idnits 2.17.1 draft-ietf-tsvwg-ecn-l4s-id-06.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- No issues found here. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year == Line 1749 has weird spacing: '...initial even...' -- The document date (March 11, 2019) is 1871 days in the past. Is this intentional? Checking references for intended status: Experimental ---------------------------------------------------------------------------- == Outdated reference: A later version (-09) exists of draft-ietf-avtcore-cc-feedback-message-03 == Outdated reference: A later version (-34) exists of draft-ietf-quic-transport-18 == Outdated reference: A later version (-28) exists of draft-ietf-tcpm-accurate-ecn-08 == Outdated reference: A later version (-15) exists of draft-ietf-tcpm-generalized-ecn-03 == Outdated reference: A later version (-15) exists of draft-ietf-tcpm-rack-04 == Outdated reference: A later version (-25) exists of draft-ietf-tsvwg-aqm-dualq-coupled-08 == Outdated reference: A later version (-22) exists of draft-ietf-tsvwg-ecn-encap-guidelines-11 == Outdated reference: A later version (-20) exists of draft-ietf-tsvwg-l4s-arch-03 == Outdated reference: A later version (-07) exists of draft-stewart-tsvwg-sctpecn-05 -- Obsolete informational reference (is this intentional?): RFC 2309 (Obsoleted by RFC 7567) -- Obsolete informational reference (is this intentional?): RFC 4960 (Obsoleted by RFC 9260) -- Obsolete informational reference (is this intentional?): RFC 8312 (Obsoleted by RFC 9438) Summary: 0 errors (**), 0 flaws (~~), 11 warnings (==), 4 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Transport Services (tsv) K. De Schepper 3 Internet-Draft Nokia Bell Labs 4 Intended status: Experimental B. Briscoe, Ed. 5 Expires: September 12, 2019 CableLabs 6 March 11, 2019 8 Identifying Modified Explicit Congestion Notification (ECN) Semantics 9 for Ultra-Low Queuing Delay (L4S) 10 draft-ietf-tsvwg-ecn-l4s-id-06 12 Abstract 14 This specification defines the identifier to be used on IP packets 15 for a new network service called low latency, low loss and scalable 16 throughput (L4S). It is similar to the original (or 'Classic') 17 Explicit Congestion Notification (ECN). 'Classic' ECN marking was 18 required to be equivalent to a drop, both when applied in the network 19 and when responded to by a transport. Unlike 'Classic' ECN marking, 20 for packets carrying the L4S identifier, the network applies marking 21 more immediately and more aggressively than drop, and the transport 22 response to each mark is reduced and smoothed relative to that for 23 drop. The two changes counterbalance each other so that the 24 throughput of an L4S flow will be roughly the same as a 'Classic' 25 flow under the same conditions. However, the much more frequent 26 control signals and the finer responses to them result in ultra-low 27 queuing delay without compromising link utilization, and low delay is 28 maintained during high load. Examples of new active queue management 29 (AQM) marking algorithms and examples of new transports (whether TCP- 30 like or real-time) are specified separately. The new L4S identifier 31 is the key piece that enables them to interwork and distinguishes 32 them from 'Classic' traffic. It gives an incremental migration path 33 so that existing 'Classic' TCP traffic will be no worse off, but it 34 can be prevented from degrading the ultra-low delay and loss of the 35 new scalable transports. 37 Status of This Memo 39 This Internet-Draft is submitted in full conformance with the 40 provisions of BCP 78 and BCP 79. 42 Internet-Drafts are working documents of the Internet Engineering 43 Task Force (IETF). Note that other groups may also distribute 44 working documents as Internet-Drafts. The list of current Internet- 45 Drafts is at https://datatracker.ietf.org/drafts/current/. 47 Internet-Drafts are draft documents valid for a maximum of six months 48 and may be updated, replaced, or obsoleted by other documents at any 49 time. It is inappropriate to use Internet-Drafts as reference 50 material or to cite them other than as "work in progress." 52 This Internet-Draft will expire on September 12, 2019. 54 Copyright Notice 56 Copyright (c) 2019 IETF Trust and the persons identified as the 57 document authors. All rights reserved. 59 This document is subject to BCP 78 and the IETF Trust's Legal 60 Provisions Relating to IETF Documents 61 (https://trustee.ietf.org/license-info) in effect on the date of 62 publication of this document. Please review these documents 63 carefully, as they describe your rights and restrictions with respect 64 to this document. Code Components extracted from this document must 65 include Simplified BSD License text as described in Section 4.e of 66 the Trust Legal Provisions and are provided without warranty as 67 described in the Simplified BSD License. 69 Table of Contents 71 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3 72 1.1. Problem . . . . . . . . . . . . . . . . . . . . . . . . . 4 73 1.2. Terminology . . . . . . . . . . . . . . . . . . . . . . . 6 74 1.3. Scope . . . . . . . . . . . . . . . . . . . . . . . . . . 6 75 2. Consensus Choice of L4S Packet Identifier: Requirements . . . 7 76 3. L4S Packet Identification at Run-Time . . . . . . . . . . . . 8 77 4. Prerequisite Transport Layer Behaviour . . . . . . . . . . . 8 78 4.1. Prerequisite Codepoint Setting . . . . . . . . . . . . . 8 79 4.2. Prerequisite Transport Feedback . . . . . . . . . . . . . 8 80 4.3. Prerequisite Congestion Response . . . . . . . . . . . . 9 81 5. Prerequisite Network Node Behaviour . . . . . . . . . . . . . 11 82 5.1. Prerequisite Classification and Re-Marking Behaviour . . 11 83 5.2. The Meaning of L4S CE Relative to Drop . . . . . . . . . 11 84 5.3. Exception for L4S Packet Identification by Network Nodes 85 with Transport-Layer Awareness . . . . . . . . . . . . . 12 86 5.4. Interaction of the L4S Identifier with other Identifiers 13 87 5.4.1. Examples of Other Identifiers Complementing L4S 88 Identifiers . . . . . . . . . . . . . . . . . . . . . 13 89 5.4.1.1. Inclusion of Additional Traffic with L4S . . . . 13 90 5.4.1.2. Exclusion of Traffic From L4S Treatment . . . . . 14 91 5.4.2. Generalized Combination of L4S and Other Identifiers 15 92 6. L4S Experiments . . . . . . . . . . . . . . . . . . . . . . . 16 93 7. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 16 94 8. Security Considerations . . . . . . . . . . . . . . . . . . . 16 95 9. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 16 96 10. References . . . . . . . . . . . . . . . . . . . . . . . . . 17 97 10.1. Normative References . . . . . . . . . . . . . . . . . . 17 98 10.2. Informative References . . . . . . . . . . . . . . . . . 17 99 Appendix A. The 'Prague L4S Requirements' . . . . . . . . . . . 23 100 A.1. Requirements for Scalable Transport Protocols . . . . . . 24 101 A.1.1. Use of L4S Packet Identifier . . . . . . . . . . . . 24 102 A.1.2. Accurate ECN Feedback . . . . . . . . . . . . . . . . 24 103 A.1.3. Fall back to Reno-friendly congestion control on 104 packet loss . . . . . . . . . . . . . . . . . . . . . 24 105 A.1.4. Fall back to Reno-friendly congestion control on 106 classic ECN bottlenecks . . . . . . . . . . . . . . . 25 107 A.1.5. Reduce RTT dependence . . . . . . . . . . . . . . . . 25 108 A.1.6. Scaling down to fractional congestion windows . . . . 26 109 A.1.7. Measuring Reordering Tolerance in Time Units . . . . 27 110 A.2. Scalable Transport Protocol Optimizations . . . . . . . . 29 111 A.2.1. Setting ECT in TCP Control Packets and 112 Retransmissions . . . . . . . . . . . . . . . . . . . 29 113 A.2.2. Faster than Additive Increase . . . . . . . . . . . . 29 114 A.2.3. Faster Convergence at Flow Start . . . . . . . . . . 30 115 Appendix B. Alternative Identifiers . . . . . . . . . . . . . . 30 116 B.1. ECT(1) and CE codepoints . . . . . . . . . . . . . . . . 31 117 B.2. ECN Plus a Diffserv Codepoint (DSCP) . . . . . . . . . . 33 118 B.3. ECN capability alone . . . . . . . . . . . . . . . . . . 35 119 B.4. Protocol ID . . . . . . . . . . . . . . . . . . . . . . . 37 120 B.5. Source or destination addressing . . . . . . . . . . . . 37 121 B.6. Summary: Merits of Alternative Identifiers . . . . . . . 37 122 Appendix C. Potential Competing Uses for the ECT(1) Codepoint . 38 123 C.1. Integrity of Congestion Feedback . . . . . . . . . . . . 38 124 C.2. Notification of Less Severe Congestion than CE . . . . . 39 125 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 40 127 1. Introduction 129 This specification defines the identifier to be used on IP packets 130 for a new network service called low latency, low loss and scalable 131 throughput (L4S). It is similar to the original (or 'Classic') 132 Explicit Congestion Notification (ECN [RFC3168]). 'Classic' ECN 133 marking was required to be equivalent to a drop, both when applied in 134 the network and when responded to by a transport. Unlike 'Classic' 135 ECN marking, the network applies L4S marking more immediately and 136 more aggressively than drop, and the transport response to each mark 137 is reduced and smoothed relative to that for drop. The two changes 138 counterbalance each other so that the throughput of an L4S flow will 139 be roughly the same as a 'Classic' flow under the same conditions. 140 Nonetheless, the much more frequent control signals and the finer 141 responses to them result in ultra-low queuing delay without 142 compromising link utilization, and low delay is maintained during 143 high load. 145 An example of a scalable congestion control that would enable the L4S 146 service is Data Centre TCP (DCTCP), which until now has been 147 applicable solely to controlled environments like data centres 148 [RFC8257], because it is too aggressive to co-exist with existing 149 TCP. The DualQ Coupled AQM, which is defined in a complementary 150 experimental specification [I-D.ietf-tsvwg-aqm-dualq-coupled], is an 151 AQM framework that enables scalable congestion controls like DCTCP to 152 co-exist with existing traffic, each getting roughly the same flow 153 rate when they compete under similar conditions. Note that a 154 transport such as DCTCP is still not safe to deploy on the Internet 155 unless it satisfies the requirements listed in Section 4. Also note 156 that L4S is not only for elastic TCP-like traffic - there are 157 scalable congestion controls for real-time media, such as the L4S 158 variant of the SCReAM [RFC8298] real-time media congestion avoidance 159 technique (RMCAT). 161 The new L4S identifier is the key piece that enables L4S hosts and 162 L4S network nodes to interwork and distinguishes their traffic from 163 'Classic' traffic. It gives an incremental migration path so that 164 existing 'Classic' TCP traffic will be no worse off, but it can be 165 prevented from degrading the ultra-low delay and loss of the new 166 scalable congestion controls. The performance improvement is so 167 great that it is motivating initial deployment of the separate parts 168 of this system. 170 1.1. Problem 172 Latency is becoming the critical performance factor for many (most?) 173 applications on the public Internet, e.g. interactive Web, Web 174 services, voice, conversational video, interactive video, interactive 175 remote presence, instant messaging, online gaming, remote desktop, 176 cloud-based applications, and video-assisted remote control of 177 machinery and industrial processes. In the developed world, further 178 increases in access network bit-rate offer diminishing returns, 179 whereas latency is still a multi-faceted problem. In the last decade 180 or so, much has been done to reduce propagation time by placing 181 caches or servers closer to users. However, queuing remains a major 182 intermittent component of latency. 184 The Diffserv architecture provides Expedited Forwarding [RFC3246], so 185 that low latency traffic can jump the queue of other traffic. 186 However, on access links dedicated to individual sites (homes, small 187 enterprises or mobile devices), often all traffic at any one time 188 will be latency-sensitive. Then Diffserv is of little use. Instead, 189 we need to remove the causes of any unnecessary delay. 191 The bufferbloat project has shown that excessively-large buffering 192 ('bufferbloat') has been introducing significantly more delay than 193 the underlying propagation time. These delays appear only 194 intermittently--only when a capacity-seeking (e.g. TCP) flow is long 195 enough for the queue to fill the buffer, making every packet in other 196 flows sharing the buffer sit through the queue. 198 Active queue management (AQM) was originally developed to solve this 199 problem (and others). Unlike Diffserv, which gives low latency to 200 some traffic at the expense of others, AQM controls latency for _all_ 201 traffic in a class. In general, AQMs introduce an increasing level 202 of discard from the buffer the longer the queue persists above a 203 shallow threshold. This gives sufficient signals to capacity-seeking 204 (aka. greedy) flows to keep the buffer empty for its intended 205 purpose: absorbing bursts. However, RED [RFC2309] and other 206 algorithms from the 1990s were sensitive to their configuration and 207 hard to set correctly. So, AQM was not widely deployed. 209 More recent state-of-the-art AQMs, e.g. fq_CoDel [RFC8290], 210 PIE [RFC8033], Adaptive RED [ARED01], are easier to configure, 211 because they define the queuing threshold in time not bytes, so it is 212 invariant for different link rates. However, no matter how good the 213 AQM, the sawtoothing rate of TCP will either cause queuing delay to 214 vary or cause the link to be under-utilized. Even with a perfectly 215 tuned AQM, the additional queuing delay will be of the same order as 216 the underlying speed-of-light delay across the network. Flow-queuing 217 can isolate one flow from another, but it cannot isolate a TCP flow 218 from the delay variations it inflicts on itself, and it has other 219 problems - it overrides the flow rate decisions of variable rate 220 video applications, it does not recognise the flows within IPSec VPN 221 tunnels and it is relatively expensive to implement. 223 Latency is not our only concern: It was known when TCP was first 224 developed that it would not scale to high bandwidth-delay products 225 [TCP-CA]. Given regular broadband bit-rates over WAN distances are 226 already [RFC3649] beyond the scaling range of 'Classic' TCP Reno, 227 'less unscalable' Cubic [RFC8312] and 228 Compound [I-D.sridharan-tcpm-ctcp] variants of TCP have been 229 successfully deployed. However, these are now approaching their 230 scaling limits. Unfortunately, fully scalable congestion controls 231 such as DCTCP [RFC8257] cause 'Classic' TCP to starve itself, which 232 is why they have been confined to private data centres or research 233 testbeds (until now). 235 It turns out that a TCP algorithm like DCTCP that solves the latency 236 problem also solves TCP's scalability problem. The finer sawteeth 237 have low amplitude, so they cause very little queuing delay variation 238 and the number of sawteeth per round trip remains invariant, which 239 maintains constant tight control as flow-rate scales. A supporting 240 paper [DCttH15] gives the full explanation of why the design solves 241 both the latency and the scaling problems, both in plain English and 242 in more precise mathematical form. The explanation is summarised 243 without the maths in the L4S architecture document 244 [I-D.ietf-tsvwg-l4s-arch]. 246 1.2. Terminology 248 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 249 "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and 250 "OPTIONAL" in this document are to be interpreted as described in 251 [RFC2119]. In this document, these words will appear with that 252 interpretation only when in ALL CAPS. Lower case uses of these words 253 are not to be interpreted as carrying RFC-2119 significance. 255 Classic service: The 'Classic' service is intended for all the 256 behaviours that currently co-exist with TCP Reno (e.g. TCP Cubic, 257 Compound, SCTP, etc). 259 Low-Latency, Low-Loss and Scalable (L4S) service: The 'L4S' service 260 is intended for traffic from scalable congestion control 261 algorithms such as Data Centre TCP. But it is also more general-- 262 it allows the set of congestion controls with similar scaling 263 properties to DCTCP to evolve (e.g. Relentless TCP [Mathis09] and 264 the L4S variant of SCREAM for real-time media [RFC8298]. 266 Both Classic and L4S services can cope with a proportion of 267 unresponsive or less-responsive traffic as well, as long as it 268 does not build a queue (e.g. DNS, VoIP, game sync datagrams, 269 etc). 271 Classic ECN: The original Explicit Congestion Notification (ECN) 272 protocol [RFC3168]. 274 1.3. Scope 276 The new L4S identifier defined in this specification is applicable 277 for IPv4 and IPv6 packets (as for classic ECN [RFC3168]). It is 278 applicable for the unicast, multicast and anycast forwarding modes. 280 The L4S identifier is an orthogonal packet classification to the 281 Differentiated Services Code Point (DSCP [RFC2474]). Section 5.4 282 explains what this means in practice. 284 This document is intended for experimental status, so it does not 285 update any standards track RFCs. Therefore it depends on [RFC8311], 286 which is a standards track specification that: 288 o updates the ECN proposed standard [RFC3168] to allow experimental 289 track RFCs to relax the requirement that an ECN mark must be 290 equivalent to a drop, both when applied by the network, and when 291 responded to by the sender; 293 o changes the status of the experimental ECN nonce [RFC3540] to 294 historic; 296 o makes consequent updates to the following additional proposed 297 standard RFCs to reflect the above two bullets: 299 * ECN for RTP [RFC6679]; 301 * the congestion control specifications of various DCCP 302 congestion control identifier (CCID) profiles [RFC4341], 303 [RFC4342], [RFC5622]. 305 2. Consensus Choice of L4S Packet Identifier: Requirements 307 This subsection briefly records the process that led to a consensus 308 choice of L4S identifier, selected from all the alternatives in 309 Appendix B. 311 Ideally, the identifier for packets using the Low Latency, Low Loss, 312 Scalable throughput (L4S) service ought to meet the following 313 requirements: 315 o it SHOULD survive end-to-end between source and destination 316 applications: across the boundary between host and network, 317 between interconnected networks, and through middleboxes; 319 o it SHOULD be common to IPv4 and IPv6 and transport-agnostic; 321 o it SHOULD be incrementally deployable; 323 o it SHOULD enable an AQM to classify packets encapsulated by outer 324 IP or lower-layer headers; 326 o it SHOULD consume minimal extra codepoints; 328 o it SHOULD be consistent on all the packets of a transport layer 329 flow, so that some packets of a flow are not served by a different 330 queue to others. 332 Whether the identifier would be recoverable if the experiment failed 333 is a factor that could be taken into account. However, this has not 334 been made a requirement, because that would favour schemes that would 335 be easier to fail, rather than those more likely to succeed. 337 It is recognised that the chosen identifier is unlikely to satisfy 338 all these requirements, particularly given the limited space left in 339 the IP header. Therefore a compromise will be necessary, which is 340 why all the requirements are expressed with the word 'SHOULD' not 341 'MUST'. Appendix B discusses the pros and cons of the compromises 342 made in various competing identification schemes against the above 343 requirements. 345 On the basis of this analysis, "ECT(1) and CE codepoints" is the best 346 compromise. Therefore this scheme is defined in detail in the 347 following sections, while Appendix B records the rationale for this 348 decision. 350 3. L4S Packet Identification at Run-Time 352 The L4S treatment is an experimental track alternative packet marking 353 treatment [RFC4774] to the classic ECN treatment [RFC3168], which has 354 been updated by [RFC8311] to allow this experiment (amongst others). 355 Like classic ECN, L4S ECN identifies both network and host behaviour: 356 it identifies the marking treatment that network nodes are expected 357 to apply to L4S packets, and it identifies packets that have been 358 sent from hosts that are expected to comply with a broad type of 359 sending behaviour. 361 For a packet to receive L4S treatment as it is forwarded, the sender 362 sets the ECN field in the IP header to the ECT(1) codepoint. See 363 Section 4 for full transport layer behaviour requirements, including 364 feedback and congestion response. 366 A network node that implements the L4S service normally classifies 367 arriving ECT(1) and CE packets for L4S treatment. See Section 5 for 368 full network element behaviour requirements, including 369 classification, ECN-marking and interaction of the L4S identifier 370 with other identifiers and per-hop behaviours. 372 4. Prerequisite Transport Layer Behaviour 374 4.1. Prerequisite Codepoint Setting 376 A sender that wishes a packet to receive L4S treatment as it is 377 forwarded, MUST set the ECN field in the IP header (v4 or v6) to the 378 ECT(1) codepoint. 380 4.2. Prerequisite Transport Feedback 382 In general, a scalable congestion control needs feedback of the 383 extent of CE marking on the forward path. When ECN was added to TCP 384 [RFC3168], the feedback method reported no more than one CE mark per 385 round trip. Some transport protocols derived from TCP mimic this 386 behaviour while others report the accurate extent of TCP marking. 387 This means that some transport protocols will need to be updated as a 388 prerequisite for scalable congestion control. The position for a few 389 well-known transport protocols is given below. 391 TCP: Support for accurate ECN feedback (AccECN 392 [I-D.ietf-tcpm-accurate-ecn]) by both ends is a prerequisite for 393 scalable congestion control. Therefore, the presence of ECT(1) in 394 the IP headers even in one direction of a TCP connection will 395 imply that both ends support AccECN. However, the converse does 396 not apply. So even if both ends support AccECN, either of the two 397 ends can choose not to use a scalable congestion control, whatever 398 the other end's choice. 400 SCTP: A suitable ECN feedback mechanism for SCTP could add a chunk 401 to report the number of received CE marks (e.g. 402 [I-D.stewart-tsvwg-sctpecn]), and update the ECN feedback protocol 403 sketched out in Appendix A of the standards track specification of 404 SCTP [RFC4960]. 406 RTP over UDP: A prerequisite for scalable congestion control is for 407 both (all) ends of one media-level hop to signal ECN support 408 [RFC6679] and use the new generic RTCP feedback format of 409 [I-D.ietf-avtcore-cc-feedback-message]. The presence of ECT(1) 410 implies that both (all) ends of that hop support ECN. However, 411 the converse does not apply, so each end of a media-level hop can 412 independently choose not to use a scalable congestion control, 413 even if both ends support ECN. 415 QUIC: Support for sufficiently fine-grained ECN feedback is provided 416 by the first IETF QUIC transport [I-D.ietf-quic-transport]. 418 DCCP: The ACK vector in DCCP [RFC4340] is already sufficient to 419 report the extent of CE marking as needed by a scalable congestion 420 control. 422 4.3. Prerequisite Congestion Response 424 As a condition for a host to send packets with the L4S identifier 425 (ECT(1)), it SHOULD implement a congestion control behaviour that 426 ensures the flow rate is inversely proportional to the proportion of 427 bytes in packets marked with the CE codepoint. This is termed a 428 scalable congestion control, because the number of control signals 429 (ECN marks) per round trip remains roughly constant for any flow 430 rate. As with all transport behaviours, a detailed specification 431 will need to be defined for each type of transport or application, 432 including the timescale over which the proportionality is averaged, 433 and control of burstiness. The inverse proportionality requirement 434 above is worded as a 'SHOULD' rather than a 'MUST' to allow 435 reasonable flexibility when defining these specifications. 437 Data Center TCP (DCTCP [RFC8257]) and the L4S variant of SCReAM 438 [RFC8298] are examples of a scalable congestion controls. 440 Each sender in a session can use a scalable congestion control 441 independently of the congestion control used by the receiver(s) when 442 they send data. Therefore there might be ECT(1) packets in one 443 direction and ECT(0) or Not-ECT in the other. 445 In order to coexist safely with other Internet traffic, a scalable 446 congestion control MUST NOT tag its packets with the ECT(1) codepoint 447 unless it complies with the following bulleted requirements. The 448 specification of a particular scalable congestion control MUST 449 describe in detail how it satisfies each requirement: 451 o A scalable congestion control MUST react to packet loss in a way 452 that will coexist safely with a TCP Reno congestion control 453 [RFC5681] (see Appendix A.1.3 for rationale). 455 o A scalable congestion control MUST react to ECN marking from a 456 non-L4S but ECN-capable bottleneck in a way that will coexist with 457 a TCP Reno congestion control [RFC5681] (see Appendix A.1.4 for 458 rationale). 460 Note that a scalable congestion control is not expected to change 461 to setting ECT(0) while it temporarily falls back to coexist with 462 Reno . However an implementer who believes this would be 463 beneficial if fall-back persists, can choose to do so, 465 o A scalable congestion control MUST reduce or eliminate RTT bias 466 over as wide a range of RTTs as possible, or at least over the 467 typical range of RTTs that will interact in the intended 468 deployment scenario (see Appendix A.1.5 for rationale). 470 o A scalable congestion control MUST remain responsive to congestion 471 when the RTT is significantly smaller than in the current public 472 Internet (see Appendix A.1.6 for rationale). 474 o A scalable congestion control MUST detect loss by counting in 475 time-based units, which is scalable, as opposed to counting in 476 units of packets (as in the 3 DupACK rule of traditional TCP), 477 which is not scalable (see Appendix A.1.7 for rationale). 479 5. Prerequisite Network Node Behaviour 481 5.1. Prerequisite Classification and Re-Marking Behaviour 483 A network node that implements the L4S service MUST classify arriving 484 ECT(1) packets for L4S treatment and, other than in the exceptional 485 case referred to next, it MUST classify arriving CE packets for L4S 486 treatment as well. CE packets might have originated as ECT(1) or 487 ECT(0), but the above rule to classify them as if they originated as 488 ECT(1) is the safe choice (see Appendix B.1 for rationale). The 489 exception is where some flow-aware in-network mechanism happens to be 490 available for distinguishing CE packets that originated as ECT(0), as 491 described in Section 5.3, but there is no implication that such a 492 mechanism is necessary. 494 An L4S AQM treatment follows similar codepoint transition rules to 495 those in RFC 3168. Specifically, the ECT(1) codepoint MUST NOT be 496 changed to any other codepoint than CE, and CE MUST NOT be changed to 497 any other codepoint. An ECT(1) packet is classified as ECN-capable 498 and, if congestion increases, an L4S AQM algorithm will mark the ECN 499 field as CE for an increasing proportion of packets, otherwise 500 forwarding packets unchanged as ECT(1). Necessary conditions for an 501 L4S marking treatment are defined in Section 5.2. Under persistent 502 overload an L4S marking treatment SHOULD turn off ECN marking, using 503 drop as a congestion signal until the overload episode has subsided, 504 as recommended for all AQMs in [RFC7567] (Section 4.2.1), which 505 follows the similar advice in RFC 3168 (Section 7). 507 For backward compatibility in uncontrolled environments, a network 508 node that implements the L4S treatment MUST also implement a classic 509 AQM treatment. It MUST classify arriving ECT(0) and Not-ECT packets 510 for treatment by the Classic AQM (see the discussion of the 511 classifier for the dual-queue coupled AQM in 512 [I-D.ietf-tsvwg-aqm-dualq-coupled]). Classic treatment means that 513 the AQM will mark ECT(0) packets under the same conditions as it 514 would drop Not-ECT packets [RFC3168]. 516 5.2. The Meaning of L4S CE Relative to Drop 518 The likelihood that an AQM drops a Not-ECT Classic packet (p_C) MUST 519 be roughly proportional to the square of the likelihood that it would 520 have marked it if it had been an L4S packet (p_L) {ToDo cross-ref to 521 new section in l4s-arch that explains the rationale for the square}. 522 That is 524 p_C ~= (p_L / k)^2 526 The constant of proportionality (k) does not have to be standardised 527 for interoperability, but a value of 2 is RECOMMENDED. 529 [I-D.ietf-tsvwg-aqm-dualq-coupled] specifies the essential aspects of 530 an L4S AQM, as well as recommending other aspects. It gives example 531 implementations in appendices. 533 The term 'likelihood' is used above to allow for marking and dropping 534 to be either probabilistic or deterministic. The example AQMs in 535 [I-D.ietf-tsvwg-aqm-dualq-coupled] drop and mark probabilistically, 536 so the drop probability is arranged to be the square of the marking 537 probability. Nonetheless, an alternative AQM that dropped and marked 538 deterministically would be valid, as long as the dropping frequency 539 was proportional to the square of the marking frequency. 541 Note that, contrary to RFC 3168, a Dual AQM implementing the L4S and 542 Classic treatments does not mark an ECT(1) packet under the same 543 conditions that it would have dropped a Not-ECT packet, as allowed by 544 [RFC8311], which updates RFC 3168. However, it does mark an ECT(0) 545 packet under the same conditions that it would have dropped a Not-ECT 546 packet. 548 5.3. Exception for L4S Packet Identification by Network Nodes with 549 Transport-Layer Awareness 551 To implement the L4S treatment, a network node does not need to 552 identify transport-layer flows. Nonetheless, if an implementer is 553 willing to identify transport-layer flows at a network node, and if 554 the most recent ECT packet in the same flow was ECT(0), the node MAY 555 classify CE packets for classic ECN [RFC3168] treatment. In all 556 other cases, a network node MUST classify all CE packets for L4S 557 treatment. Examples of such other cases are: i) if no ECT packets 558 have yet been identified in a flow; ii) if it is not desirable for a 559 network node to identify transport-layer flows; or iii) if the most 560 recent ECT packet in a flow was ECT(1). 562 If an implementer uses flow-awareness to classify CE packets, to 563 determine whether the flow is using ECT(0) or ECT(1) it only uses the 564 most recent ECT packet of a flow (this advice will need to be 565 verified as part of L4S experiments). This is because a sender might 566 switch from sending ECT(1) (L4S) packets to sending ECT(0) (Classic) 567 packets, or back again, in the middle of a transport-layer flow (e.g. 568 it might manually switch its congestion control module mid- 569 connection, or it might be deliberately attempting to confuse the 570 network). 572 5.4. Interaction of the L4S Identifier with other Identifiers 574 5.4.1. Examples of Other Identifiers Complementing L4S Identifiers 576 5.4.1.1. Inclusion of Additional Traffic with L4S 578 In a typical case for the public Internet a network element that 579 implements L4S might want to classify some low-rate but unresponsive 580 traffic (e.g. DNS, voice, game sync packets) into the low latency 581 queue to mix with L4S traffic. Such non-ECN-based packet types MUST 582 be safe to mix with L4S traffic without harming the low latency 583 service, where 'safe' is explained in Section 5.4.1.1.1 below. 585 In this case it would not be appropriate to call the queue an L4S 586 queue, because it is shared by L4S and non-L4S traffic. Instead it 587 will be called the low latency or L queue. The L queue then offers 588 two different treatments: 590 o The L4S treatment, which is a combination of the L4S AQM treatment 591 and a priority scheduling treatment; 593 o The low latency treatment, which is solely the priority scheduling 594 treatment, without ECN-marking by the AQM. 596 To identify packets for just the scheduling treatment, it would be 597 inappropriate to use the L4S ECT(1) identifier, because such traffic 598 is unresponsive to ECN marking. Therefore, a network element that 599 implements L4S MAY classify additional packets into the L queue if 600 they carry certain non-ECN identifiers. For instance: 602 o addresses of specific applications or hosts configured to be safe 603 (but for example cannot set the ECN field for some temporary 604 reason); 606 o certain protocols that are usually lightweight (e.g. ARP, DNS); 608 o specific Diffserv codepoints that indicate traffic with limited 609 burstiness such as the EF (Expedited Forwarding) and Voice-Admit 610 service classes or equivalent local-use DSCPs (see 611 [I-D.briscoe-tsvwg-l4s-diffserv]). 613 Of course, a packet that carried both the ECT(1) codepoint and a 614 relevant non-ECN identifier would also be classified into the L 615 queue. 617 For clarity, non-ECN identifiers, such as the examples itemized 618 above, might be used by some network operators who believe they 619 identify non-L4S traffic that would be safe to mix with L4S traffic. 621 They are not alternative ways for a host to indicate that it is 622 sending L4S packets. Only the ECT(1) ECN codepoint indicates to a 623 network element that a host is sending L4S packets (and CE indicates 624 that it could be). Specifically ECT(1) indicates that the host 625 claims its behaviour satisfies the per-requisite transport 626 requirements in Section 4. 628 5.4.1.1.1. 'Safe' Unresponsive Traffic 630 The above section requires unresponsive traffic to be 'safe' to mix 631 with L4S traffic. Ideally this means that the sender never sends any 632 sequence of packets at a data rate that exceeds the available 633 capacity of the bottleneck link. However, typically an unresponsive 634 transport does not even know the bottleneck capacity of the path, let 635 alone its available capacity. Nonetheless, an application can be 636 considered safe enough if it paces packets out (not necessarily 637 completely regularly) such that its maximum instantaneous data rate 638 from packet to packet stays well below a typical broadband access 639 rate. 641 This is a vague but useful definition, because it encompasses many 642 low latency applications of interest, such as DNS, voice, game sync 643 packets, RPC, ACKs, keep-alives, etc. 645 5.4.1.2. Exclusion of Traffic From L4S Treatment 647 To extend the above example, an operator might want to exclude some 648 traffic from the L4S treatment for policy reason, e.g. security 649 (traffic from malicious sources) or commercial (initially the 650 operator may wish to confine the benefits of L4S to business 651 customers). 653 In this exclusion case, the operator MUST classify on the relevant 654 locally-used identifiers (e.g. source addresses) before classifying 655 the non-matching traffic on the end-to-end L4S ECN identifier. 657 The operator MUST NOT re-mark the end-to-end L4S identifier, because 658 its decision to exclude certain traffic from L4S treatment is local- 659 only. The end-to-end L4S identifier then survives for other 660 operators to use, or indeed, they can apply their own policy, 661 independently based on their own choice of locally-used identifiers. 662 This approach also allows any operator to remove its locally-applied 663 exclusions in future, e.g. if it wishes to widen the benefit of the 664 L4S treatment to all its customers. 666 5.4.2. Generalized Combination of L4S and Other Identifiers 668 L4S concerns low latency, which it can provide for all traffic 669 without differentiation and without affecting bandwidth allocation. 670 Diffserv provides for differentiation of both bandwidth and low 671 latency, but its control of latency depends on its control of 672 bandwidth. The two can be combined if a network operator wants to 673 control bandwidth allocation but it also wants to provide low latency 674 - for any amount of traffic within one of these allocations of 675 bandwidth (rather than only providing low latency by limiting 676 bandwidth) [I-D.briscoe-tsvwg-l4s-diffserv]. 678 The examples above were framed in the context of providing the 679 default Best Efforts Per-Hop Behaviour (PHB) using two queues - a Low 680 Latency (L) queue and a Classic (C) Queue. This single DualQ 681 structure is expected to be by far the most common and useful 682 arrangement. But, more generally, an operator might choose to 683 control bandwidth allocation through a hierarchy of Diffserv PHBs at 684 a node, and to offer one (or more) of these PHBs with a low latency 685 and a classic variant. 687 In the first case, if we assume that there are no other PHBs except 688 the DualQ, if a packet carries ECT(1) or CE, a network element would 689 classify it for the L4S treatment irrespective of its DSCP. And, if 690 a packet carried (say) the EF DSCP, the network element could 691 classify it into the L queue irrespective of its ECN codepoint. 692 However, where the DualQ is in a hierarchy of other PHBs, the 693 classifier would classify some traffic into other PHBs based on DSCP 694 before classifying between the latency and classic queues (based on 695 ECT(1), CE and perhaps also the EF DSCP or other identifiers as in 696 the above example). [I-D.briscoe-tsvwg-l4s-diffserv] gives a number 697 of examples of such arrangements to address various requirements. 699 [I-D.briscoe-tsvwg-l4s-diffserv] describes how an operator might use 700 L4S to offer low latency for all L4S traffic as well as using 701 Diffserv for bandwidth differentiation. It identifies two main types 702 of approach, which can be combined: the operator might split certain 703 Diffserv PHBs between L4S and a corresponding Classic service. Or it 704 might split the L4S and/or the Classic service into multiple Diffserv 705 PHBs. In any of these cases, a packet would have to be classified on 706 its Diffserv and ECN codepoints. 708 In summary, there are numerous ways in which the L4S ECN identifier 709 (ECT(1) and CE) could be combined with other identifiers to achieve 710 particular objectives. The following categorization articulates 711 those that are valid, but it is not necessarily exhaustive. Those 712 tagged 'Global-use' could be set by the sending host or a network. 713 Those tagged 'Local-use' would only be set by a network: 715 1. Identifiers Complementing the L4S Identifier 717 A. Including More Traffic in the L Queue 718 (Global-use or Local-use) 720 B. Excluding Certain Traffic from the L Queue 721 (Local-use only) 723 2. Identifiers to place L4S classification in a PHB Hierarchy 724 (Global-use or Local-use) 726 A. PHBs Before L4S ECN Classification 728 B. PHBs After L4S ECN Classification 730 6. L4S Experiments 732 [I-D.ietf-tsvwg-aqm-dualq-coupled] sets operational and management 733 requirements for experiments with DualQ Coupled AQMs. General 734 operational and management requirements for experiments with L4S 735 congestion controls are given in Section 4 and Section 5 above, e.g. 736 co-existence and scaling requirements, incremental deployment 737 arrangements. The specification of each scalable congestion control 738 will need to include protocol-specific requirements for configuration 739 and monitoring performance during experiments. Appendix A of 740 [RFC5706] provides a helpful checklist. 742 7. IANA Considerations 744 This specification contains no IANA considerations. 746 8. Security Considerations 748 Approaches to assure the integrity of signals using the new identifer 749 are introduced in Appendix C.1. 751 The requirement to detect loss in time units prevents the ACK- 752 splitting attacks described in [Savage-TCP]. 754 9. Acknowledgements 756 Thanks to Richard Scheffenegger, John Leslie, David Taeht, Jonathan 757 Morton, Gorry Fairhurst, Michael Welzl, Mikael Abrahamsson and Andrew 758 McGregor for the discussions that led to this specification. Ing-jyh 759 (Inton) Tsang was a contributor to the early drafts of this document. 760 And thanks to Mikael Abrahamsson, Lloyd Wood, Nicolas Kuhn, Greg 761 White, David Black and Gorry Fairhurst for providing help and 762 reviewing this draft and to Ingemar Johansson for reviewing and 763 providing substantial text. Appendix A listing the Prague L4S 764 Requirements is based on text authored by Marcelo Bagnulo Braun that 765 was originally an appendix to [I-D.ietf-tsvwg-l4s-arch]. That text 766 was in turn based on the collective output of the attendees listed in 767 the minutes of a 'bar BoF' on DCTCP Evolution during IETF-94 768 [TCPPrague]. 770 The authors' contributions were part-funded by the European Community 771 under its Seventh Framework Programme through the Reducing Internet 772 Transport Latency (RITE) project (ICT-317700). Bob Briscoe was also 773 part-funded by the Research Council of Norway through the TimeIn 774 project. The views expressed here are solely those of the authors. 776 10. References 778 10.1. Normative References 780 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 781 Requirement Levels", BCP 14, RFC 2119, 782 DOI 10.17487/RFC2119, March 1997, 783 . 785 [RFC3168] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition 786 of Explicit Congestion Notification (ECN) to IP", 787 RFC 3168, DOI 10.17487/RFC3168, September 2001, 788 . 790 [RFC4774] Floyd, S., "Specifying Alternate Semantics for the 791 Explicit Congestion Notification (ECN) Field", BCP 124, 792 RFC 4774, DOI 10.17487/RFC4774, November 2006, 793 . 795 [RFC6679] Westerlund, M., Johansson, I., Perkins, C., O'Hanlon, P., 796 and K. Carlberg, "Explicit Congestion Notification (ECN) 797 for RTP over UDP", RFC 6679, DOI 10.17487/RFC6679, August 798 2012, . 800 10.2. Informative References 802 [Alizadeh-stability] 803 Alizadeh, M., Javanmard, A., and B. Prabhakar, "Analysis 804 of DCTCP: Stability, Convergence, and Fairness", ACM 805 SIGMETRICS 2011 , June 2011. 807 [ARED01] Floyd, S., Gummadi, R., and S. Shenker, "Adaptive RED: An 808 Algorithm for Increasing the Robustness of RED's Active 809 Queue Management", ACIRI Technical Report , August 2001, 810 . 812 [DCttH15] De Schepper, K., Bondarenko, O., Briscoe, B., and I. 813 Tsang, "'Data Centre to the Home': Ultra-Low Latency for 814 All", RITE Project Technical Report , 2015, 815 . 817 [I-D.briscoe-tsvwg-l4s-diffserv] 818 Briscoe, B., "Interactions between Low Latency, Low Loss, 819 Scalable Throughput (L4S) and Differentiated Services", 820 draft-briscoe-tsvwg-l4s-diffserv-02 (work in progress), 821 November 2018. 823 [I-D.ietf-avtcore-cc-feedback-message] 824 Sarker, Z., Perkins, C., Singh, V., and M. Ramalho, "RTP 825 Control Protocol (RTCP) Feedback for Congestion Control", 826 draft-ietf-avtcore-cc-feedback-message-03 (work in 827 progress), December 2018. 829 [I-D.ietf-quic-transport] 830 Iyengar, J. and M. Thomson, "QUIC: A UDP-Based Multiplexed 831 and Secure Transport", draft-ietf-quic-transport-18 (work 832 in progress), January 2019. 834 [I-D.ietf-tcpm-accurate-ecn] 835 Briscoe, B., Kuehlewind, M., and R. Scheffenegger, "More 836 Accurate ECN Feedback in TCP", draft-ietf-tcpm-accurate- 837 ecn-08 (work in progress), March 2019. 839 [I-D.ietf-tcpm-generalized-ecn] 840 Bagnulo, M. and B. Briscoe, "ECN++: Adding Explicit 841 Congestion Notification (ECN) to TCP Control Packets", 842 draft-ietf-tcpm-generalized-ecn-03 (work in progress), 843 October 2018. 845 [I-D.ietf-tcpm-rack] 846 Cheng, Y., Cardwell, N., Dukkipati, N., and P. Jha, "RACK: 847 a time-based fast loss detection algorithm for TCP", 848 draft-ietf-tcpm-rack-04 (work in progress), July 2018. 850 [I-D.ietf-tsvwg-aqm-dualq-coupled] 851 Schepper, K., Briscoe, B., Bondarenko, O., and I. Tsang, 852 "DualQ Coupled AQMs for Low Latency, Low Loss and Scalable 853 Throughput (L4S)", draft-ietf-tsvwg-aqm-dualq-coupled-08 854 (work in progress), November 2018. 856 [I-D.ietf-tsvwg-ecn-encap-guidelines] 857 Briscoe, B., Kaippallimalil, J., and P. Thaler, 858 "Guidelines for Adding Congestion Notification to 859 Protocols that Encapsulate IP", draft-ietf-tsvwg-ecn- 860 encap-guidelines-11 (work in progress), November 2018. 862 [I-D.ietf-tsvwg-l4s-arch] 863 Briscoe, B., Schepper, K., and M. Bagnulo, "Low Latency, 864 Low Loss, Scalable Throughput (L4S) Internet Service: 865 Architecture", draft-ietf-tsvwg-l4s-arch-03 (work in 866 progress), October 2018. 868 [I-D.sridharan-tcpm-ctcp] 869 Sridharan, M., Tan, K., Bansal, D., and D. Thaler, 870 "Compound TCP: A New TCP Congestion Control for High-Speed 871 and Long Distance Networks", draft-sridharan-tcpm-ctcp-02 872 (work in progress), November 2008. 874 [I-D.stewart-tsvwg-sctpecn] 875 Stewart, R., Tuexen, M., and X. Dong, "ECN for Stream 876 Control Transmission Protocol (SCTP)", draft-stewart- 877 tsvwg-sctpecn-05 (work in progress), January 2014. 879 [Mathis09] 880 Mathis, M., "Relentless Congestion Control", PFLDNeT'09 , 881 May 2009, . 884 [Paced-Chirping] 885 Misund, J., "Rapid Acceleration in TCP Prague", Masters 886 Thesis , May 2018, 887 . 890 [QV] Briscoe, B. and P. Hurtig, "Up to Speed with Queue View", 891 RITE Technical Report D2.3; Appendix C.2, August 2015, 892 . 895 [RFC2309] Braden, B., Clark, D., Crowcroft, J., Davie, B., Deering, 896 S., Estrin, D., Floyd, S., Jacobson, V., Minshall, G., 897 Partridge, C., Peterson, L., Ramakrishnan, K., Shenker, 898 S., Wroclawski, J., and L. Zhang, "Recommendations on 899 Queue Management and Congestion Avoidance in the 900 Internet", RFC 2309, DOI 10.17487/RFC2309, April 1998, 901 . 903 [RFC2474] Nichols, K., Blake, S., Baker, F., and D. Black, 904 "Definition of the Differentiated Services Field (DS 905 Field) in the IPv4 and IPv6 Headers", RFC 2474, 906 DOI 10.17487/RFC2474, December 1998, 907 . 909 [RFC2983] Black, D., "Differentiated Services and Tunnels", 910 RFC 2983, DOI 10.17487/RFC2983, October 2000, 911 . 913 [RFC3246] Davie, B., Charny, A., Bennet, J., Benson, K., Le Boudec, 914 J., Courtney, W., Davari, S., Firoiu, V., and D. 915 Stiliadis, "An Expedited Forwarding PHB (Per-Hop 916 Behavior)", RFC 3246, DOI 10.17487/RFC3246, March 2002, 917 . 919 [RFC3540] Spring, N., Wetherall, D., and D. Ely, "Robust Explicit 920 Congestion Notification (ECN) Signaling with Nonces", 921 RFC 3540, DOI 10.17487/RFC3540, June 2003, 922 . 924 [RFC3649] Floyd, S., "HighSpeed TCP for Large Congestion Windows", 925 RFC 3649, DOI 10.17487/RFC3649, December 2003, 926 . 928 [RFC4340] Kohler, E., Handley, M., and S. Floyd, "Datagram 929 Congestion Control Protocol (DCCP)", RFC 4340, 930 DOI 10.17487/RFC4340, March 2006, 931 . 933 [RFC4341] Floyd, S. and E. Kohler, "Profile for Datagram Congestion 934 Control Protocol (DCCP) Congestion Control ID 2: TCP-like 935 Congestion Control", RFC 4341, DOI 10.17487/RFC4341, March 936 2006, . 938 [RFC4342] Floyd, S., Kohler, E., and J. Padhye, "Profile for 939 Datagram Congestion Control Protocol (DCCP) Congestion 940 Control ID 3: TCP-Friendly Rate Control (TFRC)", RFC 4342, 941 DOI 10.17487/RFC4342, March 2006, 942 . 944 [RFC4960] Stewart, R., Ed., "Stream Control Transmission Protocol", 945 RFC 4960, DOI 10.17487/RFC4960, September 2007, 946 . 948 [RFC5562] Kuzmanovic, A., Mondal, A., Floyd, S., and K. 949 Ramakrishnan, "Adding Explicit Congestion Notification 950 (ECN) Capability to TCP's SYN/ACK Packets", RFC 5562, 951 DOI 10.17487/RFC5562, June 2009, 952 . 954 [RFC5622] Floyd, S. and E. Kohler, "Profile for Datagram Congestion 955 Control Protocol (DCCP) Congestion ID 4: TCP-Friendly Rate 956 Control for Small Packets (TFRC-SP)", RFC 5622, 957 DOI 10.17487/RFC5622, August 2009, 958 . 960 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion 961 Control", RFC 5681, DOI 10.17487/RFC5681, September 2009, 962 . 964 [RFC5706] Harrington, D., "Guidelines for Considering Operations and 965 Management of New Protocols and Protocol Extensions", 966 RFC 5706, DOI 10.17487/RFC5706, November 2009, 967 . 969 [RFC5925] Touch, J., Mankin, A., and R. Bonica, "The TCP 970 Authentication Option", RFC 5925, DOI 10.17487/RFC5925, 971 June 2010, . 973 [RFC6077] Papadimitriou, D., Ed., Welzl, M., Scharf, M., and B. 974 Briscoe, "Open Research Issues in Internet Congestion 975 Control", RFC 6077, DOI 10.17487/RFC6077, February 2011, 976 . 978 [RFC6660] Briscoe, B., Moncaster, T., and M. Menth, "Encoding Three 979 Pre-Congestion Notification (PCN) States in the IP Header 980 Using a Single Diffserv Codepoint (DSCP)", RFC 6660, 981 DOI 10.17487/RFC6660, July 2012, 982 . 984 [RFC7560] Kuehlewind, M., Ed., Scheffenegger, R., and B. Briscoe, 985 "Problem Statement and Requirements for Increased Accuracy 986 in Explicit Congestion Notification (ECN) Feedback", 987 RFC 7560, DOI 10.17487/RFC7560, August 2015, 988 . 990 [RFC7567] Baker, F., Ed. and G. Fairhurst, Ed., "IETF 991 Recommendations Regarding Active Queue Management", 992 BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015, 993 . 995 [RFC7713] Mathis, M. and B. Briscoe, "Congestion Exposure (ConEx) 996 Concepts, Abstract Mechanism, and Requirements", RFC 7713, 997 DOI 10.17487/RFC7713, December 2015, 998 . 1000 [RFC8033] Pan, R., Natarajan, P., Baker, F., and G. White, 1001 "Proportional Integral Controller Enhanced (PIE): A 1002 Lightweight Control Scheme to Address the Bufferbloat 1003 Problem", RFC 8033, DOI 10.17487/RFC8033, February 2017, 1004 . 1006 [RFC8257] Bensley, S., Thaler, D., Balasubramanian, P., Eggert, L., 1007 and G. Judd, "Data Center TCP (DCTCP): TCP Congestion 1008 Control for Data Centers", RFC 8257, DOI 10.17487/RFC8257, 1009 October 2017, . 1011 [RFC8290] Hoeiland-Joergensen, T., McKenney, P., Taht, D., Gettys, 1012 J., and E. Dumazet, "The Flow Queue CoDel Packet Scheduler 1013 and Active Queue Management Algorithm", RFC 8290, 1014 DOI 10.17487/RFC8290, January 2018, 1015 . 1017 [RFC8298] Johansson, I. and Z. Sarker, "Self-Clocked Rate Adaptation 1018 for Multimedia", RFC 8298, DOI 10.17487/RFC8298, December 1019 2017, . 1021 [RFC8311] Black, D., "Relaxing Restrictions on Explicit Congestion 1022 Notification (ECN) Experimentation", RFC 8311, 1023 DOI 10.17487/RFC8311, January 2018, 1024 . 1026 [RFC8312] Rhee, I., Xu, L., Ha, S., Zimmermann, A., Eggert, L., and 1027 R. Scheffenegger, "CUBIC for Fast Long-Distance Networks", 1028 RFC 8312, DOI 10.17487/RFC8312, February 2018, 1029 . 1031 [Savage-TCP] 1032 Savage, S., Cardwell, N., Wetherall, D., and T. Anderson, 1033 "TCP Congestion Control with a Misbehaving Receiver", ACM 1034 SIGCOMM Computer Communication Review 29(5):71--78, 1035 October 1999. 1037 [TCP-CA] Jacobson, V. and M. Karels, "Congestion Avoidance and 1038 Control", Laurence Berkeley Labs Technical Report , 1039 November 1988, . 1041 [TCP-sub-mss-w] 1042 Briscoe, B. and K. De Schepper, "Scaling TCP's Congestion 1043 Window for Small Round Trip Times", BT Technical Report 1044 TR-TUB8-2015-002, May 2015, 1045 . 1048 [TCPPrague] 1049 Briscoe, B., "Notes: DCTCP evolution 'bar BoF': Tue 21 Jul 1050 2015, 17:40, Prague", tcpprague mailing list archive , 1051 July 2015, . 1054 [VCP] Xia, Y., Subramanian, L., Stoica, I., and S. Kalyanaraman, 1055 "One more bit is enough", Proc. SIGCOMM'05, ACM CCR 1056 35(4)37--48, 2005, 1057 . 1059 Appendix A. The 'Prague L4S Requirements' 1061 This appendix is informative, not normative. It gives a list of 1062 modifications to current scalable congestion controls so that they 1063 can be deployed over the public Internet and coexist safely with 1064 existing traffic. The list complements the normative requirements in 1065 Section 4 that a sender has to comply with before it can set the L4S 1066 identifier in packets it sends into the Internet. As well as 1067 necessary safety improvements (requirements) this appendix also 1068 includes preferable performance improvements (optimizations). 1070 These recommendations have become know as the Prague L4S 1071 Requirements, because they were originally identified at an ad hoc 1072 meeting during IETF-94 in Prague [TCPPrague]. The wording has been 1073 generalized to apply to all scalable congestion controls, not just 1074 TCP congestion control specifically. They were originally called the 1075 'TCP Prague Requirements', but they are not solely applicable to TCP, 1076 so the name has been generalized, and TCP Prague is now used for a 1077 specific implementation of the requirements. 1079 DCTCP [RFC8257] is currently the most widely used scalable transport 1080 protocol. In its current form, DCTCP is specified to be deployable 1081 only in controlled environments. Deploying it in the public Internet 1082 would lead to a number of issues, both from the safety and the 1083 performance perspective. The modifications and additional mechanisms 1084 listed in this section will be necessary for its deployment over the 1085 global Internet. Where an example is needed, DCTCP is used as a 1086 base, but it is likely that most of these requirements equally apply 1087 to other scalable congestion controls. 1089 A.1. Requirements for Scalable Transport Protocols 1091 A.1.1. Use of L4S Packet Identifier 1093 Description: A scalable congestion control needs to distinguish the 1094 packets it sends from those sent by classic congestion controls. 1096 Motivation: It needs to be possible for a network node to classify 1097 L4S packets without flow state into a queue that applies an L4S ECN 1098 marking behaviour and isolates L4S packets from the queuing delay of 1099 classic packets. 1101 A.1.2. Accurate ECN Feedback 1103 Description: The transport protocol for a scalable congestion control 1104 needs to provide timely, accurate feedback about the extent of ECN 1105 marking experienced by all packets. 1107 Motivation: Classic congestion controls only need feedback about the 1108 existence of a congestion episode within a round trip, not precisely 1109 how many packets were marked with ECN or dropped. Therefore, in 1110 2001, when ECN feedback was added to TCP [RFC3168], it could not 1111 inform the sender of more than one ECN mark per RTT. Since then, 1112 requirements for more accurate ECN feedback in TCP have been defined 1113 in [RFC7560] and [I-D.ietf-tcpm-accurate-ecn] specifies an 1114 experimental change to the TCP wire protocol to satisfy these 1115 requirements. Most other transport protocols already satisfy this 1116 requirement. 1118 A.1.3. Fall back to Reno-friendly congestion control on packet loss 1120 Description: A scalable congestion control needs to react to packet 1121 loss in a way that will coexist safely with a TCP Reno congestion 1122 control [RFC5681]. 1124 Motivation: Part of the safety conditions for deploying a scalable 1125 congestion control on the public Internet is to make sure that it 1126 behaves properly when it builds a queue at a network bottleneck that 1127 has not been upgraded to support L4S. Packet loss can have many 1128 causes, but it usually has to be conservatively assumed that it is a 1129 sign of congestion. Therefore, on detecting packet loss, a scalable 1130 congestion control will need to fall back to classic congestion 1131 control behaviour. If it does not comply with this requirement it 1132 could starve classic traffic. 1134 A scalable congestion control can be used for different types of 1135 transport, e.g. for real-time media or for reliable bulk transport 1136 like TCP. Therefore, the particular classic congestion control 1137 behaviour to fall back on will need to be part of the congestion 1138 control specification of the relevant transport. In the particular 1139 case of DCTCP, the current DCTCP specification states that "It is 1140 RECOMMENDED that an implementation deal with loss episodes in the 1141 same way as conventional TCP." For safe deployment of a scalable 1142 congestion control in the public Internet, the above requirement 1143 would need to be defined as a "MUST". 1145 Packet loss might (rarely) occur in the case that the bottleneck is 1146 L4S capable. In this case, the sender may receive a high number of 1147 packets marked with the CE bit set and also experience a loss. 1148 Current DCTCP implementations react differently to this situation. 1149 At least one implementation reacts only to the drop signal (e.g. by 1150 halving the CWND) and at least another DCTCP implementation reacts to 1151 both signals (e.g. by halving the CWND due to the drop and also 1152 further reducing the CWND based on the proportion of marked packet). 1153 We believe that further experimentation is needed to understand what 1154 is the best behaviour for the public Internet, which may or not be 1155 one of these existing approaches. 1157 A.1.4. Fall back to Reno-friendly congestion control on classic ECN 1158 bottlenecks 1160 Description: A scalable congestion control needs to react to ECN 1161 marking from a non-L4S but ECN-capable bottleneck in a way that will 1162 coexist with a TCP Reno congestion control [RFC5681]. 1164 Motivation: Similarly to the requirement in Appendix A.1.3, this 1165 requirement is a safety condition to ensure a scalable congestion 1166 control behaves properly when it builds a queue at a network 1167 bottleneck that has not been upgraded to support L4S. On detecting 1168 classic ECN marking (see below), a scalable congestion control will 1169 need to fall back to classic congestion control behaviour. If it 1170 does not comply with this requirement it could starve classic 1171 traffic. 1173 It would take time for endpoints to distinguish classic and L4S ECN 1174 marking. An increase in queuing delay or in delay variation would be 1175 a tell-tale sign, but it is not yet clear where a line would be drawn 1176 between the two behaviours. It might be possible to cache what was 1177 learned about the path to help subsequent attempts to detect the type 1178 of marking. 1180 A.1.5. Reduce RTT dependence 1182 Description: A scalable congestion control needs to reduce or 1183 eliminate RTT bias over as wide a range of RTTs as possible, or at 1184 least over the typical range of RTTs that will interact in the 1185 intended deployment scenario. 1187 Motivation: Classic TCP's throughput is known to be inversely 1188 proportional to RTT, so one would expect flows over very low RTT 1189 paths to nearly starve flows over larger RTTs. However, Classic TCP 1190 has never allowed a very low RTT path to exist because it induces a 1191 large queue. For instance, consider two paths with base RTT 1ms and 1192 100ms. If Classic TCP induces a 100ms queue, it turns these RTTs 1193 into 101ms and 200ms leading to a throughput ratio of about 2:1. 1194 Whereas if a Scalable TCP induces only a 1ms queue, the ratio is 1195 2:101, leading to a throughput ratio of about 50:1. 1197 Therefore, with very small queues, long RTT flows will essentially 1198 starve, unless scalable congestion controls comply with this 1199 requirement. 1201 A.1.6. Scaling down to fractional congestion windows 1203 Description: A scalable congestion control needs to remain responsive 1204 to congestion when RTTs are significantly smaller than in the current 1205 public Internet. 1207 Motivation: As currently specified, the minimum required congestion 1208 window of TCP (and its derivatives) is set to 2 maximum segment sizes 1209 (MSS) (see equation (4) in [RFC5681]). Once the congestion window 1210 reaches this minimum, all current TCP algorithms become unresponsive 1211 to congestion signals. No matter how much drop or ECN marking, the 1212 congestion window no longer reduces. Instead, TCP forces the queue 1213 to grow, overriding any AQM and increasing queuing delay. 1215 L4S mechanisms significantly reduce queueing delay so, over the same 1216 path, the RTT becomes lower. Then this problem becomes surprisingly 1217 common [TCP-sub-mss-w]. This is because, for the same link capacity, 1218 smaller RTT implies a smaller window. For instance, consider a 1219 residential setting with an upstream broadband Internet access of 8 1220 Mb/s, assuming a max segment size of 1500 B. Two upstream flows will 1221 each have the minimum window of 2 MSS if the RTT is 6ms or less, 1222 which is quite common when accessing a nearby data centre. So, any 1223 more than two such parallel TCP flows will become unresponsive and 1224 increase queuing delay. 1226 Unless scalable congestion controls are required to comply with this 1227 requirement from the start, they will frequently become unresponsive, 1228 negating the low latency benefit of L4S, for themselves and for 1229 others. One possible sub-MSS window mechanism is described in 1230 [TCP-sub-mss-w], and other approaches are likely to be feasible. 1232 A.1.7. Measuring Reordering Tolerance in Time Units 1234 Description: A scalable congestion control needs to detect loss by 1235 counting in time-based units, which is scalable, rather than counting 1236 in units of packets, which is not. 1238 Motivation: If it is known that all L4S senders using a link obey 1239 this rule, then link technologies that support L4S can remove the 1240 head-of-line blocking delay they have to introduce while trying to 1241 keep packets in tight order to avoid triggering loss detection based 1242 on counting packets. 1244 End-systems cannot know whether a missing packet is due to loss or 1245 reordering, except in hindsight - if it appears later. If senders 1246 deem that loss has occurred by counting reordered packets (e.g. the 3 1247 Duplicate ACK rule of Classic TCP), the time over which the network 1248 has to keep packets in order scales down as packet rates scale up 1249 over the years. In contrast, if senders allow a reordering window in 1250 time-based units before they deem there has been a loss, the time 1251 over which the network has to keep packets in order stays constant. 1253 The potential benefit for links comes in two parts: i) switching the 1254 unit from packet count to time-based; ii) potentially relaxing the 1255 amount of time available for re-ordering. The initial switch to 1256 time-based units offers no immediate benefit, but as the years 1257 progress it stops the reordering requirement becoming tighter. The 1258 secondary relaxation might be possible where some transport protocols 1259 find they can tolerate more re-ordering (e.g. more than the 3 DupACK 1260 rule, perhaps because it was reasonable when packet rates were low, 1261 but it is now far too tight). 1263 Tolerance of reordering over a small duration could allow parallel 1264 (e.g. bonded-channel) link technologies to relax their need to 1265 deliver packets strictly in order. Such links typically give 1266 arriving packets a link-level sequence number and introduce delay 1267 while buffering packets at the receiving end until they can be 1268 delivered in the same order. For radio links, this delay usually 1269 includes the time allowed for link-layer retransmissions. 1271 Note, however, that a link technology will only be able to relax its 1272 ordering requirement if it is certain that it will not degrade the 1273 transport /most/ sensitive to reordering that might use the link. 1274 Also note that in some controlled environments no reordering is 1275 tolerated by some transports (e.g. RoCEv2 ussed for RDMA), therefore 1276 a switch to time-based units could not be exploited to relax 1277 reordering. 1279 For receivers that need their packets in order, it would seem that 1280 relaxing network ordering would simply shift this reordering delay 1281 from the network to the receiver. However, that is not true in the 1282 general case because links generally do not recognize transport layer 1283 flows and often cannot even see application layer streams within the 1284 flows (as in SCTP, HTTP/2 or QUIC). So a link will often be holding 1285 back packets from one flow or stream while waiting for those from 1286 another. Relaxing strict ordering in the network will remove this 1287 head-of-line blocking delay. {ToDo: this is being quantified 1288 experimentally - will need to add the figures here.} 1290 Classic TCP implementations are switching over to the time-based 1291 approach of RACK (Recent ACKnowledgements [I-D.ietf-tcpm-rack]). 1292 However, it will be many years (decades?) before networks no longer 1293 have to allow for the presence of traditional TCP senders still using 1294 the 3 DupACK rule. This specification (Section 4.3) says that 1295 senders are not entitled to identify packets as L4S in the IP/ECN 1296 field unless they use the time-based approach. Then networks that 1297 identify L4S traffic separately (e.g. using 1298 [I-D.ietf-tsvwg-aqm-dualq-coupled]) can know for certain that all L4S 1299 traffic is using the scalable time-based approach. 1301 This will allow networks to remove the head-of-line blocking delay of 1302 resequencing straight away, but only for L4S traffic. Classic 1303 traffic will have to wait for many years until incremental deployment 1304 of RACK has become near-universal. Nonetheless, experience with RACK 1305 will determine how much reordering tolerance networks will be 1306 reasonable for L4S traffic. 1308 Performance Optimization as well as Safety Improvement: The delay 1309 benefit would be lost if any L4S sender did not follow the time-based 1310 approach. Therefore, the time-based approach is made a normative 1311 requirement (a necessary safety improvement). Nonetheless, the time- 1312 based approach also enables a throughput benefit that a flow can 1313 enjoy independently of others (a performance optimization), explained 1314 next. 1316 Given the requirement for a scalable congestion control to fall-back 1317 to Reno or Cubic on a loss (see Appendix A.1.3), it is important that 1318 a scalable congestion control does not deem that a loss has occurred 1319 too soon. If, later within the same round trip, an out-of-order 1320 acknowledgement fills the gap, the sender would have halved its rate 1321 spuriously (as well as retransmitting spuriously). With a RACK-like 1322 approach, allowing longer before a loss is deemed to have occurred 1323 maintains higher throughput in the presence of reordering {ToDo: 1324 Quantify this statement}. 1326 On the other hand, it is also important not to wait too long before 1327 deeming that a gap is due to a loss (termed a long reordering 1328 window), otherwise loss recovery would be slow. 1330 The speed of loss recovery is much more significant for short flows 1331 than long, therefore a good compromise would adapt the reordering 1332 window; from a small fraction of the RTT at the start of a flow, to a 1333 larger fraction of the RTT for flows that continue for many round 1334 trips. This is the approach adopted by TCP RACK (Recent 1335 ACKnowledgements) [I-D.ietf-tcpm-rack] and recommended for all L4S 1336 senders, whether using TCP or another transport protocol. 1338 The requirement to detect loss in time units also prevents the ACK- 1339 splitting attacks described in [Savage-TCP]. 1341 A.2. Scalable Transport Protocol Optimizations 1343 A.2.1. Setting ECT in TCP Control Packets and Retransmissions 1345 Description: This item only concerns TCP and its derivatives (e.g. 1346 SCTP), because the original specification of ECN for TCP precluded 1347 the use of ECN on control packets and retransmissions. To improve 1348 performance, scalable transport protocols ought to enable ECN at the 1349 IP layer in TCP control packets (SYN, SYN-ACK, pure ACKs, etc.) and 1350 in retransmitted packets. The same is true for derivatives of TCP, 1351 e.g. SCTP. 1353 Motivation: RFC 3168 prohibits the use of ECN on these types of TCP 1354 packet, based on a number of arguments. This means these packets are 1355 not protected from congestion loss by ECN, which considerably harms 1356 performance, particularly for short flows. 1357 [I-D.ietf-tcpm-generalized-ecn] counters each argument in RFC 3168 in 1358 turn, showing it was over-cautious. Instead it proposes experimental 1359 use of ECN on all types of TCP packet as long as AccECN feedback 1360 [I-D.ietf-tcpm-accurate-ecn] is available (which is itself a 1361 prerequisite for using a scalable congestion control). 1363 A.2.2. Faster than Additive Increase 1365 Description: It would improve performance if scalable congestion 1366 controls did not limit their congestion window increase to the 1367 traditional additive increase of 1 MSS per round trip [RFC5681] 1368 during congestion avoidance. The same is true for derivatives of TCP 1369 congestion control, including similar approaches used for real-time 1370 media. 1372 Motivation: As currently defined, DCTCP uses the traditional TCP Reno 1373 additive increase in congestion avoidance phase. When the available 1374 capacity suddenly increases (e.g. when another flow finishes, or if 1375 radio capacity increases) it can take very many round trips to take 1376 advantage of the new capacity. In the steady state, DCTCP induces 1377 about 2 ECN marks per round trip, so it should be possible to quickly 1378 detect when these signals have disappeared and seek available 1379 capacity more rapidly. It will of course be necessary to minimize 1380 the impact on other flows (classic and scalable). 1382 TCP Cubic was designed to solve this problem, but as flow rates have 1383 continued to increase, the delay accelerating into available capacity 1384 has become prohibitive. For instance, with RTT=20 ms, to increase 1385 flow rate from 100Mb/s to 200Mb/s Cubic takes between 50 and 100 1386 round trips. Every 8x increase in flow rate leads to 2x more 1387 acceleration delay. 1389 A.2.3. Faster Convergence at Flow Start 1391 Description: Particularly when a flow starts, scalable congestion 1392 controls need to converge (reach their steady-state share of the 1393 capacity) at least as fast as classic TCP and preferably faster. 1394 This does not just affect TCP Prague, but also the flow start 1395 behaviour of any L4S congestion control derived from a Classic 1396 transport that uses TCP slow start, including those for real-time 1397 media. 1399 Motivation: As an example, a new DCTCP flow takes longer than classic 1400 TCP to obtain its share of the capacity of the bottleneck when there 1401 are already ongoing flows using the bottleneck capacity. In a data 1402 centre environment DCTCP takes about a factor of 1.5 to 2 longer to 1403 converge due to the much higher typical level of ECN marking that 1404 DCTCP background traffic induces, which causes new flows to exit slow 1405 start early [Alizadeh-stability]. In testing for use over the public 1406 Internet the convergence time of DCTCP relative to regular TCP is 1407 even less favourable [Paced-Chirping]). It is exacerbated by the 1408 typically greater mismatch between the link rate of the sending host 1409 and typical Internet access bottlenecks, in combination with the 1410 shallow ECN marking threshold needed for L4S. This problem is 1411 detrimental in general, but would particularly harm the performance 1412 of short flows relative to classic TCP. 1414 Appendix B. Alternative Identifiers 1416 This appendix is informative, not normative. It records the pros and 1417 cons of various alternative ways to identify L4S packets to record 1418 the rationale for the choice of ECT(1) (Appendix B.1) as the L4S 1419 identifier. At the end, Appendix B.6 summarises the distinguishing 1420 features of the leading alternatives. It is intended to supplement, 1421 not replace the detailed text. 1423 The leading solutions all use the ECN field, sometimes in combination 1424 with the Diffserv field. Both the ECN and Diffserv fields have the 1425 additional advantage that they are no different in either IPv4 or 1426 IPv6. A couple of alternatives that use other fields are mentioned 1427 at the end, but it is quickly explained why they are not serious 1428 contenders. 1430 B.1. ECT(1) and CE codepoints 1432 Definition: 1434 Packets with ECT(1) and conditionally packets with CE would 1435 signify L4S semantics as an alternative to the semantics of 1436 classic ECN [RFC3168], specifically: 1438 * The ECT(1) codepoint would signify that the packet was sent by 1439 an L4S-capable sender; 1441 * Given shortage of codepoints, both L4S and classic ECN sides of 1442 an AQM would have to use the same CE codepoint to indicate that 1443 a packet had experienced congestion. If a packet that had 1444 already been marked CE in an upstream buffer arrived at a 1445 subsequent AQM, this AQM would then have to guess whether to 1446 classify CE packets as L4S or classic ECN. Choosing the L4S 1447 treatment would be a safer choice, because then a few classic 1448 packets might arrive early, rather than a few L4S packets 1449 arriving late; 1451 * Additional information might be available if the classifier 1452 were transport-aware. Then it could classify a CE packet for 1453 classic ECN treatment if the most recent ECT packet in the same 1454 flow had been marked ECT(0). However, the L4S service ought 1455 not to need tranport-layer awareness; 1457 Cons: 1459 Consumes the last ECN codepoint: The L4S service is intended to 1460 supersede the service provided by classic ECN, therefore using 1461 ECT(1) to identify L4S packets could ultimately mean that the 1462 ECT(0) codepoint was 'wasted' purely to distinguish one form of 1463 ECN from its successor; 1465 ECN hard in some lower layers: It is not always possible to support 1466 ECN in an AQM acting in a buffer below the IP layer 1467 [I-D.ietf-tsvwg-ecn-encap-guidelines]. In such cases, the L4S 1468 service would have to drop rather than mark frames even though 1469 they might contain an ECN-capable packet. However, such cases 1470 would be unusual. 1472 Risk of reordering classic CE packets: Having to classify all CE 1473 packets as L4S risks some classic CE packets being wrongly 1474 classified as L4S and arriving early, which is a form of 1475 reordering. Reordering can cause the TCP sender to retransmit 1476 spuriously. However, the risk of spurious retransmissions would 1477 be extremely low, because: 1479 1. it is quite unusual to experience more than one bottleneck 1480 queue on a path. 1482 2. It would be even more unusual for the first bottleneck to 1483 support classic ECN marking and for the second to support L4S 1484 ECN marking 1486 3. even then, reordering would only occur if there was 1487 simultaneous mixing of classic and L4S traffic, which would be 1488 more unlikely in an access link, which is where most 1489 bottlenecks are located; 1491 4. even then, spurious retransmissions would only occur if a 1492 contiguous sequence of three or more packets in one classic 1493 ECN flow were all CE-marked at the first bottleneck; 1495 5. even then, a spurious retransmission would only occur if the 1496 source did not support RACK [I-D.ietf-tcpm-rack], which is 1497 already widely supported. Otherwise a whole reordering window 1498 within one classic ECN flow would have to be marked CE at the 1499 first bottleneck to cause a spurious retransmission. 1501 It is extremely unlikely that a set of 5 eventualities that are 1502 each unusual in themselves would all happen simultaneously. But, 1503 even if they did, it would only cause spurious retransmission of a 1504 packet. 1506 Non-L4S service for control packets: The classic ECN RFCs [RFC3168] 1507 and [RFC5562] require a sender to clear the ECN field to Not-ECT 1508 for retransmissions and certain control packets specifically pure 1509 ACKs, window probes and SYNs. When L4S packets are classified by 1510 the ECN field alone, these control packets would not be classified 1511 into an L4S queue, and could therefore be delayed relative to the 1512 other packets in the flow. This would not cause re-ordering 1513 (because retransmissions are already out of order, and the control 1514 packets carry no data). However, it would make critical control 1515 packets more vulnerable to loss and delay. To address this 1516 problem, [I-D.ietf-tcpm-generalized-ecn] proposes an experiment in 1517 which all TCP control packets and retransmissions are ECN-capable 1518 as long as ECN feedback is available. 1520 Pros: 1522 Should work e2e: The ECN field generally works end-to-end across the 1523 Internet. Unlike the DSCP, the setting of the ECN field is at 1524 least forwarded unchanged by networks that do not support ECN, and 1525 networks rarely clear it to zero; 1527 Should work in tunnels: Unlike Diffserv, ECN is defined to always 1528 work across tunnels. However, tunnels do not always implement ECN 1529 processing as they should do, particularly because IPsec tunnels 1530 were defined differently for a few years. 1532 Could migrate to one codepoint: If all classic ECN senders 1533 eventually evolve to use the L4S service, the ECT(0) codepoint 1534 could be reused for some future purpose, but only once use of 1535 ECT(0) packets had reduced to zero, or near-zero, which might 1536 never happen. 1538 B.2. ECN Plus a Diffserv Codepoint (DSCP) 1540 Definition: 1542 For packets with a defined DSCP, all codepoints of the ECN field 1543 (except Not-ECT) would signify alternative L4S semantics to those 1544 for classic ECN [RFC3168], specifically: 1546 * The L4S DSCP would signifiy that the packet came from an L4S- 1547 capable sender; 1549 * ECT(0) and ECT(1) would both signify that the packet was 1550 travelling between transport endpoints that were both ECN- 1551 capable; 1553 * CE would signify that the packet had been marked by an AQM 1554 implementing the L4S service. 1556 Use of a DSCP is the only approach for alternative ECN semantics 1557 given as an example in [RFC4774]. However, it was perhaps considered 1558 more for controlled environments than new end-to-end services; 1560 Cons: 1562 Consumes DSCP pairs: A DSCP is obviously not orthogonal to Diffserv. 1563 Therefore, wherever the L4S service is applied to multiple 1564 Diffserv scheduling behaviours, it would be necessary to replace 1565 each DSCP with a pair of DSCPs. 1567 Uses critical lower-layer header space: The resulting increased 1568 number of DSCPs might be hard to support for some lower layer 1569 technologies, e.g. 802.1p and MPLS both offer only 3-bits for a 1570 maximum of 8 traffic class identifiers. Although L4S should 1571 reduce and possibly remove the need for some DSCPs intended for 1572 differentiated queuing delay, it will not remove the need for 1573 Diffserv entirely, because Diffserv is also used to allocate 1574 bandwidth, e.g. by prioritising some classes of traffic over 1575 others when traffic exceeds available capacity. 1577 Not end-to-end (host-network): Very few networks honour a DSCP set 1578 by a host. Typically a network will zero (bleach) the Diffserv 1579 field from all hosts. Sometimes networks will attempt to identify 1580 applications by some form of packet inspection and, based on 1581 network policy, they will set the DSCP considered appropriate for 1582 the identified application. Network-based application 1583 identification might use some combination of protocol ID, port 1584 numbers(s), application layer protocol headers, IP address(es), 1585 VLAN ID(s) and even packet timing. 1587 Not end-to-end (network-network): Very few networks honour a DSCP 1588 received from a neighbouring network. Typically a network will 1589 zero (bleach) the Diffserv field from all neighbouring networks at 1590 an interconnection point. Sometimes bilateral arrangements are 1591 made between networks, such that the receiving network remarks 1592 some DSCPs to those it uses for roughly equivalent services. The 1593 likelihood that a DSCP will be bleached or ignored depends on the 1594 type of DSCP: 1596 Local-use DSCP: These tend to be used to implement application- 1597 specific network policies, but a bilateral arrangement to 1598 remark certain DSCPs is often applied to DSCPs in the local-use 1599 range simply because it is easier not to change all of a 1600 network's internal configurations when a new arrangement is 1601 made with a neighbour; 1603 Global-use DSCP: These do not tend to be honoured across network 1604 interconnections more than local-use DSCPs. However, if two 1605 networks decide to honour certain of each other's DSCPs, the 1606 reconfiguration is a little easier if both of their globally 1607 recognised services are already represented by the relevant 1608 global-use DSCPs. 1610 Note that today a global-use DSCP gives little more assurance 1611 of end-to-end service than a local-use DSCP. In future the 1612 global-use range might give more assurance of end-to-end 1613 service than local-use, but it is unlikely that either 1614 assurance will be high, particularly given the hosts are 1615 included in the end-to-end path. 1617 Not all tunnels: Diffserv codepoints are often not propagated to the 1618 outer header when a packet is encapsulated by a tunnel header. 1619 DSCPs are propagated to the outer of uniform mode tunnels, but not 1620 pipe mode [RFC2983], and pipe mode is fairly common. 1622 ECN hard in some lower layers:: Because this approach uses both the 1623 Diffserv and ECN fields, an AQM wil only work at a lower layer if 1624 both can be supported. If individual network operators wished to 1625 deploy an AQM at a lower layer, they would usually propagate an IP 1626 Diffserv codepoint to the lower layer, using for example IEEE 1627 802.1p. However, the ECN capability is harder to propagate down 1628 to lower layers because few lower layers support it. 1630 Pros: 1632 Could migrate to e2e: If all usage of classic ECN migrates to usage 1633 of L4S, the DSCP would become redundant, and the ECN capability 1634 alone could eventually identify L4S packets without the 1635 interconnection problems of Diffserv detailed above, and without 1636 having permanently consumed more than one codepoint in the IP 1637 header. Although the DSCP does not generally function as an end- 1638 to-end identifier (see above), it could be used initially by 1639 individual ISPs to introduce the L4S service for their own locally 1640 generated traffic; 1642 B.3. ECN capability alone 1644 Definition: 1646 This approach uses ECN capability alone as the L4S identifier. It 1647 is only feasible if classic ECN is not widely deployed. The 1648 specific definition of codepoints would be: 1650 * Any ECN codepoint other than Not-ECT would signify an L4S- 1651 capable sender; 1653 * ECN codepoints would not be used for classic [RFC3168] ECN, and 1654 the classic network service would only be used for Not-ECT 1655 packets. 1657 This approach would only be feasible if 1659 A. it was generally agreed that there was little chance of any 1660 classic [RFC3168] ECN deployment in any network nodes; 1662 B. it was generally agreed that there was little chance of any 1663 client devices being deployed with classic [RFC3168] TCP-ECN 1664 on by default (note that classic TCP-ECN is already on-by- 1665 default on many servers); 1667 C. for TCP connections, developers of client OSs would all have 1668 to agree not to encourage further deployment of classic ECN. 1669 Specifically, at the start of a TCP connection classic ECN 1670 could be disabled during negotation of the ECN capability: 1672 + an L4S-capable host would have to disable ECN if the 1673 corresponding host did not support accurate ECN feedback 1674 [RFC7560], which is a prerequisite for the L4S service; 1676 + developers of operating systems for user devices would only 1677 enable ECN by default for TCP once the stack implemented 1678 L4S and accurate ECN feedback [RFC7560] including 1679 requesting accurate ECN feedback by default. 1681 Cons: 1683 Near-infeasible deployment constraints: The constraints for 1684 deployment above represent a highly unlikely, but not completely 1685 impossible, set of circumstances. If, despite the above measures, 1686 a pair of hosts did negotiate to use classic ECN, their packets 1687 would be classified into the same queue as L4S traffic, and if 1688 they had to compete with a long-running L4S flow they would get a 1689 very small capacity share; 1691 ECN hard in some lower layers: See the same issue with "ECT(1) and 1692 CE codepoints" (Appendix B.1); 1694 Non-L4S service for control packets: See the same issue with "ECT(1) 1695 and CE codepoints" (Appendix B.1). 1697 Pros: 1699 Consumes no additional codepoints: The ECT(1) codepoint and all 1700 spare Diffserv codepoints would remain available for future use; 1702 Should work e2e: As with "ECT(1) and CE codepoints" (Appendix B.1); 1704 Should work in tunnels: As with "ECT(1) and CE codepoints" 1705 (Appendix B.1). 1707 B.4. Protocol ID 1709 It has been suggested that a new ID in the IPv4 Protocol field or the 1710 IPv6 Next Header field could identify L4S packets. However this 1711 approach is ruled out by numerous problems: 1713 o A new protocol ID would need to be paired with the old one for 1714 each transport (TCP, SCTP, UDP, etc.); 1716 o In IPv6, there can be a sequence of Next Header fields, and it 1717 would not be obvious which one would be expected to identify a 1718 network service like L4S; 1720 o A new protocol ID would rarely provide an end-to-end service, 1721 because It is well-known that new protocol IDs are often blocked 1722 by numerous types of middlebox; 1724 o The approach is not a solution for AQMs below the IP layer; 1726 B.5. Source or destination addressing 1728 Locally, a network operator could arrange for L4S service to be 1729 applied based on source or destination addressing, e.g. packets from 1730 its own data centre and/or CDN hosts, packets to its business 1731 customers, etc. It could use addressing at any layer, e.g. IP 1732 addresses, MAC addresses, VLAN IDs, etc. Although addressing might 1733 be a useful tactical approach for a single ISP, it would not be a 1734 feasible approach to identify an end-to-end service like L4S. Even 1735 for a single ISP, it would require packet classifiers in buffers to 1736 be dependent on changing topology and address allocation decisions 1737 elsewhere in the network. Therefore this approach is not a feasible 1738 solution. 1740 B.6. Summary: Merits of Alternative Identifiers 1742 Table 1 provides a very high level summary of the pros and cons 1743 detailed against the schemes described respectively in Appendix B.2, 1744 Appendix B.3 and Appendix B.1, for six issues that set them apart. 1746 +--------------+--------------------+---------+--------------------+ 1747 | Issue | DSCP + ECN | ECN | ECT(1) + CE | 1748 +--------------+--------------------+---------+--------------------+ 1749 | | initial eventual | initial | initial eventual | 1750 | | | | | 1751 | end-to-end | N . . . ? . | . . Y | . . Y . . Y | 1752 | tunnels | . O . . O . | . . ? | . . ? . . Y | 1753 | lower layers | N . . . ? . | . O . | . O . . . ? | 1754 | codepoints | N . . . . ? | . . Y | N . . . . ? | 1755 | reordering | . . Y . . Y | . . Y | . O . . . ? | 1756 | ctrl pkts | . . Y . . Y | . O . | . O . . . ? | 1757 | | | | | 1758 | | | Note 1 | | 1759 +--------------+--------------------+---------+--------------------+ 1761 Note 1: Only feasible if classic ECN is obsolete. 1763 Table 1: Comparison of the Merits of Three Alternative Identifiers 1765 The schemes are scored based on both their capabilities now 1766 ('initial') and in the long term ('eventual'). The 'ECN' scheme 1767 shares the 'eventual' scores of the 'ECT(1) + CE' scheme. The scores 1768 are one of 'N, O, Y', meaning 'Poor', 'Ordinary', 'Good' 1769 respectively. The same scores are aligned vertically to aid the eye. 1770 A score of "?" in one of the positions means that this approach might 1771 optimisitically become this good, given sufficient effort. The table 1772 summarises the text and is not meant to be understandable without 1773 having read the text. 1775 Appendix C. Potential Competing Uses for the ECT(1) Codepoint 1777 The ECT(1) codepoint of the ECN field has already been assigned once 1778 for the ECN nonce [RFC3540], which has now been categorized as 1779 historic [RFC8311]. ECN is probably the only remaining field in the 1780 Internet Protocol that is common to IPv4 and IPv6 and still has 1781 potential to work end-to-end, with tunnels and with lower layers. 1782 Therefore, ECT(1) should not be reassigned to a different 1783 experimental use (L4S) without carefully assessing competing 1784 potential uses. These fall into the following categories: 1786 C.1. Integrity of Congestion Feedback 1788 Receiving hosts can fool a sender into downloading faster by 1789 suppressing feedback of ECN marks (or of losses if retransmissions 1790 are not necessary or available otherwise). 1792 The historic ECN nonce protocol [RFC3540] proposed that a TCP sender 1793 could set either of ECT(0) or ECT(1) in each packet of a flow and 1794 remember the sequence it had set. If any packet was lost or 1795 congestion marked, the receiver would miss that bit of the sequence. 1796 An ECN Nonce receiver had to feed back the least significant bit of 1797 the sum, so it could not suppress feedback of a loss or mark without 1798 a 50-50 chance of guessing the sum incorrectly. 1800 It is highly unlikely that ECT(1) will be needed for integrity 1801 protection in future. The ECN Nonce RFC [RFC3540] as been 1802 reclassified as historic, partly because other ways have been 1803 developed to protect TCP feedback integrity [RFC8311] that do not 1804 consume a codepoint in the IP header. For instance: 1806 o the sender can test the integrity of the receiver's feedback by 1807 occasionally setting the IP-ECN field to a value normally only set 1808 by the network. Then it can test whether the receiver's feedback 1809 faithfully reports what it expects (see para 2 of Section 20.2 of 1810 [RFC3168]. This works for loss and it will work for the accurate 1811 ECN feedback [RFC7560] intended for L4S; 1813 o A network can enforce a congestion response to its ECN markings 1814 (or packet losses) by auditing congestion exposure (ConEx) 1815 [RFC7713]. Whether the receiver or a downstream network is 1816 suppressing congestion feedback or the sender is unresponsive to 1817 the feedback, or both, ConEx audit can neutralise any advantage 1818 that any of these three parties would otherwise gain. 1820 o The TCP authentication option (TCP-AO [RFC5925]) can be used to 1821 detect any tampering with TCP congestion feedback (whether 1822 malicious or accidental). TCP's congestion feedback fields are 1823 immutable end-to-end, so they are amenable to TCP-AO protection, 1824 which covers the main TCP header and TCP options by default. 1825 However, TCP-AO is often too brittle to use on many end-to-end 1826 paths, where middleboxes can make verification fail in their 1827 attempts to improve performance or security, e.g. by 1828 resegmentation or shifting the sequence space. 1830 C.2. Notification of Less Severe Congestion than CE 1832 Various researchers have proposed to use ECT(1) as a less severe 1833 congestion notification than CE, particularly to enable flows to fill 1834 available capacity more quickly after an idle period, when another 1835 flow departs or when a flow starts, e.g. VCP [VCP], Queue View (QV) 1836 [QV]. 1838 Before assigning ECT(1) as an identifer for L4S, we must carefully 1839 consider whether it might be better to hold ECT(1) in reserve for 1840 future standardisation of rapid flow acceleration, which is an 1841 important and enduring problem [RFC6077]. 1843 Pre-Congestion Notification (PCN) is another scheme that assigns 1844 alternative semantics to the ECN field. It uses ECT(1) to signify a 1845 less severe level of pre-congestion notification than CE [RFC6660]. 1846 However, the ECN field only takes on the PCN semantics if packets 1847 carry a Diffserv codepoint defined to indicate PCN marking within a 1848 controlled environment. PCN is required to be applied solely to the 1849 outer header of a tunnel across the controlled region in order not to 1850 interfere with any end-to-end use of the ECN field. Therefore a PCN 1851 region on the path would not interfere with any of the L4S service 1852 identifiers proposed in Appendix B. 1854 Authors' Addresses 1856 Koen De Schepper 1857 Nokia Bell Labs 1858 Antwerp 1859 Belgium 1861 Email: koen.de_schepper@nokia.com 1862 URI: https://www.bell-labs.com/usr/koen.de_schepper 1864 Bob Briscoe (editor) 1865 CableLabs 1866 UK 1868 Email: ietf@bobbriscoe.net 1869 URI: http://bobbriscoe.net/