idnits 2.17.1 draft-ietf-aqm-recommendation-05.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 : ---------------------------------------------------------------------------- ** There is 1 instance of too long lines in the document, the longest one being 2 characters in excess of 72. -- The draft header indicates that this document obsoletes RFC2309, but the abstract doesn't seem to directly say this. It does mention RFC2309 though, so this could be OK. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year -- The document date (June 24, 2014) is 3586 days in the past. Is this intentional? Checking references for intended status: Best Current Practice ---------------------------------------------------------------------------- (See RFCs 3967 and 4897 for information about using normative references to lower-maturity documents in RFCs) == Unused Reference: 'Jain94' is defined on line 1029, but no explicit reference was found in the text == Unused Reference: 'Nic12' is defined on line 1049, but no explicit reference was found in the text == Unused Reference: 'Sut99' is defined on line 1126, but no explicit reference was found in the text ** Obsolete normative reference: RFC 5405 (Obsoleted by RFC 8085) -- Obsolete informational reference (is this intentional?): RFC 793 (Obsoleted by RFC 9293) -- Obsolete informational reference (is this intentional?): RFC 896 (Obsoleted by RFC 7805) -- Obsolete informational reference (is this intentional?): RFC 2309 (Obsoleted by RFC 7567) -- Obsolete informational reference (is this intentional?): RFC 2460 (Obsoleted by RFC 8200) -- Obsolete informational reference (is this intentional?): RFC 4960 (Obsoleted by RFC 9260) Summary: 2 errors (**), 0 flaws (~~), 4 warnings (==), 7 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Network Working Group F. Baker, Ed. 3 Internet-Draft Cisco Systems 4 Obsoletes: 2309 (if approved) G. Fairhurst, Ed. 5 Intended status: Best Current Practice University of Aberdeen 6 Expires: December 24, 2014 June 24, 2014 8 IETF Recommendations Regarding Active Queue Management 9 draft-ietf-aqm-recommendation-05 11 Abstract 13 This memo presents recommendations to the Internet community 14 concerning measures to improve and preserve Internet performance. It 15 presents a strong recommendation for testing, standardization, and 16 widespread deployment of active queue management (AQM) in network 17 devices, to improve the performance of today's Internet. It also 18 urges a concerted effort of research, measurement, and ultimate 19 deployment of AQM mechanisms to protect the Internet from flows that 20 are not sufficiently responsive to congestion notification. 22 The note largely repeats the recommendations of RFC 2309, updated 23 after fifteen years of experience and new research. 25 Status of This Memo 27 This Internet-Draft is submitted in full conformance with the 28 provisions of BCP 78 and BCP 79. 30 Internet-Drafts are working documents of the Internet Engineering 31 Task Force (IETF). Note that other groups may also distribute 32 working documents as Internet-Drafts. The list of current Internet- 33 Drafts is at http://datatracker.ietf.org/drafts/current/. 35 Internet-Drafts are draft documents valid for a maximum of six months 36 and may be updated, replaced, or obsoleted by other documents at any 37 time. It is inappropriate to use Internet-Drafts as reference 38 material or to cite them other than as "work in progress." 40 This Internet-Draft will expire on December 18, 2014. 42 Copyright Notice 44 Copyright (c) 2014 IETF Trust and the persons identified as the 45 document authors. All rights reserved. 47 This document is subject to BCP 78 and the IETF Trust's Legal 48 Provisions Relating to IETF Documents 49 (http://trustee.ietf.org/license-info) in effect on the date of 50 publication of this document. Please review these documents 51 carefully, as they describe your rights and restrictions with respect 52 to this document. Code Components extracted from this document must 53 include Simplified BSD License text as described in Section 4.e of 54 the Trust Legal Provisions and are provided without warranty as 55 described in the Simplified BSD License. 57 Table of Contents 59 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2 60 1.1. Requirements Language . . . . . . . . . . . . . . . . . . 4 61 2. The Need For Active Queue Management . . . . . . . . . . . . 4 62 2.1. AQM and Multiple Queues . . . . . . . . . . . . . . . . . 8 63 2.2. AQM and Explicit Congestion Marking (ECN) . . . . . . . . 8 64 2.3. AQM and Buffer Size . . . . . . . . . . . . . . . . . . . 9 65 3. Managing Aggressive Flows . . . . . . . . . . . . . . . . . . 9 66 4. Conclusions and Recommendations . . . . . . . . . . . . . . . 12 67 4.1. Operational deployments SHOULD use AQM procedures . . . . 13 68 4.2. Signaling to the transport endpoints . . . . . . . . . . 13 69 4.2.1. AQM and ECN . . . . . . . . . . . . . . . . . . . . . 14 70 4.3. AQM algorithms deployed SHOULD NOT require operational 71 tuning . . . . . . . . . . . . . . . . . . . . . . . . . 15 72 4.4. AQM algorithms SHOULD respond to measured congestion, not 73 application profiles. . . . . . . . . . . . . . . . . . . 17 74 4.5. AQM algorithms SHOULD NOT be dependent on specific 75 transport protocol behaviours . . . . . . . . . . . . . . 17 76 4.6. Interactions with congestion control algorithms . . . . . 18 77 4.7. The need for further research . . . . . . . . . . . . . . 19 78 5. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 20 79 6. Security Considerations . . . . . . . . . . . . . . . . . . . 20 80 7. Privacy Considerations . . . . . . . . . . . . . . . . . . . 20 81 8. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 20 82 9. References . . . . . . . . . . . . . . . . . . . . . . . . . 21 83 9.1. Normative References . . . . . . . . . . . . . . . . . . 21 84 9.2. Informative References . . . . . . . . . . . . . . . . . 22 85 Appendix A. Change Log . . . . . . . . . . . . . . . . . . . . . 25 86 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 26 88 1. Introduction 90 The Internet protocol architecture is based on a connectionless end- 91 to-end packet service using the Internet Protocol, whether IPv4 92 [RFC0791] or IPv6 [RFC2460]. The advantages of its connectionless 93 design: flexibility and robustness, have been amply demonstrated. 94 However, these advantages are not without cost: careful design is 95 required to provide good service under heavy load. In fact, lack of 96 attention to the dynamics of packet forwarding can result in severe 97 service degradation or "Internet meltdown". This phenomenon was 98 first observed during the early growth phase of the Internet in the 99 mid 1980s [RFC0896][RFC0970], and is technically called "congestive 100 collapse". 102 The original fix for Internet meltdown was provided by Van Jacobsen. 103 Beginning in 1986, Jacobsen developed the congestion avoidance 104 mechanisms [Jacobson88] that are now required for implementations of 105 the Transport Control Protocol (TCP) [RFC0768] [RFC1122]. These 106 mechanisms operate in Internet hosts to cause TCP connections to 107 "back off" during congestion. We say that TCP flows are "responsive" 108 to congestion signals (i.e., marked or dropped packets) from the 109 network. It is primarily these TCP congestion avoidance algorithms 110 that prevent the congestive collapse of today's Internet. Similar 111 algorithms are specified for other non-TCP transports. 113 However, that is not the end of the story. Considerable research has 114 been done on Internet dynamics since 1988, and the Internet has 115 grown. It has become clear that the congestion avoidance mechanisms 116 [RFC5681], while necessary and powerful, are not sufficient to 117 provide good service in all circumstances. Basically, there is a 118 limit to how much control can be accomplished from the edges of the 119 network. Some mechanisms are needed in the network devices to 120 complement the endpoint congestion avoidance mechanisms. These 121 mechanisms may be implemented in network devices that include 122 routers, switches, and other network middleboxes. 124 It is useful to distinguish between two classes of algorithms related 125 to congestion control: "queue management" versus "scheduling" 126 algorithms. To a rough approximation, queue management algorithms 127 manage the length of packet queues by marking or dropping packets 128 when necessary or appropriate, while scheduling algorithms determine 129 which packet to send next and are used primarily to manage the 130 allocation of bandwidth among flows. While these two mechanisms are 131 closely related, they address different performance issues and 132 operate on different timescales. Both may be used in combination. 134 This memo highlights two performance issues: 136 The first issue is the need for an advanced form of queue management 137 that we call "Active Queue Management", AQM. Section 2 summarizes 138 the benefits that active queue management can bring. A number of AQM 139 procedures are described in the literature, with different 140 characteristics. This document does not recommend any of them in 141 particular, but does make recommendations that ideally would affect 142 the choice of procedure used in a given implementation. 144 The second issue, discussed in Section 3 of this memo, is the 145 potential for future congestive collapse of the Internet due to flows 146 that are unresponsive, or not sufficiently responsive, to congestion 147 indications. Unfortunately, while scheduling can mitigate some of 148 the side-effects of sharing a network queue with an unresponsive 149 flow, there is currently no consensus solution to controlling the 150 congestion caused by such aggressive flows. Methods such as 151 congestion exposure (ConEx) [RFC6789] offer a framework [CONEX] that 152 can update network devices to alleviate these effcects. Significant 153 research and engineering will be required before any solution will be 154 available. It is imperative that work to mitigate the impact of 155 unresponsive flows is energetically pursued, to ensure the future 156 stability of the Internet. 158 Section 4 concludes the memo with a set of recommendations to the 159 Internet community concerning these topics. 161 The discussion in this memo applies to "best-effort" traffic, which 162 is to say, traffic generated by applications that accept the 163 occasional loss, duplication, or reordering of traffic in flight. It 164 also applies to other traffic, such as real-time traffic that can 165 adapt its sending rate to reduce loss and/or delay. It is most 166 effective when the adaption occurs on time scales of a single Round 167 Trip Time (RTT) or a small number of RTTs, for elastic traffic 168 [RFC1633]. 170 [RFC2309] resulted from past discussions of end-to-end performance, 171 Internet congestion, and Random Early Discard (RED) in the End-to-End 172 Research Group of the Internet Research Task Force (IRTF). This 173 update results from experience with this and other algorithms, and 174 the AQM discussion within the IETF[AQM-WG]. 176 1.1. Requirements Language 178 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 179 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 180 document are to be interpreted as described in [RFC2119]. 182 2. The Need For Active Queue Management 184 Active Queue Management (AQM) is a method that allows network devices 185 to control the queue length or the mean time that a packet spends in 186 a queue. Although AQM can be applied across a range of deployment 187 enviroments, the recommendations in this document are directed to use 188 in the general Internet. It is expected that the principles and 189 guidance are also applicable to a wide range of environments, but may 190 require tuning for specific types of link/network (e.g. to 191 accommodate the traffic patterns found in data centres, the 192 challenges of wireless infrastructure, or the higher delay 193 encountered on satellite Internet links). The remainder of this 194 section identifies the need for AQM and the advantages of deploying 195 the method. 197 The traditional technique for managing the queue length in a network 198 device is to set a maximum length (in terms of packets) for each 199 queue, accept packets for the queue until the maximum length is 200 reached, then reject (drop) subsequent incoming packets until the 201 queue decreases because a packet from the queue has been transmitted. 202 This technique is known as "tail drop", since the packet that arrived 203 most recently (i.e., the one on the tail of the queue) is dropped 204 when the queue is full. This method has served the Internet well for 205 years, but it has two important drawbacks: 207 1. Full Queues 209 The tail drop discipline allows queues to maintain a full (or, 210 almost full) status for long periods of time, since tail drop 211 signals congestion (via a packet drop) only when the queue has 212 become full. It is important to reduce the steady-state queue 213 size, and this is perhaps the most important goal for queue 214 management. 216 The naive assumption might be that there is a simple tradeoff 217 between delay and throughput, and that the recommendation that 218 queues be maintained in a "non-full" state essentially translates 219 to a recommendation that low end-to-end delay is more important 220 than high throughput. However, this does not take into account 221 the critical role that packet bursts play in Internet 222 performance. For example, even though TCP constrains the 223 congestion window of a flow, packets often arrive at network 224 devices in bursts [Leland94]. If the queue is full or almost 225 full, an arriving burst will cause multiple packets to be 226 dropped. This can result in a global synchronization of flows 227 throttling back, followed by a sustained period of lowered link 228 utilization, reducing overall throughput. 230 The point of buffering in the network is to absorb data bursts 231 and to transmit them during the (hopefully) ensuing bursts of 232 silence. This is essential to permit transmission of bursts of 233 data. Normally small queues are preferred in network devices, 234 with sufficient queue capacity to absorb the bursts. The 235 counter-intuitive result is that maintaining normally-small 236 queues can result in higher throughput as well as lower end-to- 237 end delay. In summary, queue limits should not reflect the 238 steady state queues we want to be maintained in the network; 239 instead, they should reflect the size of bursts that a network 240 device needs to absorb. 242 2. Lock-Out 244 In some situations tail drop allows a single connection or a few 245 flows to monopolize the queue space starving other connection 246 preventing them from getting room in the queue. 248 3. Control loop synchronisation 250 Congestion control, like other end-to-end mechanisms, introduces 251 a control loop between hosts. Sessions that share a common network 252 bottleneck can therefore become synchronised, introducing 253 periodic disruption (e.g. jitter/loss). "lock-out" is often also 254 the result of synchronization or other timing effects. 256 Besides tail drop, two alternative queue management disciplines that 257 can be applied when a queue becomes full are "random drop on full" or 258 "head drop on full". When a new packet arrives at a full queue using 259 the random drop on full discipline, the network device drops a 260 randomly selected packet from the queue (which can be an expensive 261 operation, since it naively requires an O(N) walk through the packet 262 queue). When a new packet arrives at a full queue using the head 263 drop on full discipline, the network device drops the packet at the 264 front of the queue [Lakshman96]. Both of these solve the lock-out 265 problem, but neither solves the full-queues problem described above. 267 We know in general how to solve the full-queues problem for 268 "responsive" flows, i.e., those flows that throttle back in response 269 to congestion notification. In the current Internet, dropped packets 270 provide a critical mechanism indicating congestion notification to 271 hosts. The solution to the full-queues problem is for network 272 devices to drop packets before a queue becomes full, so that hosts 273 can respond to congestion before buffers overflow. We call such a 274 proactive approach AQM. By dropping packets before buffers overflow, 275 AQM allows network devices to control when and how many packets to 276 drop. 278 In summary, an active queue management mechanism can provide the 279 following advantages for responsive flows. 281 1. Reduce number of packets dropped in network devices 283 Packet bursts are an unavoidable aspect of packet networks 284 [Willinger95]. If all the queue space in a network device is 285 already committed to "steady state" traffic or if the buffer 286 space is inadequate, then the network device will have no ability 287 to buffer bursts. By keeping the average queue size small, AQM 288 will provide greater capacity to absorb naturally-occurring 289 bursts without dropping packets. 291 Furthermore, without AQM, more packets will be dropped when a 292 queue does overflow. This is undesirable for several reasons. 293 First, with a shared queue and the tail drop discipline, this can 294 result in unnecessary global synchronization of flows, resulting 295 in lowered average link utilization, and hence lowered network 296 throughput. Second, unnecessary packet drops represent a waste 297 of network capacity on the path before the drop point. 299 While AQM can manage queue lengths and reduce end-to-end latency 300 even in the absence of end-to-end congestion control, it will be 301 able to reduce packet drops only in an environment that continues 302 to be dominated by end-to-end congestion control. 304 2. Provide a lower-delay interactive service 306 By keeping a small average queue size, AQM will reduce the delays 307 experienced by flows. This is particularly important for 308 interactive applications such as short web transfers, POP/IMAP, 309 DNS, terminal traffic (telnet, ssh, mosh, RDP, etc), gaming or 310 interactive audio-video sessions, whose subjective (and 311 objective) performance is better when the end-to-end delay is 312 low. 314 3. Avoid lock-out behavior 316 AQM can prevent lock-out behavior by ensuring that there will 317 almost always be a buffer available for an incoming packet. For 318 the same reason, AQM can prevent a bias against low capacity, but 319 highly bursty, flows. 321 Lock-out is undesirable because it constitutes a gross unfairness 322 among groups of flows. However, we stop short of calling this 323 benefit "increased fairness", because general fairness among 324 flows requires per-flow state, which is not provided by queue 325 management. For example, in a network device using AQM with only 326 FIFO scheduling, two TCP flows may receive very different share 327 of the network capacity simply because they have different round- 328 trip times [Floyd91], and a flow that does not use congestion 329 control may receive more capacity than a flow that does. AQM can 330 therefore be combined with a scheduling mechanism that divides 331 network traffic between multiple queues (section 2.1). 333 4. Reduce the probability of control loop synchronisation 334 The probability of network control loop synchronisation can be 335 reduced by introducing randomness in the AQM functions used by 336 network devices that trigger congestion avoidance at the sending 337 host. 339 2.1. AQM and Multiple Queues 341 A network device may use per-flow or per-class queuing with a 342 scheduling algorithm to either prioritise certain applications or 343 classes of traffic, or to provide isolation between different traffic 344 flows within a common class. For example, a router may maintain per- 345 flow state to achieve general fairness by a per-flow scheduling 346 algorithm such as various forms of Fair Queueing (FQ) [Dem90], 347 including Weighted Fair Queuing (WFQ), Stochastic Fairness Queueing 348 (SFQ) [McK90] Deficit Round Robin (DRR) [Shr96] and/or a Class-Based 349 Queue scheduling algorithm such as CBQ [Floyd95]. Hierarchical 350 queues may also be used e.g., as a part of a Hierarchical Token 351 Bucket (HTB), or Hierarchical Fair Service Curve (HFSC) [Sto97] . 352 These methods are also used to realise a range of Quality of Service 353 (QoS) behaviours designed to + meet the need of traffic classes (e.g. 354 using the integrated or differentiated service models). 356 AQM is needed even for network devices that use per-flow or per-class 357 queuing, because scheduling algorithms by themselves do not control 358 the overall queue size or the size of individual queues. AQM 359 mechanisms need to control the overall queue sizes, to ensure that 360 arriving bursts can be accommodated without dropping packets. AQM 361 should also be used to control the queue size for each individual 362 flow or class, so that they do not experience unnecessarily high 363 delay. Using a combination of AQM and scheduling between multiple 364 queues has been shown to offer good results in experimental and some 365 types of operational use. 367 In short, scheduling algorithms and queue management should be seen 368 as complementary, not as replacements for each other. 370 2.2. AQM and Explicit Congestion Marking (ECN) 372 An AQM method may use Explicit Congestion Notification (ECN) 373 [RFC3168] instead of dropping to mark packets under mild or moderate 374 congestion. ECN-marking can allow a network device to signal 375 congestion at a point before a transport experiences congestion loss 376 or additional queuing delay [ECN-Benefit]. Section 4.2.1 describes 377 some of the benefits of using ECN with AQM. 379 2.3. AQM and Buffer Size 381 It is important to differentiate the choice of buffer size for a 382 queue in a switch/router or other network device, and the 383 threshold(s) and other parameters that determine how and when an AQM 384 algorithm operates. One the one hand, the optimum buffer size is a 385 function of operational requirements and should generally be sized to 386 be sufficient to buffer the largest normal traffic burst that is 387 expected. This size depends on the number and burstiness of traffic 388 arriving at the queue and the rate at which traffic leaves the queue. 389 Different types of traffic and deployment scenarios will lead to 390 different requirements. 392 AQM frees a designer from having to the limit buffer space to achieve 393 acceptable performance, allowing allocation of sufficient buffering 394 to satisfy the needs of the particular traffic pattern. On the other 395 hand, the choice of AQM algorithm and associated parameters is a 396 function of the way in which congestion is experienced and the 397 required reaction to achieve acceptable performance. This latter 398 topic is the primary topic of the following sections. 400 3. Managing Aggressive Flows 402 One of the keys to the success of the Internet has been the 403 congestion avoidance mechanisms of TCP. Because TCP "backs off" 404 during congestion, a large number of TCP connections can share a 405 single, congested link in such a way that link bandwidth is shared 406 reasonably equitably among similarly situated flows. The equitable 407 sharing of bandwidth among flows depends on all flows running 408 compatible congestion avoidance algorithms, i.e., methods conformant 409 with the current TCP specification [RFC5681]. 411 In this document a flow is known as "TCP-friendly" when it has a 412 congestion response that approximates the average response expected 413 of a TCP flow. One example method of a TCP-friendly scheme is the 414 TCP-Friendly Rate Control algorithm [RFC5348]. In this document, the 415 term is used more generally to describe this and other algorithms 416 that meet these goals. 418 It is convenient to divide flows into three classes: (1) TCP Friendly 419 flows, (2) unresponsive flows, i.e., flows that do not slow down when 420 congestion occurs, and (3) flows that are responsive but are not TCP- 421 friendly. The last two classes contain more aggressive flows that 422 pose significant threats to Internet performance, which we will now 423 discuss. 425 1. TCP-Friendly flows 426 A TCP-friendly flow responds to congestion notification within a 427 small number of path Round Trip Times (RTT), and in steady-state 428 it uses no more capacity than a conformant TCP running under 429 comparable conditions (drop rate, RTT, packet size, etc.). This 430 is described in the remainder of the document. 432 2. Non-Responsive Flows 434 The User Datagram Protocol (UDP) [RFC0768] provides a minimal, 435 best-effort transport to applications and upper-layer protocols 436 (both simply called "applications" in the remainder of this 437 document) and does not itself provide mechanisms to prevent 438 congestion collapse and establish a degree of fairness [RFC5405]. 440 There is a growing set of UDP-based applications whose congestion 441 avoidance algorithms are inadequate or nonexistent (i.e, a flow 442 that does not throttle its sending rate when it experiences 443 congestion). Examples include some UDP streaming applications 444 for packet voice and video, and some multicast bulk data 445 transport. If no action is taken, such unresponsive flows could 446 lead to a new congestive collapse [RFC2309]. 447 In general, UDP-based applications need to incorporate effective 448 congestion avoidance mechanisms [RFC5405]. Further research and 449 development of ways to accomplish congestion avoidance for 450 presently unresponsive applications continue to be important. 451 Network devices need to be able to protect themselves against 452 unresponsive flows, and mechanisms to accomplish this must be 453 developed and deployed. Deployment of such mechanisms would 454 provide an incentive for all applications to become responsive by 455 either using a congestion-controlled transport (e.g. TCP, SCTP 456 [RFC4960] and DCCP [RFC4340].) or by incorporating their own 457 congestion control in the application [RFC5405]. 458 Lastly, some applications (e.g. current web browsers) open a 459 large numbers of short TCP flows for a single session. This can 460 lead to each individual flow spending the majority of time in the 461 exponential TCP slow start phase, rather than in TCP congestion 462 avoidance. The resulting traffic aggregate can therefore be much 463 less responsive than a single standard TCP flow. 465 3. Non-TCP-friendly Transport Protocols 467 A second threat is posed by transport protocol implementations 468 that are responsive to congestion, but, either deliberately or 469 through faulty implementation, are not TCP-friendly. Such 470 applications may gain an unfair share of the available network 471 capacity. 473 For example, the popularity of the Internet has caused a 474 proliferation in the number of TCP implementations. Some of 475 these may fail to implement the TCP congestion avoidance 476 mechanisms correctly because of poor implementation. Others may 477 deliberately be implemented with congestion avoidance algorithms 478 that are more aggressive in their use of capacity than other TCP 479 implementations; this would allow a vendor to claim to have a 480 "faster TCP". The logical consequence of such implementations 481 would be a spiral of increasingly aggressive TCP implementations, 482 leading back to the point where there is effectively no 483 congestion avoidance and the Internet is chronically congested. 485 Another example could be an RTP/UDP video flow that uses an 486 adaptive codec, but responds incompletely to indications of 487 congestion or responds over an excessively long time period. 488 Such flows are unlikely to be responsive to congestion signals in 489 a timeframe comparable to a small number of end-to-end 490 transmission delays. However, over a longer timescale, perhaps 491 seconds in duration, they could moderate their speed, or increase 492 their speed if they determine capacity to be available. 494 Tunneled traffic aggregates carrying multiple (short) TCP flows 495 can be more aggressive than standard bulk TCP. Applications 496 (e.g. web browsers and peer-to-peer file-sharing) have exploited 497 this by opening multiple connections to the same endpoint. 499 The projected increase in the fraction of total Internet traffic for 500 more aggressive flows in classes 2 and 3 clearly poses a threat to 501 future Internet stability. There is an urgent need for measurements 502 of current conditions and for further research into the ways of 503 managing such flows. This raises many difficult issues in 504 identifying and isolating unresponsive or non-TCP-friendly flows at 505 an acceptable overhead cost. Finally, there is as yet little 506 measurement or simulation evidence available about the rate at which 507 these threats are likely to be realized, or about the expected 508 benefit of algorithms for managing such flows. 510 Another topic requiring consideration is the appropriate 511 granugranularity of a "flow" when considering a queue management 512 method. There are a few "natural" answers: 1) a transport (e.g. TCP 513 or UDP) flow (source address/port, destination address/port, 514 protocol); 2) Differentiated Services Code Point, DSCP; 3) a source/ 515 destination host pair (IP address); 4) a given source host or a given 516 destination host, or various combinations of the above. 518 The source/destination host pair gives an appropriate granularity in 519 many circumstances, However, different vendors/providers use 520 different granularities for defining a flow (as a way of 521 "distinguishing" themselves from one another), and different 522 granularities may be chosen for different places in the network. It 523 may be the case that the granularity is less important than the fact 524 that a network device needs to be able to deal with more unresponsive 525 flows at *some* granularity. The granularity of flows for congestion 526 management is, at least in part, a question of policy that needs to 527 be addressed in the wider IETF community. 529 4. Conclusions and Recommendations 531 The IRTF, in publishing [RFC2309], and the IETF in subsequent 532 discussion, has developed a set of specific recommendations regarding 533 the implementation and operational use of AQM procedures. The 534 updated recommendations provided by this document are summarised as: 536 1. Network devices SHOULD implement some AQM mechanism to manage 537 queue lengths, reduce end-to-end latency, and avoid lock-out 538 phenomena within the Internet. 540 2. Deployed AQM algorithms SHOULD support Explicit Congestion 541 Notification (ECN) as well as loss to signal congestion to 542 endpoints. 544 3. The algorithms that the IETF recommends SHOULD NOT require 545 operational (especially manual) configuration or tuning. 547 4. AQM algorithms SHOULD respond to measured congestion, not 548 application profiles. 550 5. AQM algorithms SHOULD NOT interpret specific transport protocol 551 behaviours. 553 6. Transport protocol congestion control algorithms SHOULD maximize 554 their use of available capacity (when there is data to send) 555 without incurring undue loss or undue round trip delay. 557 7. Research, engineering, and measurement efforts are needed 558 regarding the design of mechanisms to deal with flows that are 559 unresponsive to congestion notification or are responsive, but 560 are more aggressive than present TCP. 562 These recommendations are expressed using the word "SHOULD". This is 563 in recognition that there may be use cases that have not been 564 envisaged in this document in which the recommendation does not 565 apply. Therefore, care should be taken in concluding that one's use 566 case falls in that category; during the life of the Internet, such 567 use cases have been rarely if ever observed and reported. To the 568 contrary, available research [Choi04] says that even high speed links 569 in network cores that are normally very stable in depth and behavior 570 experience occasional issues that need moderation. The 571 recommendations are detailed in the following sections. 573 4.1. Operational deployments SHOULD use AQM procedures 575 AQM procedures are designed to minimize the delay and buffer 576 exhaustion induced in the network by queues that have filled as a 577 result of host behavior. Marking and loss behaviors provide a signal 578 that buffers within network devices are becoming unnecessarily full, 579 and that the sender would do well to moderate its behavior. 581 The use of scheduling mechanisms, such as priority queuing, classful 582 queuing, and fair queuing, is often effective in networks to help a 583 network serve the needs of a range of applications. Network 584 operators can use these methods to manage traffic passing a choke 585 point. This is discussed in [RFC2474] and [RFC2475]. When 586 scheduling is used AQM should be applied across the classes or flows 587 as well as within each class or flow: 589 o AQM mechanisms need to control the overall queue sizes, to ensure 590 that arriving bursts can be accommodated without dropping packets. 592 o AQM mechanisms need to allow combination with other mechanisms, 593 such as scheduling, to allow implementation of polices for 594 providing fairness between different flows. 596 o AQM should be used to control the queue size for each individual 597 flow or class, so that they do not experience unnecessarily high 598 delay. 600 4.2. Signaling to the transport endpoints 602 There are a number of ways a network device may signal to the end 603 point that the network is becoming congested and trigger a reduction 604 in rate. The signalling methods include: 606 o Delaying transport segments (packets) in flight, such as in a 607 queue. 609 o Dropping transport segments (packets) in transit. 611 o Marking transport segments (packets), such as using Explicit 612 Congestion Control[RFC3168] [RFC4301] [RFC4774] [RFC6040] 613 [RFC6679]. 615 Increased network latency is used as an implicit signal of 616 congestion. E.g., in TCP additional delay can affect ACK Clocking 617 and has the result of reducing the rate of transmission of new data. 618 In the Real Time Protocol (RTP), network latency impacts the RTCP- 619 reported RTT and increased latency can trigger a sender to adjust its 620 rate. Methods such as Low Extra Delay Background Transport (LEDBAT) 621 [RFC6817] assume increased latency as a primary signal of congestion. 622 Appropriate use of delay-based methods and the implications of AQM 623 presently remains an area for further research. 625 It is essential that all Internet hosts respond to loss [RFC5681], 626 [RFC5405][RFC4960][RFC4340]. Packet dropping by network devices that 627 are under load has two effects: It protects the network, which is the 628 primary reason that network devices drop packets. The detection of 629 loss also provides a signal to a reliable transport (e.g. TCP, SCTP) 630 that there is potential congestion using a pragmatic heuristic; "when 631 the network discards a message in flight, it may imply the presence 632 of faulty equipment or media in a path, and it may imply the presence 633 of congestion. To be conservative, a transport must assume it may be 634 the latter." Unreliable transports (e.g. using UDP) need to 635 similarly react to loss [RFC5405] 637 Network devices SHOULD use an AQM algorithm to determine the packets 638 that are marked or discarded due to congestion. Procedures for 639 dropping or marking packets within the network need to avoid 640 increasing synchronisation events, and hence randomness SHOULD be 641 introduced in the algorithms that generate these congestion signals 642 to the endpoints. 644 Loss also has an effect on the efficiency of a flow and can 645 significantly impact some classes of application. In reliable 646 transports the dropped data must be subsequently retransmitted. 647 While other applications/transports may adapt to the absence of lost 648 data, this still implies inefficient use of available capacity and 649 the dropped traffic can affect other flows. Hence, congestion 650 signalling by loss is not entirely positive; it is a necessary evil. 652 4.2.1. AQM and ECN 654 Explicit Congestion Notification (ECN) [RFC4301] [RFC4774] [RFC6040] 655 [RFC6679] is a network-layer function that allows a transport to 656 receive network congestion information from a network device without 657 incurring the unintended consequences of loss. ECN includes both 658 transport mechanisms and functions implemented in network devices, 659 the latter rely upon using AQM to decider when and whether to ECN- 660 mark. 662 Congestion for ECN-capable transports is signalled by a network 663 device setting the "Congestion Experienced (CE)" codepoint in the IP 664 header. This codepoint is noted by the remote receiving end point 665 and signalled back to the sender using a transport protocol 666 mechanism, allowing the sender to trigger timely congestion control. 667 The decision to set the CE codepoint requires an AQM algorithm 668 configured with a threshold. Non-ECN capable flows (the default) are 669 dropped under congestion. 671 Network devices SHOULD use an AQM algorithm that marks ECN-capable 672 traffic when making decisions about the response to congestion. 673 Network devices need to implement this method by marking ECN-capable 674 traffic or by dropping non-ECN-capable traffic. 676 Safe deployment of ECN requires that network devices drop excessive 677 traffic, even when marked as originating from an ECN-capable 678 transport. This is a necessary safety precaution because: 680 1. A non-conformant, broken or malicious receiver could conceal an 681 ECN mark, and not report this to the sender; 683 2. A non-conformant, broken or malicious sender could ignore a 684 reported ECN mark, as it could ignore a loss without using ECN; 686 3. A malfunctioning or non-conforming network device may "hide" an 687 ECN mark (or fail to correctly set the ECN codepoint at an egress 688 of a network tunnel). 690 In normal operation, such cases should be very uncommon, however 691 overload protection is desirable to protect traffic from 692 misconfigured or malicious use of ECN (e.g. a denial-of-service 693 attack that generates ECN-capable traffic that is unresponsive to CE- 694 marking). 696 An AQM algorithm that supports ECN needs to define the threshold and 697 algorithm for ECN-marking. This threshold MAY differ from that used 698 for dropping packets that are not marked as ECN-capable, and SHOULD 699 be configurable. 701 Network devices SHOULD use an algorithm to drop excessive traffic 702 (e.g. at some level above the threshold for CE-marking), even when 703 the packets are marked as originating from an ECN-capable transport. 705 4.3. AQM algorithms deployed SHOULD NOT require operational tuning 707 A number of AQM algorithms have been proposed. Many require some 708 form of tuning or setting of parameters for initial network 709 conditions. This can make these algorithms difficult to use in 710 operational networks. 712 AQM algorithms need to consider both "initial conditions" and 713 "operational conditions". The former includes values that exist 714 before any experience is gathered about the use of the algorithm, 715 such as the configured speed of interface, support for full duplex 716 communication, interface MTU and other properties of the link. The 717 latter includes information observed from monitoring the size of the 718 queue, experienced queueing delay, rate of packet discard, etc. 720 This document therefore specifies that AQM algorithms that are 721 proposed for deployment in the Internet have the following 722 properties: 724 o SHOULD NOT require tuning of initial or configuration parameters. 725 An algorithm needs to provide a default behaviour that auto-tunes 726 to a reasonable performance for typical network operational 727 conditions. This is expected to ease deployment and operation. 728 Initial conditions, such as the interface rate and MTU size or 729 other values derived from these, MAY be required by an AQM 730 algorithm. 732 o MAY support further manual tuning that could improve performance 733 in a specific deployed network. Algorithms that lack such 734 variables are acceptable, but if such variables exist, they SHOULD 735 be externalized (made visible to the operator). Guidance needs to 736 be provided on the cases where auto-tuning is unlikely to achieve 737 satisfactory performance and to identify the set of parameters 738 that can be tuned. For example, the expected response of an 739 algorithm may need to be configured to accommodate the largest 740 expected Path RTT, since this value can not be known at 741 initialisation. This guidance is expected to enable the algorithm 742 to be deployed in networks that have specific characteristics 743 (paths with variable/larger delay; networks where capacity is 744 impacted by interactions with lower layer mechanisms, etc). 746 o MAY provide logging and alarm signals to assist in identifying if 747 an algorithm using manual or auto-tuning is functioning as 748 expected. (e.g., this could be based on an internal consistency 749 check between input, output, and mark/drop rates over time). This 750 is expected to encourage deployment by default and allow operators 751 to identify potential interactions with other network functions. 753 Hence, self-tuning algorithms are to be preferred. Algorithms 754 recommended for general Internet deployment by the IETF need to be 755 designed so that they do not require operational (especially manual) 756 configuration or tuning. 758 4.4. AQM algorithms SHOULD respond to measured congestion, not 759 application profiles. 761 Not all applications transmit packets of the same size. Although 762 applications may be characterized by particular profiles of packet 763 size this should not be used as the basis for AQM (see next section). 764 Other methods exist, e.g. Differentiated Services queueing, Pre- 765 Congestion Notification (PCN) [RFC5559], that can be used to 766 differentiate and police classes of application. Network devices may 767 combine AQM with these traffic classification mechanisms and perform 768 AQM only on specific queues within a network device. 770 An AQM algorithm should not deliberately try to prejudice the size of 771 packet that performs best (i.e. Preferentially drop/mark based only 772 on packet size). Procedures for selecting packets to mark/drop 773 SHOULD observe the actual or projected time that a packet is in a 774 queue (bytes at a rate being an analog to time). When an AQM 775 algorithm decides whether to drop (or mark) a packet, it is 776 RECOMMENDED that the size of the particular packet should not be 777 taken into account [Byte-pkt]. 779 Applications (or transports) generally know the packet size that they 780 are using and can hence make their judgments about whether to use 781 small or large packets based on the data they wish to send and the 782 expected impact on the delay or throughput, or other performance 783 parameter. When a transport or application responds to a dropped or 784 marked packet, the size of the rate reduction should be proportionate 785 to the size of the packet that was sent [Byte-pkt]. 787 AQM-enabled system MAY instantiate different instances of an AQM 788 algorithm to be applied within the same traffic class. Traffic 789 classes may be differentiated based on an Access Control List (ACL), 790 the packet Differentiated Services Code Point (DSCP) [RFC5559], 791 enabling use of the ECN field (i.e. any of ECT(0), ECT(1) or 792 CE)[RFC3168] [RFC4774], a multi-field (MF) classifier that combines 793 the values of a set of protocol fields (e.g. IP address, transport, 794 ports) or an equivalent codepoint at a lower layer. This 795 recommendation goes beyond what is defined in RFC 3168, by allowing 796 that an implementation MAY use more than one instance of an AQM 797 algorithm to handle both ECN-capable and non-ECN-capable packets. 799 4.5. AQM algorithms SHOULD NOT be dependent on specific transport 800 protocol behaviours 802 In deploying AQM, network devices need to support a range of Internet 803 traffic and SHOULD NOT make implicit assumptions about the 804 characteristics desired by the set transports/applications the 805 network supports. That is, AQM methods should be opaque to the 806 choice of transport and application. 808 AQM algorithms are often evaluated by considering TCP [RFC0793] with 809 a limited number of applications. Although TCP is the predominant 810 transport in the Internet today, this no longer represents a 811 sufficient selection of traffic for verification. There is 812 significant use of UDP [RFC0768] in voice and video services, and 813 some applications find utility in SCTP [RFC4960] and DCCP [RFC4340]. 814 Hence, AQM algorithms should also demonstrate operation with 815 transports other than TCP and need to consider a variety of 816 applications. Selection of AQM algorithms also needs to consider use 817 of tunnel encapsulations that may carry traffic aggregates. 819 AQM algorithms SHOULD NOT target or derive implicit assumptions about 820 the characteristics desired by specific transports/applications. 821 Transports and applications need to respond to the congestion signals 822 provided by AQM (i.e. dropping or ECN-marking) in a timely manner 823 (within a few RTT at the latest). 825 4.6. Interactions with congestion control algorithms 827 Applications and transports need to react to received implicit or 828 explicit signals that indicate the presence of congestion. This 829 section identifies issues that can impact the design of transport 830 protocols when using paths that use AQM. 832 Transport protocols and applications need timely signals of 833 congestion. The time taken to detect and respond to congestion is 834 increased when network devices queue packets in buffers. It can be 835 difficult to detect tail losses at a higher layer and this may 836 sometimes require transport timers or probe packets to detect and 837 respond to such loss. Loss patterns may also impact timely 838 detection, e.g. the time may be reduced when network devices do not 839 drop long runs of packets from the same flow. 841 A common objective of an elastic transport congestion control 842 protocol is to allow an application to deliver the maximum rate of 843 data without inducing excessive delays when packets are queued in a 844 buffers within the network. To achieve this, a transport should try 845 to operate at rate below the inflexion point of the load/delay curve 846 (the bend of what is sometimes called a "hockey-stick" curve). When 847 the congestion window allows the load to approach this bend, the end- 848 to-end delay starts to rise - a result of congestion, as packets 849 probabilistically arrive at non-overlapping times. On the one hand, 850 a transport that operates above this point can experience congestion 851 loss and could also trigger operator activities, such as those 852 discussed in [RFC6057]. On the other hand, a flow may achieve both 853 near-maximum throughput and low latency when it operates close to 854 this knee point, with minimal contribution to router congestion. 855 Choice of an appropriate rate/congestion window can therefore 856 significantly impact the loss and delay experienced by a flow and 857 will impact other flows that share a common network queue. 859 Some applications may send less than permitted by the congestion 860 control window (or rate). Examples include multimedia codecs that 861 stream at some natural rate (or set of rates) or an application that 862 is naturally interactive (e.g., some web applications, gaming, 863 transaction-based protocols). Such applications may have different 864 objectives. They may not wish to maximize throughput, but may desire 865 a lower loss rate or bounded delay. 867 The correct operation of an AQM-enabled network device MUST NOT rely 868 upon specific transport responses to congestion signals. 870 4.7. The need for further research 872 The second recommendation of [RFC2309] called for further research 873 into the interaction between network queues and host applications, 874 and the means of signaling between them. This research has occurred, 875 and we as a community have learned a lot. However, we are not done. 877 We have learned that the problems of congestion, latency and buffer- 878 sizing have not gone away, and are becoming more important to many 879 users. A number of self-tuning AQM algorithms have been found that 880 offer significant advantages for deployed networks. There is also 881 renewed interest in deploying AQM and the potential of ECN. 883 In 2013, an obvious example of further research is the need to 884 consider the use of Map/Reduce applications in data centers; do we 885 need to extend our taxonomy of TCP/SCTP sessions to include not only 886 "mice" and "elephants", but "lemmings". "Lemmings" are flash crowds 887 of "mice" that the network inadvertently try to signal to as if they 888 were elephant flows, resulting in head of line blocking in data 889 center applications. 891 Examples of other required research include: 893 o Research into new AQM and scheduling algorithms. 895 o Appropriate use of delay-based methods and the implications of 896 AQM. 898 o Research into the use of and deployment of ECN alongside AQM. 900 o Tools for enabling AQM (and ECN) deployment and measuring the 901 performance. 903 o Methods for mitigating the impact of non-conformant and malicious 904 flows. 906 o Research to understand the implications of using new network and 907 transport methods on applications. 909 Hence, this document therefore reiterates the call of RFC 2309: we 910 need continuing research as applications develop. 912 5. IANA Considerations 914 This memo asks the IANA for no new parameters. 916 6. Security Considerations 918 While security is a very important issue, it is largely orthogonal to 919 the performance issues discussed in this memo. 921 Many deployed network devices use queueing methods that allow 922 unresponsive traffic to capture network capacity, denying access to 923 other traffic flows. This could potentially be used as a denial-of- 924 service attack. This threat could be reduced in network devices 925 deploy AQM or some form of scheduling. We note, however, that a 926 denial-of-service attack that results in unresponsive traffic flows 927 may be indistinguishable from other traffic flows (e.g. tunnels 928 carrying aggregates of short flows, high-rate isochronous 929 applications). New methods therefore may remain vulnerable, and this 930 document recommends that ongoing research should consider ways to 931 mitigate such attacks. 933 7. Privacy Considerations 935 This document, by itself, presents no new privacy issues. 937 8. Acknowledgements 939 The original recommendation in [RFC2309] was written by the End-to- 940 End Research Group, which is to say Bob Braden, Dave Clark, Jon 941 Crowcroft, Bruce Davie, Steve Deering, Deborah Estrin, Sally Floyd, 942 Van Jacobson, Greg Minshall, Craig Partridge, Larry Peterson, KK 943 Ramakrishnan, Scott Shenker, John Wroclawski, and Lixia Zhang. This 944 is an edited version of that document, with much of its text and 945 arguments unchanged. 947 The need for an updated document was agreed to in the tsvarea meeting 948 at IETF 86. This document was reviewed on the aqm@ietf.org list. 949 Comments were received from Colin Perkins, Richard Scheffenegger, 950 Dave Taht, John Leslie, David Collier-Brown and many others. 952 Gorry Fairhurst was in part supported by the European Community under 953 its Seventh Framework Programme through the Reducing Internet 954 Transport Latency (RITE) project (ICT-317700). 956 9. References 958 9.1. Normative References 960 [Byte-pkt] 961 and Internet Engineering Task Force, Work in Progress, 962 "Byte and Packet Congestion Notification (draft-ietf- 963 tsvwg-byte-pkt-congest)", July 2013. 965 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 966 Requirement Levels", BCP 14, RFC 2119, March 1997. 968 [RFC3168] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition 969 of Explicit Congestion Notification (ECN) to IP", RFC 970 3168, September 2001. 972 [RFC4301] Kent, S. and K. Seo, "Security Architecture for the 973 Internet Protocol", RFC 4301, December 2005. 975 [RFC4774] Floyd, S., "Specifying Alternate Semantics for the 976 Explicit Congestion Notification (ECN) Field", BCP 124, 977 RFC 4774, November 2006. 979 [RFC5405] Eggert, L. and G. Fairhurst, "Unicast UDP Usage Guidelines 980 for Application Designers", BCP 145, RFC 5405, November 981 2008. 983 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion 984 Control", RFC 5681, September 2009. 986 [RFC6040] Briscoe, B., "Tunnelling of Explicit Congestion 987 Notification", RFC 6040, November 2010. 989 [RFC6679] Westerlund, M., Johansson, I., Perkins, C., O'Hanlon, P., 990 and K. Carlberg, "Explicit Congestion Notification (ECN) 991 for RTP over UDP", RFC 6679, August 2012. 993 9.2. Informative References 995 [AQM-WG] "IETF AQM WG", . 997 [CONEX] Mathis, M. and B. Briscoe, "The Benefits to Applications 998 of using Explicit Congestion Notification (ECN)", IETF 999 (Work-in-Progress) draft-ietf-conex-abstract-mech, March 1000 2014. 1002 [Choi04] Sprint ATL, Burlingame, CA, , , , and , "Analysis of 1003 Point-To-Point Packet Delay In an Operational Network", 1004 March 2004. 1006 [Dem90] Demers, A., Keshav, S., and S. Shenker, "Analysis and 1007 Simulation of a Fair Queueing Algorithm, Internetworking: 1008 Research and Experience", SIGCOMM Symposium proceedings on 1009 Communications architectures and protocols , 1990. 1011 [ECN-Benefit] 1012 Welzl, M. and G. Fairhurst, "The Benefits to Applications 1013 of using Explicit Congestion Notification (ECN)", IETF 1014 (Work-in-Progress) , February 2014. 1016 [Floyd91] Floyd, S., "Connections with Multiple Congested Gateways 1017 in Packet-Switched Networks Part 1: One-way Traffic.", 1018 Computer Communications Review , October 1991. 1020 [Floyd95] Floyd, S. and V. Jacobson, "Link-sharing and Resource 1021 Management Models for Packet Networks", IEEE/ACM 1022 Transactions on Networking , August 1995. 1024 [Jacobson88] 1025 Jacobson, V., "Congestion Avoidance and Control", SIGCOMM 1026 Symposium proceedings on Communications architectures and 1027 protocols , August 1988. 1029 [Jain94] Jain, Raj., Ramakrishnan, KK., and Chiu. Dah-Ming, 1030 "Congestion avoidance scheme for computer networks", US 1031 Patent Office 5377327, December 1994. 1033 [Lakshman96] 1034 Lakshman, TV., Neidhardt, A., and T. Ott, "The Drop From 1035 Front Strategy in TCP Over ATM and Its Interworking with 1036 Other Control Features", IEEE Infocomm , 1996. 1038 [Leland94] 1039 Leland, W., Taqqu, M., Willinger, W., and D. Wilson, "On 1040 the Self-Similar Nature of Ethernet Traffic (Extended 1041 Version)", IEEE/ACM Transactions on Networking , February 1042 1994. 1044 [McK90] McKenney, PE. and G. Varghese, "Stochastic Fairness 1045 Queuing", 1046 http://www2.rdrop.com/~paulmck/scalability/paper/ 1047 sfq.2002.06.04.pdf , 1990. 1049 [Nic12] Nichols, K., "Controlling Queue Delay", Communications of 1050 the ACM Vol. 55 No. 11, July, 2012, pp.42-50. , July 2002. 1052 [RFC0768] Postel, J., "User Datagram Protocol", STD 6, RFC 768, 1053 August 1980. 1055 [RFC0791] Postel, J., "Internet Protocol", STD 5, RFC 791, September 1056 1981. 1058 [RFC0793] Postel, J., "Transmission Control Protocol", STD 7, RFC 1059 793, September 1981. 1061 [RFC0896] Nagle, J., "Congestion control in IP/TCP internetworks", 1062 RFC 896, January 1984. 1064 [RFC0970] Nagle, J., "On packet switches with infinite storage", RFC 1065 970, December 1985. 1067 [RFC1122] Braden, R., "Requirements for Internet Hosts - 1068 Communication Layers", STD 3, RFC 1122, October 1989. 1070 [RFC1633] Braden, B., Clark, D., and S. Shenker, "Integrated 1071 Services in the Internet Architecture: an Overview", RFC 1072 1633, June 1994. 1074 [RFC2309] Braden, B., Clark, D., Crowcroft, J., Davie, B., Deering, 1075 S., Estrin, D., Floyd, S., Jacobson, V., Minshall, G., 1076 Partridge, C., Peterson, L., Ramakrishnan, K., Shenker, 1077 S., Wroclawski, J., and L. Zhang, "Recommendations on 1078 Queue Management and Congestion Avoidance in the 1079 Internet", RFC 2309, April 1998. 1081 [RFC2460] Deering, S. and R. Hinden, "Internet Protocol, Version 6 1082 (IPv6) Specification", RFC 2460, December 1998. 1084 [RFC2474] Nichols, K., Blake, S., Baker, F., and D. Black, 1085 "Definition of the Differentiated Services Field (DS 1086 Field) in the IPv4 and IPv6 Headers", RFC 2474, December 1087 1998. 1089 [RFC2475] Blake, S., Black, D., Carlson, M., Davies, E., Wang, Z., 1090 and W. Weiss, "An Architecture for Differentiated 1091 Services", RFC 2475, December 1998. 1093 [RFC4340] Kohler, E., Handley, M., and S. Floyd, "Datagram 1094 Congestion Control Protocol (DCCP)", RFC 4340, March 2006. 1096 [RFC4960] Stewart, R., "Stream Control Transmission Protocol", RFC 1097 4960, September 2007. 1099 [RFC5348] Floyd, S., Handley, M., Padhye, J., and J. Widmer, "TCP 1100 Friendly Rate Control (TFRC): Protocol Specification", RFC 1101 5348, September 2008. 1103 [RFC5559] Eardley, P., "Pre-Congestion Notification (PCN) 1104 Architecture", RFC 5559, June 2009. 1106 [RFC6057] Bastian, C., Klieber, T., Livingood, J., Mills, J., and R. 1107 Woundy, "Comcast's Protocol-Agnostic Congestion Management 1108 System", RFC 6057, December 2010. 1110 [RFC6789] Briscoe, B., Woundy, R., and A. Cooper, "Congestion 1111 Exposure (ConEx) Concepts and Use Cases", RFC 6789, 1112 December 2012. 1114 [RFC6817] Shalunov, S., Hazel, G., Iyengar, J., and M. Kuehlewind, 1115 "Low Extra Delay Background Transport (LEDBAT)", RFC 6817, 1116 December 2012. 1118 [Shr96] Shreedhar, M. and G. Varghese, "Efficient Fair Queueing 1119 Using Deficit Round Robin", IEEE/ACM Transactions on 1120 Networking Vol 4, No. 3 , July 1996. 1122 [Sto97] Stoica, I. and H. Zhang, "A Hierarchical Fair Service 1123 Curve algorithm for Link sharing, real-time and priority 1124 services", ACM SIGCOMM , 1997. 1126 [Sut99] Suter, B., "Buffer Management Schemes for Supporting TCP 1127 in Gigabit Routers with Per-flow Queueing", IEEE Journal 1128 on Selected Areas in Communications Vol. 17 Issue 6, June, 1129 1999, pp. 1159-1169. , 1999. 1131 [Willinger95] 1132 Willinger, W., Taqqu, M., Sherman, R., Wilson, D., and V. 1133 Jacobson, "Self-Similarity Through High-Variability: 1134 Statistical Analysis of Ethernet LAN Traffic at the Source 1135 Level", SIGCOMM Symposium proceedings on Communications 1136 architectures and protocols , August 1995. 1138 Appendix A. Change Log 1140 Initial Version: March 2013 1142 Minor update of the algorithms that the IETF recommends SHOULD NOT 1143 require operational (especially manual) configuration or tuningdate: 1145 April 2013 1147 Major surgery. This draft is for discussion at IETF-87 and expected 1148 to be further updated. 1149 July 2013 1151 -00 WG Draft - Updated transport recommendations; revised deployment 1152 configuration section; numerous minor edits. 1153 Oct 2013 1155 -01 WG Draft - Updated transport recommendations; revised deployment 1156 configuration section; numerous minor edits. 1157 Jan 2014 - Feedback from WG. 1159 -02 WG Draft - Minor edits Feb 2014 - Mainly language fixes. 1161 -03 WG Draft - Minor edits Feb 2013 - Comments from David Collier- 1162 Brown and David Taht. 1164 -04 WG Draft - Minor edits May 2014 - Comments during WGLC: Provided 1165 some introductory subsections to help people (with subsections and 1166 better text). - Written more on the role scheduling. - Clarified 1167 that ECN mark threshold needs to be configurable. - Reworked your 1168 "knee" para. Various updates in response to feedback. 1170 -05 WG Draft - Minor edits June 2014 - New text added to address 1171 further comments, and improve introduction - adding context, 1172 reference to Conex, linking between sections, added text on 1173 synchronisation. 1175 Authors' Addresses 1177 Fred Baker (editor) 1178 Cisco Systems 1179 Santa Barbara, California 93117 1180 USA 1182 Email: fred@cisco.com 1184 Godred Fairhurst (editor) 1185 University of Aberdeen 1186 School of Engineering 1187 Fraser Noble Building 1188 Aberdeen, Scotland AB24 3UE 1189 UK 1191 Email: gorry@erg.abdn.ac.uk 1192 URI: http://www.erg.abdn.ac.uk