is true for derivatives of TCP congestion control, including similar approaches used for real-time media.
Motivation: As currently defined, DCTCP uses the traditional TCP Reno additive increase in congestion avoidance phase. When the available capacity suddenly increases (e.g. when another flow finishes, or if radio capacity increases) it can take very many round trips to take advantage of the new capacity. In the steady state, DCTCP induces about 2 ECN marks per round trip, so it should be possible to quickly detect when these signals have disappeared and seek available capacity more rapidly. It will of course be necessary to minimize the impact on other flows (classic and scalable).
TCP Cubic was designed to solve this problem, but as flow rates have continued to increase, the delay accelerating into available capacity has become prohibitive. For instance, with RTT=20 ms, to increase flow rate from 100Mb/s to 200Mb/s Cubic takes between 50 and 100 round trips. Every 8x increase in flow rate leads to 2x more acceleration delay.
Description: Particularly when a flow starts, scalable congestion controls need to converge (reach their steady-state share of the capacity) at least as fast as classic TCP and preferably faster. This does not just affect TCP Prague, but also the flow start behaviour of any L4S congestion control derived from a Classic transport that uses TCP slow start, including those for real-time media.
Motivation: As an example, a new DCTCP flow takes longer than classic TCP to obtain its share of the capacity of the bottleneck when there are already ongoing flows using the bottleneck capacity. In a data centre environment DCTCP takes about a factor of 1.5 to 2 longer to converge due to the much higher typical level of ECN marking that DCTCP background traffic induces, which causes new flows to exit slow start early [Alizadeh-stability]. In testing for use over the public Internet the convergence time of DCTCP relative to regular TCP is even less favourable [Paced-Chirping]). It is exacerbated by the typically greater mismatch between the link rate of the sending host and typical Internet access bottlenecks, in combination with the shallow ECN marking threshold needed for L4S. This problem is detrimental in general, but would particularly harm the performance of short flows relative to classic TCP.
This appendix is informative, not normative. It records the pros and cons of various alternative ways to identify L4S packets to record the rationale for the choice of ECT(1) (Appendix B.1) as the L4S identifier. At the end, Appendix B.6 summarises the distinguishing features of the leading alternatives. It is intended to supplement, not replace the detailed text.
The leading solutions all use the ECN field, sometimes in combination with the Diffserv field. Both the ECN and Diffserv fields have the additional advantage that they are no different in either IPv4 or IPv6. A couple of alternatives that use other fields are mentioned at the end, but it is quickly explained why they are not serious contenders.
Definition:
Cons:
It is extremely unlikely that the above set of 5 eventualities that are each unusual in themselves would all happen simultaneously. But, even if they did, the consequences would hardly be dire: the odd spurious fast retransmission. Admittedly TCP reduces its congestion window when it deems there has been a loss, but even this can be recovered once the sender detects that the retransmission was spurious.
Pros:
Definition:
Use of a DSCP is the only approach for alternative ECN semantics given as an example in [RFC4774]. However, it was perhaps considered more for controlled environments than new end-to-end services.
Cons:
Pros:
Definition:
This approach would only be feasible if
Cons:
Pros:
It has been suggested that a new ID in the IPv4 Protocol field or the IPv6 Next Header field could identify L4S packets. However this approach is ruled out by numerous problems:
Locally, a network operator could arrange for L4S service to be applied based on source or destination addressing, e.g. packets from its own data centre and/or CDN hosts, packets to its business customers, etc. It could use addressing at any layer, e.g. IP addresses, MAC addresses, VLAN IDs, etc. Although addressing might be a useful tactical approach for a single ISP, it would not be a feasible approach to identify an end-to-end service like L4S. Even for a single ISP, it would require packet classifiers in buffers to be dependent on changing topology and address allocation decisions elsewhere in the network. Therefore this approach is not a feasible solution.
Table 1 provides a very high level summary of the pros and cons detailed against the schemes described respectively in Appendix B.2, Appendix B.3 and Appendix B.1, for six issues that set them apart.
| Issue | DSCP + ECN | ECN | ECT(1) + CE |
|---|---|---|---|
| initial eventual | initial | initial eventual | |
| end-to-end | N . . . ? . | . . Y | . . Y . . Y |
| tunnels | . O . . O . | . . ? | . . ? . . Y |
| lower layers | N . . . ? . | . O . | . O . . . ? |
| codepoints | N . . . . ? | . . Y | N . . . . ? |
| reordering | . . Y . . Y | . . Y | . O . . . ? |
| ctrl pkts | . . Y . . Y | . O . | . O . . . ? |
| Note 1 |
Note 1: Only feasible if classic ECN is obsolete.
The schemes are scored based on both their capabilities now ('initial') and in the long term ('eventual'). The 'ECN' scheme shares the 'eventual' scores of the 'ECT(1) + CE' scheme. The scores are one of 'N, O, Y', meaning 'Poor', 'Ordinary', 'Good' respectively. The same scores are aligned vertically to aid the eye. A score of "?" in one of the positions means that this approach might optimisitically become this good, given sufficient effort. The table summarises the text and is not meant to be understandable without having read the text.
The ECT(1) codepoint of the ECN field has already been assigned once for the ECN nonce [RFC3540], which has now been categorized as historic [RFC8311]. ECN is probably the only remaining field in the Internet Protocol that is common to IPv4 and IPv6 and still has potential to work end-to-end, with tunnels and with lower layers. Therefore, ECT(1) should not be reassigned to a different experimental use (L4S) without carefully assessing competing potential uses. These fall into the following categories:
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).
The historic ECN nonce protocol [RFC3540] proposed that a TCP sender could set either of ECT(0) or ECT(1) in each packet of a flow and remember the sequence it had set. If any packet was lost or congestion marked, the receiver would miss that bit of the sequence. An ECN Nonce receiver had to feed back the least significant bit of the sum, so it could not suppress feedback of a loss or mark without a 50-50 chance of guessing the sum incorrectly.
It is highly unlikely that ECT(1) will be needed for integrity protection in future. The ECN Nonce RFC [RFC3540] as been reclassified as historic, partly because other ways have been developed to protect TCP feedback integrity [RFC8311] that do not consume a codepoint in the IP header. For instance:
Various researchers have proposed to use ECT(1) as a less severe congestion notification than CE, particularly to enable flows to fill available capacity more quickly after an idle period, when another flow departs or when a flow starts, e.g. VCP [VCP], Queue View (QV) [QV].
Before assigning ECT(1) as an identifer for L4S, we must carefully consider whether it might be better to hold ECT(1) in reserve for future standardisation of rapid flow acceleration, which is an important and enduring problem [RFC6077].
Pre-Congestion Notification (PCN) is another scheme that assigns alternative semantics to the ECN field. It uses ECT(1) to signify a less severe level of pre-congestion notification than CE [RFC6660]. However, the ECN field only takes on the PCN semantics if packets carry a Diffserv codepoint defined to indicate PCN marking within a controlled environment. PCN is required to be applied solely to the outer header of a tunnel across the controlled region in order not to interfere with any end-to-end use of the ECN field. Therefore a PCN region on the path would not interfere with any of the L4S service identifiers proposed in Appendix B.