idnits 2.17.1 draft-ietf-tsvwg-aqm-dualq-coupled-03.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 -- The document date (January 24, 2018) is 2284 days in the past. Is this intentional? Checking references for intended status: Experimental ---------------------------------------------------------------------------- -- Looks like a reference, but probably isn't: '0' on line 1452 -- Looks like a reference, but probably isn't: '1' on line 1452 == Outdated reference: A later version (-29) exists of draft-ietf-tsvwg-ecn-l4s-id-00 == Outdated reference: A later version (-20) exists of draft-ietf-tsvwg-l4s-arch-00 -- Obsolete informational reference (is this intentional?): RFC 2309 (Obsoleted by RFC 7567) Summary: 0 errors (**), 0 flaws (~~), 3 warnings (==), 4 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Transport Area working group (tsvwg) K. De Schepper 3 Internet-Draft Nokia Bell Labs 4 Intended status: Experimental B. Briscoe, Ed. 5 Expires: July 28, 2018 CableLabs 6 O. Bondarenko 7 Simula Research Lab 8 I. Tsang 9 Nokia 10 January 24, 2018 12 DualQ Coupled AQMs for Low Latency, Low Loss and Scalable Throughput 13 (L4S) 14 draft-ietf-tsvwg-aqm-dualq-coupled-03 16 Abstract 18 Data Centre TCP (DCTCP) was designed to provide predictably low 19 queuing latency, near-zero loss, and throughput scalability using 20 explicit congestion notification (ECN) and an extremely simple 21 marking behaviour on switches. However, DCTCP does not co-exist with 22 existing TCP traffic---DCTCP is so aggressive that existing TCP 23 algorithms approach starvation. So, until now, DCTCP could only be 24 deployed where a clean-slate environment could be arranged, such as 25 in private data centres. This specification defines `DualQ Coupled 26 Active Queue Management (AQM)' to allow scalable congestion controls 27 like DCTCP to safely co-exist with classic Internet traffic. The 28 Coupled AQM ensures that a flow runs at about the same rate whether 29 it uses DCTCP or TCP Reno/Cubic, but without inspecting transport 30 layer flow identifiers. When tested in a residential broadband 31 setting, DCTCP achieved sub-millisecond average queuing delay and 32 zero congestion loss under a wide range of mixes of DCTCP and 33 `Classic' broadband Internet traffic, without compromising the 34 performance of the Classic traffic. The solution also reduces 35 network complexity and eliminates network configuration. 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 July 28, 2018. 54 Copyright Notice 56 Copyright (c) 2018 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 and Scope . . . . . . . . . . . . . . . . . . . . 3 73 1.2. Terminology . . . . . . . . . . . . . . . . . . . . . . . 5 74 1.3. Features . . . . . . . . . . . . . . . . . . . . . . . . 6 75 2. DualQ Coupled AQM . . . . . . . . . . . . . . . . . . . . . . 7 76 2.1. Coupled AQM . . . . . . . . . . . . . . . . . . . . . . . 7 77 2.2. Dual Queue . . . . . . . . . . . . . . . . . . . . . . . 8 78 2.3. Traffic Classification . . . . . . . . . . . . . . . . . 8 79 2.4. Overall DualQ Coupled AQM Structure . . . . . . . . . . . 8 80 2.5. Normative Requirements for a DualQ Coupled AQM . . . . . 11 81 2.5.1. Functional Requirements . . . . . . . . . . . . . . . 11 82 2.5.2. Management Requirements . . . . . . . . . . . . . . . 12 83 3. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 13 84 4. Security Considerations . . . . . . . . . . . . . . . . . . . 13 85 4.1. Overload Handling . . . . . . . . . . . . . . . . . . . . 13 86 4.1.1. Avoiding Classic Starvation: Sacrifice L4S Throughput 87 or Delay? . . . . . . . . . . . . . . . . . . . . . . 14 88 4.1.2. Congestion Signal Saturation: Introduce L4S Drop or 89 Delay? . . . . . . . . . . . . . . . . . . . . . . . 15 90 4.1.3. Protecting against Unresponsive ECN-Capable Traffic . 16 91 5. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 16 92 6. References . . . . . . . . . . . . . . . . . . . . . . . . . 16 93 6.1. Normative References . . . . . . . . . . . . . . . . . . 17 94 6.2. Informative References . . . . . . . . . . . . . . . . . 17 95 Appendix A. Example DualQ Coupled PI2 Algorithm . . . . . . . . 20 96 A.1. Pass #1: Core Concepts . . . . . . . . . . . . . . . . . 20 97 A.2. Pass #2: Overload Details . . . . . . . . . . . . . . . . 25 98 Appendix B. Example DualQ Coupled Curvy RED Algorithm . . . . . 28 99 Appendix C. Guidance on Controlling Throughput Equivalence . . . 34 100 Appendix D. Open Issues . . . . . . . . . . . . . . . . . . . . 35 101 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 36 103 1. Introduction 105 1.1. Problem and Scope 107 Latency is becoming the critical performance factor for many (most?) 108 applications on the public Internet, e.g. interactive Web, Web 109 services, voice, conversational video, interactive video, interactive 110 remote presence, instant messaging, online gaming, remote desktop, 111 cloud-based applications, and video-assisted remote control of 112 machinery and industrial processes. In the developed world, further 113 increases in access network bit-rate offer diminishing returns, 114 whereas latency is still a multi-faceted problem. In the last decade 115 or so, much has been done to reduce propagation time by placing 116 caches or servers closer to users. However, queuing remains a major 117 component of latency. 119 The Diffserv architecture provides Expedited Forwarding [RFC3246], so 120 that low latency traffic can jump the queue of other traffic. 121 However, on access links dedicated to individual sites (homes, small 122 enterprises or mobile devices), often all traffic at any one time 123 will be latency-sensitive and, if all the traffic on a link is marked 124 as EF, Diffserv cannot reduce the delay of any of it. In contrast, 125 the Low Latency Low Loss Scalable throughput (L4S) approach removes 126 the causes of any unnecessary queuing delay. 128 The bufferbloat project has shown that excessively-large buffering 129 (`bufferbloat') has been introducing significantly more delay than 130 the underlying propagation time. These delays appear only 131 intermittently--only when a capacity-seeking (e.g. TCP) flow is long 132 enough for the queue to fill the buffer, making every packet in other 133 flows sharing the buffer sit through the queue. 135 Active queue management (AQM) was originally developed to solve this 136 problem (and others). Unlike Diffserv, which gives low latency to 137 some traffic at the expense of others, AQM controls latency for _all_ 138 traffic in a class. In general, AQMs introduce an increasing level 139 of discard from the buffer the longer the queue persists above a 140 shallow threshold. This gives sufficient signals to capacity-seeking 141 (aka. greedy) flows to keep the buffer empty for its intended 142 purpose: absorbing bursts. However, RED [RFC2309] and other 143 algorithms from the 1990s were sensitive to their configuration and 144 hard to set correctly. So, AQM was not widely deployed. 146 More recent state-of-the-art AQMs, e.g. fq_CoDel [RFC8290], 147 PIE [RFC8033], Adaptive RED [ARED01], are easier to configure, 148 because they define the queuing threshold in time not bytes, so it is 149 invariant for different link rates. However, no matter how good the 150 AQM, the sawtoothing rate of TCP will either cause queuing delay to 151 vary or cause the link to be under-utilized. Even with a perfectly 152 tuned AQM, the additional queuing delay will be of the same order as 153 the underlying speed-of-light delay across the network. Flow-queuing 154 can isolate one flow from another, but it cannot isolate a TCP flow 155 from the delay variations it inflicts on itself, and it has other 156 problems - it overrides the flow rate decisions of variable rate 157 video applications, it does not recognise the flows within IPSec VPN 158 tunnels and it is relatively expensive to implement. 160 It seems that further changes to the network alone will now yield 161 diminishing returns. Data Centre TCP (DCTCP [RFC8257]) teaches us 162 that a small but radical change to TCP is needed to cut two major 163 outstanding causes of queuing delay variability: 165 1. the `sawtooth' varying rate of TCP itself; 167 2. the smoothing delay deliberately introduced into AQMs to permit 168 bursts without triggering losses. 170 The former causes a flow's round trip time (RTT) to vary from about 1 171 to 2 times the base RTT between the machines in question. The latter 172 delays the system's response to change by a worst-case 173 (transcontinental) RTT, which could be hundreds of times the actual 174 RTT of typical traffic from localized CDNs. 176 Latency is not our only concern: 178 3. It was known when TCP was first developed that it would not scale 179 to high bandwidth-delay products. 181 Given regular broadband bit-rates over WAN distances are 182 already [RFC3649] beyond the scaling range of `classic' TCP Reno, 183 `less unscalable' Cubic [I-D.ietf-tcpm-cubic] and 184 Compound [I-D.sridharan-tcpm-ctcp] variants of TCP have been 185 successfully deployed. However, these are now approaching their 186 scaling limits. Unfortunately, fully scalable TCPs such as DCTCP 187 cause `classic' TCP to starve itself, which is why they have been 188 confined to private data centres or research testbeds (until now). 190 This document specifies a `DualQ Coupled AQM' extension that solves 191 the problem of coexistence between scalable and classic flows, 192 without having to inspect flow identifiers. The AQM is not like 193 flow-queuing approaches [RFC8290] that classify packets by flow 194 identifier into numerous separate queues in order to isolate sparse 195 flows from the higher latency in the queues assigned to heavier flow. 196 In contrast, the AQM exploits the behaviour of scalable congestion 197 controls like DCTCP so that every packet in every flow sharing the 198 queue for DCTCP-like traffic can be served with very low latency. 200 This AQM extension can be combined with any single queue AQM that 201 generates a statistical or deterministic mark/drop probability driven 202 by the queue dynamics. In many cases it simplifies the basic control 203 algorithm, and requires little extra processing. Therefore it is 204 believed the Coupled AQM would be applicable and easy to deploy in 205 all types of buffers; buffers in cost-reduced mass-market residential 206 equipment; buffers in end-system stacks; buffers in carrier-scale 207 equipment including remote access servers, routers, firewalls and 208 Ethernet switches; buffers in network interface cards, buffers in 209 virtualized network appliances, hypervisors, and so on. 211 The overall L4S architecture is described in 212 [I-D.ietf-tsvwg-l4s-arch]. The supporting papers [PI2] and [DCttH15] 213 give the full rationale for the AQM's design, both discursively and 214 in more precise mathematical form. 216 1.2. Terminology 218 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 219 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 220 document are to be interpreted as described in [RFC2119]. In this 221 document, these words will appear with that interpretation only when 222 in ALL CAPS. Lower case uses of these words are not to be 223 interpreted as carrying RFC-2119 significance. 225 The DualQ Coupled AQM uses two queues for two services. Each of the 226 following terms identifies both the service and the queue that 227 provides the service: 229 Classic (denoted by subscript C): The `Classic' service is intended 230 for all the behaviours that currently co-exist with TCP Reno (TCP 231 Cubic, Compound, SCTP, etc). 233 Low-Latency, Low-Loss and Scalable (L4S, denoted by subscript L): 234 The `L4S' service is intended for a set of congestion controls 235 with scalable properties such as DCTCP (e.g. 236 Relentless [Mathis09]). 238 Either service can cope with a proportion of unresponsive or less- 239 responsive traffic as well (e.g. DNS, VoIP, etc), just as a single 240 queue AQM can. The DualQ Coupled AQM behaviour is similar to a 241 single FIFO queue with respect to unresponsive and overload traffic. 243 1.3. Features 245 The AQM couples marking and/or dropping across the two queues such 246 that a flow will get roughly the same throughput whichever it uses. 247 Therefore both queues can feed into the full capacity of a link and 248 no rates need to be configured for the queues. The L4S queue enables 249 scalable congestion controls like DCTCP to give stunningly low and 250 predictably low latency, without compromising the performance of 251 competing 'Classic' Internet traffic. Thousands of tests have been 252 conducted in a typical fixed residential broadband setting. Typical 253 experiments used base round trip delays up to 100ms between the data 254 centre and home network, and large amounts of background traffic in 255 both queues. For every L4S packet, the AQM kept the average queuing 256 delay below 1ms (or 2 packets if serialization delay is bigger for 257 slow links), and no losses at all were introduced by the AQM. 258 Details of the extensive experiments will be made available [PI2] 259 [DCttH15]. 261 Subjective testing was also conducted using a demanding panoramic 262 interactive video application run over a stack with DCTCP enabled and 263 deployed on the testbed. Each user could pan or zoom their own high 264 definition (HD) sub-window of a larger video scene from a football 265 match. Even though the user was also downloading large amounts of 266 L4S and Classic data, latency was so low that the picture appeared to 267 stick to their finger on the touchpad (all the L4S data achieved the 268 same ultra-low latency). With an alternative AQM, the video 269 noticeably lagged behind the finger gestures. 271 Unlike Diffserv Expedited Forwarding, the L4S queue does not have to 272 be limited to a small proportion of the link capacity in order to 273 achieve low delay. The L4S queue can be filled with a heavy load of 274 capacity-seeking flows like DCTCP and still achieve low delay. The 275 L4S queue does not rely on the presence of other traffic in the 276 Classic queue that can be 'overtaken'. It gives low latency to L4S 277 traffic whether or not there is Classic traffic, and the latency of 278 Classic traffic does not suffer when a proportion of the traffic is 279 L4S. The two queues are only necessary because DCTCP-like flows 280 cannot keep latency predictably low and keep utilization high if they 281 are mixed with legacy TCP flows, 283 The experiments used the Linux implementation of DCTCP that is 284 deployed in private data centres, without any modification despite 285 its known deficiencies. Nonetheless, certain modifications will be 286 necessary before DCTCP is safe to use on the Internet, which are 287 recorded in Appendix A of [I-D.ietf-tsvwg-ecn-l4s-id]. However, the 288 focus of this specification is to get the network service in place. 289 Then, without any management intervention, applications can exploit 290 it by migrating to scalable controls like DCTCP, which can then 291 evolve _while_ their benefits are being enjoyed by everyone on the 292 Internet. 294 2. DualQ Coupled AQM 296 There are two main aspects to the approach: 298 o the Coupled AQM that addresses throughput equivalence between 299 Classic (e.g. Reno, Cubic) flows and L4S (e.g. DCTCP) flows 301 o the Dual Queue structure that provides latency separation for L4S 302 flows to isolate them from the typically large Classic queue. 304 2.1. Coupled AQM 306 In the 1990s, the `TCP formula' was derived for the relationship 307 between TCP's congestion window, cwnd, and its drop probability, p. 308 To a first order approximation, cwnd of TCP Reno is inversely 309 proportional to the square root of p. 311 TCP Cubic implements a Reno-compatibility mode, which is the only 312 relevant mode for typical RTTs under 20ms as long as the throughput 313 of a single flow is less than about 500Mb/s. Therefore it can be 314 assumed that Cubic traffic behaves similarly to Reno (but with a 315 slightly different constant of proportionality), and the term 316 'Classic' will be used for the collection of Reno-friendly traffic 317 including Cubic in Reno mode. 319 The supporting paper [PI2] includes the derivation of the equivalent 320 rate equation for DCTCP, for which cwnd is inversely proportional to 321 p (not the square root), where in this case p is the ECN marking 322 probability. DCTCP is not the only congestion control that behaves 323 like this, so the term 'L4S' traffic will be used for all similar 324 behaviour. 326 In order to make a DCTCP flow run at roughly the same rate as a Reno 327 TCP flow (all other factors being equal), the drop or marking 328 probability for Classic traffic, p_C has to be distinct from the 329 marking probability for L4S traffic, p_L (in contrast to RFC3168 330 which requires them to be the same). It is necessary to make the 331 Classic drop probability p_C proportional to the square of the L4S 332 marking probability p_L. This makes the Reno flow rate roughly equal 333 the DCTCP flow rate, because it squares the square root of p_C in the 334 Reno rate equation to make it proportional to the straight p_L in the 335 DCTCP rate equation. 337 Stating this as a formula, the relation between Classic drop 338 probability, p_C, and L4S marking probability, p_L needs to take the 339 form: 341 p_C = ( p_L / k )^2 (1) 343 where k is the constant of proportionality. 345 2.2. Dual Queue 347 Classic traffic typically builds a large queue to prevent under- 348 utilization. Therefore a separate queue is provided for L4S traffic, 349 and it is scheduled with priority over Classic. Priority is 350 conditional to prevent starvation of Classic traffic. 352 Nonetheless, coupled marking ensures that giving priority to L4S 353 traffic still leaves the right amount of spare scheduling time for 354 Classic flows to each get equivalent throughput to DCTCP flows (all 355 other factors such as RTT being equal). The algorithm achieves this 356 without having to inspect flow identifiers. 358 2.3. Traffic Classification 360 Both the Coupled AQM and DualQ mechanisms need an identifier to 361 distinguish L and C packets. A separate draft 362 [I-D.ietf-tsvwg-ecn-l4s-id] recommends using the ECT(1) codepoint of 363 the ECN field as this identifier, having assessed various 364 alternatives. An additional process document has proved necessary to 365 make the ECT(1) codepoint available for experimentation [RFC8311]. 367 2.4. Overall DualQ Coupled AQM Structure 369 Figure 1 shows the overall structure that any DualQ Coupled AQM is 370 likely to have. This schematic is intended to aid understanding of 371 the current designs of DualQ Coupled AQMs. However, it is not 372 intended to preclude other innovative ways of satisfying the 373 normative requirements in Section 2.5 that minimally define a DualQ 374 Coupled AQM. 376 The classifier on the left separates incoming traffic between the two 377 queues (L and C). Each queue has its own AQM that determines the 378 likelihood of dropping or marking (p_L and p_C). Nonetheless, the 379 AQM for Classic traffic is implemented in two stages: i) a base stage 380 that outputs an internal probability p' (pronounced p-prime); and ii) 381 a squaring stage that outputs p_C, where 383 p_C = (p')^2. (2) 385 This allows p_L to be coupled to p_C by marking L4S traffic 386 proportionately to the intermediate output from the first stage. 387 Specifically, the output of the base AQM is coupled across to the L 388 queue in proportion to the output of the base AQM: 390 p_CL = k*p', (3) 392 where k is the constant coupling factor (see Appendix C) and p_CL is 393 the output from the coupling between the C queue and the L queue. 395 It can be seen in the following that these two transformations of p' 396 implement the required coupling given in equation (1) earlier. 397 Substituting for p' from equation (3) into (2): 399 p_C = ( p_CL / k )^2. 401 The actual L4S marking probability p_L is the maximum of the coupled 402 output (p_CL) and the output of a native L4S AQM (p'L), shown as 403 '(MAX)' in the schematic. While the output of the Native L4S AQM is 404 high (p'L > p_CL) it will dominate the way L traffic is marked. When 405 the native L4S AQM output is lower, the way L traffic is marked will 406 be driven by the coupling, that is p_L = p_CL. So, whenever the 407 coupling is needed, as required from equation (1): 409 p_C = ( p_L / k )^2. 411 _________ 412 | | ,------. 413 L4S queue | |===>| ECN | 414 ,'| _______|_| |marker|\ 415 <' | | `------'\\ 416 //`' v ^ p_L \\ 417 // ,-------. | \\ 418 // |Native |p'L | \\,. 419 // | L4S |-->(MAX) < | ___ 420 ,----------.// | AQM | ^ p_CL `\|.'Cond-`. 421 | IP-ECN |/ `-------' | / itional \ 422 ==>|Classifier| ,-------. (k*p') [ priority]==> 423 | |\ | Base | | \scheduler/ 424 `----------'\\ | AQM |--->: ,'|`-.___.-' 425 \\ | |p' | <' | 426 \\ `-------' (p'^2) //`' 427 \\ ^ | // 428 \\,. | v p_C // 429 < | _________ .------.// 430 `\| | | | Drop |/ 431 Classic |queue |===>|/mark | 432 __|______| `------' 434 Legend: ===> traffic flow; ---> control dependency. 436 Figure 1: DualQ Coupled AQM Schematic 438 After the AQMs have applied their dropping or marking, the scheduler 439 forwards their packets to the link, giving priority to L4S traffic. 440 Priority has to be conditional in some way (see Section 4.1). Simple 441 strict priority is inappropriate otherwise it could lead the L4S 442 queue to starve the Classic queue. For example, consider the case 443 where a continually busy L4S queue blocks a DNS request in the 444 Classic queue, arbitrarily delaying the start of a new Classic flow. 446 Example DualQ Coupled AQM algorithms called DualPI2 and Curvy RED are 447 given in Appendix A and Appendix B. Either example AQM can be used 448 to couple packet marking and dropping across a dual Q. 450 DualPI2 uses a Proportional-Integral (PI) controller as the Base AQM. 451 Indeed, this Base AQM with just the squared output and no L4S queue 452 can be used as a drop-in replacement for PIE [RFC8033], in which case 453 we call it just PI2 [PI2]. PI2 is a principled simplification of PIE 454 that is both more responsive and more stable in the face of 455 dynamically varying load. 457 Curvy RED is derived from RED [RFC2309], but its configuration 458 parameters are insensitive to link rate and it requires less 459 operations per packet. However, DualPI2 is more responsive and 460 stable over a wider range of RTTs than Curvy RED. As a consequence, 461 DualPI2 has attracted more development attention than Curvy RED, 462 leaving the Curvy RED design incomplete and not so fully evaluated. 464 Both AQMs regulate their queue in units of time not bytes. As 465 already explained, this ensures configuration can be invariant for 466 different drain rates. With AQMs in a dualQ structure this is 467 particularly important because the drain rate of each queue can vary 468 rapidly as flows for the two queues arrive and depart, even if the 469 combined link rate is constant. 471 It would be possible to control the queues with other alternative 472 AQMs, as long as the normative requirements (those expressed in 473 capitals) in Section 2.5 are observed. 475 2.5. Normative Requirements for a DualQ Coupled AQM 477 The following requirements are intended to capture only the essential 478 aspects of a DualQ Coupled AQM. They are intended to be independent 479 of the particular AQMs used for each queue. 481 2.5.1. Functional Requirements 483 In the Dual Queue, L4S packets MUST be given priority over Classic, 484 although priority MUST be bounded in order not to starve Classic 485 traffic. 487 All L4S traffic MUST be ECN-capable. Some Classic traffic might also 488 be ECN-capable. 490 Whatever identifier is used for L4S experiments, 491 [I-D.ietf-tsvwg-ecn-l4s-id] defines the meaning of an ECN marking on 492 L4S traffic, relative to drop of Classic traffic. In order to 493 prevent starvation of Classic traffic by scalable L4S traffic, it 494 says, "The likelihood that an AQM drops a Not-ECT Classic packet 495 (p_C) MUST be roughly proportional to the square of the likelihood 496 that it would have marked it if it had been an L4S packet (p_L)." In 497 other words, in any DualQ Coupled AQM, the power to which p_L is 498 raised in Eqn. (1) MUST be 2. The term 'likelihood' is used to allow 499 for marking and dropping to be either probabilistic or deterministic. 501 The constant of proportionality, k, in Eqn (1) determines the 502 relative flow rates of Classic and L4S flows when the AQM concerned 503 is the bottleneck (all other factors being equal). 504 [I-D.ietf-tsvwg-ecn-l4s-id] says, "The constant of proportionality 505 (k) does not have to be standardised for interoperability, but a 506 value of 2 is RECOMMENDED." 508 Assuming scalable congestion controls for the Internet will be as 509 aggressive as DCTCP, this will ensure their congestion window will be 510 roughly the same as that of a standards track TCP congestion control 511 (Reno) [RFC5681] and other so-called TCP-friendly controls, such as 512 TCP Cubic in its TCP-friendly mode. 514 {ToDo: The requirements for scalable congestion controls on the 515 Internet (termed the TCP Prague requirements) 516 [I-D.ietf-tsvwg-ecn-l4s-id] are not necessarily final. If the 517 aggressiveness of DCTCP is not defined as the benchmark for scalable 518 controls on the Internet, the recommended value of k will also be 519 subject to change.} 521 The choice of k is a matter of operator policy, and operators MAY 522 choose a different value using Table 1 and the guidelines in 523 Appendix C. 525 If multiple users share capacity at a bottleneck (e.g. in the 526 Internet access link of a campus network), the operator's choice of k 527 will determine capacity sharing between the flows of different users. 528 However, on the public Internet, access network operators typically 529 isolate customers from each other with some form of layer-2 530 multiplexing (TDM in DOCSIS, CDMA in 3G) or L3 scheduling (WRR in 531 DSL), rather than relying on TCP to share capacity between customers 532 [RFC0970]. In such cases, the choice of k will solely affect 533 relative flow rates within each customer's access capacity, not 534 between customers. Also, k will not affect relative flow rates at 535 any times when all flows are Classic or all L4S, and it will not 536 affect small flows. 538 2.5.2. Management Requirements 540 By default, a DualQ Coupled AQM SHOULD NOT need any configuration for 541 use at a bottleneck on the public Internet [RFC7567]. The following 542 parameters MAY be operator-configurable, e.g. to tune for non- 543 Internet settings: 545 o Optional packet classifier(s) to use in addition to the ECN field 546 {ToDo: e.g. ARP}; 548 o Expected typical RTT (a parameter for typical or target queuing 549 delay in each queue might be configurable instead); 551 o Expected maximum RTT (a stability parameter that depends on 552 maximum RTT might be configurable instead); 554 o Coupling factor, k; 556 o The limit to the conditional priority of L4S (scheduler-dependent, 557 e.g. the scheduler weight for WRR, or the time-shift for time- 558 shifted FIFO); 560 o The maximum Classic ECN marking probability, p_Cmax, before 561 switching over to drop. 563 An experimental DualQ Coupled AQM SHOULD allow the operator to 564 monitor the following operational statistics: 566 o Bits forwarded (total and per queue per sample interval), from 567 which utilization can be calculated 569 o Q delay (per queue over sample interval) 571 o Total packets arriving, enqueued and dequeued (per queue per 572 sample interval) 574 o ECN packets marked, non-ECN packets dropped, ECN packets dropped 575 (per queue per sample interval), from which marking and dropping 576 probabilities can be calculated 578 o Time and duration of each overload event. 580 The type of statistics produced for variables like Q delay (mean, 581 percentiles, etc.) will depend on implementation constraints. 583 3. IANA Considerations 585 This specification contains no IANA considerations. 587 4. Security Considerations 589 4.1. Overload Handling 591 Where the interests of users or flows might conflict, it could be 592 necessary to police traffic to isolate any harm to the performance of 593 individual flows. However it is hard to avoid unintended side- 594 effects with policing, and in a trusted environment policing is not 595 necessary. Therefore per-flow policing needs to be separable from a 596 basic AQM, as an option under policy control. 598 However, a basic DualQ AQM does at least need to handle overload. A 599 useful objective would be for the overload behaviour of the DualQ AQM 600 to be at least no worse than a single queue AQM. However, a trade- 601 off needs to be made between complexity and the risk of either 602 traffic class harming the other. In each of the following three 603 subsections, an overload issue specific to the DualQ is described, 604 followed by proposed solution(s). 606 Under overload the higher priority L4S service will have to sacrifice 607 some aspect of its performance. Alternative solutions are provided 608 below that each relax a different factor: e.g. throughput, delay, 609 drop. Some of these choices might need to be determined by operator 610 policy or by the developer, rather than by the IETF. {ToDo: Reach 611 consensus on which it is to be in each case.} 613 4.1.1. Avoiding Classic Starvation: Sacrifice L4S Throughput or Delay? 615 Priority of L4S is required to be conditional to avoid total 616 throughput starvation of Classic by heavy L4S traffic. This raises 617 the question of whether to sacrifice L4S throughput or L4S delay (or 618 some other policy) to mitigate starvation of Classic: 620 Sacrifice L4S throughput: By using weighted round robin as the 621 conditional priority scheduler, the L4S service can sacrifice some 622 throughput during overload to guarantee a minimum throughput 623 service for Classic traffic. The scheduling weight of the Classic 624 queue should be small (e.g. 1/16). Then, in most traffic 625 scenarios the scheduler will not interfere and it will not need to 626 - the coupling mechanism and the end-systems will share out the 627 capacity across both queues as if it were a single pool. However, 628 because the congestion coupling only applies in one direction 629 (from C to L), if L4S traffic is over-aggressive or unresponsive, 630 the scheduler weight for Classic traffic will at least be large 631 enough to ensure it does not starve. 633 In cases where the ratio of L4S to Classic flows (e.g. 19:1) is 634 greater than the ratio of their scheduler weights (e.g. 15:1), the 635 L4S flows will get less than an equal share of the capacity, but 636 only slightly. For instance, with the example numbers given, each 637 L4S flow will get (15/16)/19 = 4.9% when ideally each would get 638 1/20=5%. In the rather specific case of an unresponsive flow 639 taking up a large part of the capacity set aside for L4S, using 640 WRR could significantly reduce the capacity left for any 641 responsive L4S flows. 643 Sacrifice L4S Delay: To control milder overload of responsive 644 traffic, particularly when close to the maximum congestion signal, 645 the operator could choose to control overload of the Classic queue 646 by allowing some delay to 'leak' across to the L4S queue. The 647 scheduler can be made to behave like a single First-In First-Out 648 (FIFO) queue with different service times by implementing a very 649 simple conditional priority scheduler that could be called a 650 "time-shifted FIFO" (see the Modifier Earliest Deadline First 651 (MEDF) scheduler of [MEDF]). This scheduler adds tshift to the 652 queue delay of the next L4S packet, before comparing it with the 653 queue delay of the next Classic packet, then it selects the packet 654 with the greater adjusted queue delay. Under regular conditions, 655 this time-shifted FIFO scheduler behaves just like a strict 656 priority scheduler. But under moderate or high overload it 657 prevents starvation of the Classic queue, because the time-shift 658 (tshift) defines the maximum extra queuing delay of Classic 659 packets relative to L4S. 661 The example implementation in Appendix A can implement either policy. 663 4.1.2. Congestion Signal Saturation: Introduce L4S Drop or Delay? 665 To keep the throughput of both L4S and Classic flows roughly equal 666 over the full load range, a different control strategy needs to be 667 defined above the point where one AQM first saturates to a 668 probability of 100% leaving no room to push back the load any harder. 669 If k>1, L4S will saturate first, but saturation can be caused by 670 unresponsive traffic in either queue. 672 The term 'unresponsive' includes cases where a flow becomes 673 temporarily unresponsive, for instance, a real-time flow that takes a 674 while to adapt its rate in response to congestion, or a TCP-like flow 675 that is normally responsive, but above a certain congestion level it 676 will not be able to reduce its congestion window below the minimum of 677 2 segments, effectively becoming unresponsive. (Note that L4S 678 traffic ought to remain responsive below a window of 2 segments (see 679 [I-D.ietf-tsvwg-ecn-l4s-id]). 681 Saturation raises the question of whether to relieve congestion by 682 introducing some drop into the L4S queue or by allowing delay to grow 683 in both queues (which could eventually lead to tail drop too): 685 Drop on Saturation: Saturation can be avoided by setting a maximum 686 threshold for L4S ECN marking (assuming k>1) before saturation 687 starts to make the flow rates of the different traffic types 688 diverge. Above that the drop probability of Classic traffic is 689 applied to all packets of all traffic types. Then experiments 690 have shown that queueing delay can be kept at the target in any 691 overload situation, including with unresponsive traffic, and no 692 further measures are required. 694 Delay on Saturation: When L4S marking saturates, instead of 695 switching to drop, the drop and marking probabilities could be 696 capped. Beyond that, delay will grow either solely in the queue 697 with unresponsive traffic (if WRR is used), or in both queues (if 698 time-shifted FIFO is used). In either case, the higher delay 699 ought to control temporary high congestion. If the overload is 700 more persistent, eventually the combined DualQ will overflow and 701 tail drop will control congestion. 703 The example implementation in Appendix A applies only the "drop on 704 saturation" policy. 706 4.1.3. Protecting against Unresponsive ECN-Capable Traffic 708 Unresponsive traffic has a greater advantage if it is also ECN- 709 capable. The advantage is undetectable at normal low levels of drop/ 710 marking, but it becomes significant with the higher levels of drop/ 711 marking typical during overload. This is an issue whether the ECN- 712 capable traffic is L4S or Classic. 714 This raises the question of whether and when to switch off ECN 715 marking and use solely drop instead, as required by both Section 7 of 716 [RFC3168] and Section 4.2.1 of [RFC7567]. 718 Experiments with the DualPI2 AQM (Appendix A) have shown that 719 introducing 'drop on saturation' at 100% L4S marking addresses this 720 problem with unresponsive ECN as well as addressing the saturation 721 problem. It leaves only a small range of congestion levels where 722 unresponsive traffic gains any advantage from using the ECN 723 capability, and the advantage is hardly detectable [DualQ-Test]. 725 5. Acknowledgements 727 Thanks to Anil Agarwal, Sowmini Varadhan's and Gabi Bracha for 728 detailed review comments particularly of the appendices and 729 suggestions on how to make our explanation clearer. Thanks also to 730 Greg White and Tom Henderson for insights on the choice of schedulers 731 and queue delay measurement techniques. 733 The authors' contributions were originally part-funded by the 734 European Community under its Seventh Framework Programme through the 735 Reducing Internet Transport Latency (RITE) project (ICT-317700). Bob 736 Briscoe's contribution was also part-funded by the Research Council 737 of Norway through the TimeIn project. The views expressed here are 738 solely those of the authors. 740 6. References 741 6.1. Normative References 743 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 744 Requirement Levels", BCP 14, RFC 2119, 745 DOI 10.17487/RFC2119, March 1997, 746 . 748 6.2. Informative References 750 [ARED01] Floyd, S., Gummadi, R., and S. Shenker, "Adaptive RED: An 751 Algorithm for Increasing the Robustness of RED's Active 752 Queue Management", ACIRI Technical Report , August 2001, 753 . 755 [CoDel] Nichols, K. and V. Jacobson, "Controlling Queue Delay", 756 ACM Queue 10(5), May 2012, 757 . 759 [CRED_Insights] 760 Briscoe, B., "Insights from Curvy RED (Random Early 761 Detection)", BT Technical Report TR-TUB8-2015-003, July 762 2015, 763 . 765 [DCttH15] De Schepper, K., Bondarenko, O., Briscoe, B., and I. 766 Tsang, "`Data Centre to the Home': Ultra-Low Latency for 767 All", 2015, . 770 (Under submission) 772 [DualQ-Test] 773 Steen, H., "Destruction Testing: Ultra-Low Delay using 774 Dual Queue Coupled Active Queue Management", Masters 775 Thesis, Dept of Informatics, Uni Oslo , May 2017. 777 [I-D.ietf-tcpm-cubic] 778 Rhee, I., Xu, L., Ha, S., Zimmermann, A., Eggert, L., and 779 R. Scheffenegger, "CUBIC for Fast Long-Distance Networks", 780 draft-ietf-tcpm-cubic-07 (work in progress), November 781 2017. 783 [I-D.ietf-tsvwg-ecn-l4s-id] 784 Schepper, K., Briscoe, B., and I. Tsang, "Identifying 785 Modified Explicit Congestion Notification (ECN) Semantics 786 for Ultra-Low Queuing Delay", draft-ietf-tsvwg-ecn-l4s- 787 id-00 (work in progress), November 2016. 789 [I-D.ietf-tsvwg-l4s-arch] 790 Briscoe, B., Schepper, K., and M. Bagnulo, "Low Latency, 791 Low Loss, Scalable Throughput (L4S) Internet Service: 792 Architecture", draft-ietf-tsvwg-l4s-arch-00 (work in 793 progress), November 2016. 795 [I-D.sridharan-tcpm-ctcp] 796 Sridharan, M., Tan, K., Bansal, D., and D. Thaler, 797 "Compound TCP: A New TCP Congestion Control for High-Speed 798 and Long Distance Networks", draft-sridharan-tcpm-ctcp-02 799 (work in progress), November 2008. 801 [Mathis09] 802 Mathis, M., "Relentless Congestion Control", PFLDNeT'09 , 803 May 2009, . 806 [MEDF] Menth, M., Schmid, M., Heiss, H., and T. Reim, "MEDF - a 807 simple scheduling algorithm for two real-time transport 808 service classes with application in the UTRAN", Proc. IEEE 809 Conference on Computer Communications (INFOCOM'03) Vol.2 810 pp.1116-1122, March 2003. 812 [PI2] De Schepper, K., Bondarenko, O., Briscoe, B., and I. 813 Tsang, "PI2: A Linearized AQM for both Classic and 814 Scalable TCP", ACM CoNEXT'16 , December 2016, 815 . 818 (To appear) 820 [RFC0970] Nagle, J., "On Packet Switches With Infinite Storage", 821 RFC 970, DOI 10.17487/RFC0970, December 1985, 822 . 824 [RFC2309] Braden, B., Clark, D., Crowcroft, J., Davie, B., Deering, 825 S., Estrin, D., Floyd, S., Jacobson, V., Minshall, G., 826 Partridge, C., Peterson, L., Ramakrishnan, K., Shenker, 827 S., Wroclawski, J., and L. Zhang, "Recommendations on 828 Queue Management and Congestion Avoidance in the 829 Internet", RFC 2309, DOI 10.17487/RFC2309, April 1998, 830 . 832 [RFC3168] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition 833 of Explicit Congestion Notification (ECN) to IP", 834 RFC 3168, DOI 10.17487/RFC3168, September 2001, 835 . 837 [RFC3246] Davie, B., Charny, A., Bennet, J., Benson, K., Le Boudec, 838 J., Courtney, W., Davari, S., Firoiu, V., and D. 839 Stiliadis, "An Expedited Forwarding PHB (Per-Hop 840 Behavior)", RFC 3246, DOI 10.17487/RFC3246, March 2002, 841 . 843 [RFC3649] Floyd, S., "HighSpeed TCP for Large Congestion Windows", 844 RFC 3649, DOI 10.17487/RFC3649, December 2003, 845 . 847 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion 848 Control", RFC 5681, DOI 10.17487/RFC5681, September 2009, 849 . 851 [RFC7567] Baker, F., Ed. and G. Fairhurst, Ed., "IETF 852 Recommendations Regarding Active Queue Management", 853 BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015, 854 . 856 [RFC8033] Pan, R., Natarajan, P., Baker, F., and G. White, 857 "Proportional Integral Controller Enhanced (PIE): A 858 Lightweight Control Scheme to Address the Bufferbloat 859 Problem", RFC 8033, DOI 10.17487/RFC8033, February 2017, 860 . 862 [RFC8034] White, G. and R. Pan, "Active Queue Management (AQM) Based 863 on Proportional Integral Controller Enhanced PIE) for 864 Data-Over-Cable Service Interface Specifications (DOCSIS) 865 Cable Modems", RFC 8034, DOI 10.17487/RFC8034, February 866 2017, . 868 [RFC8257] Bensley, S., Thaler, D., Balasubramanian, P., Eggert, L., 869 and G. Judd, "Data Center TCP (DCTCP): TCP Congestion 870 Control for Data Centers", RFC 8257, DOI 10.17487/RFC8257, 871 October 2017, . 873 [RFC8290] Hoeiland-Joergensen, T., McKenney, P., Taht, D., Gettys, 874 J., and E. Dumazet, "The Flow Queue CoDel Packet Scheduler 875 and Active Queue Management Algorithm", RFC 8290, 876 DOI 10.17487/RFC8290, January 2018, 877 . 879 [RFC8311] Black, D., "Relaxing Restrictions on Explicit Congestion 880 Notification (ECN) Experimentation", RFC 8311, 881 DOI 10.17487/RFC8311, January 2018, 882 . 884 Appendix A. Example DualQ Coupled PI2 Algorithm 886 As a first concrete example, the pseudocode below gives the DualPI2 887 algorithm. DualPI2 follows the structure of the DualQ Coupled AQM 888 framework in Figure 1. A simple step threshold (in units of queuing 889 time) is used for the Native L4S AQM, but a ramp is also described as 890 an alternative. And the PI2 algorithm [PI2] is used for the Classic 891 AQM. PI2 is an improved variant of the PIE AQM [RFC8033]. 893 We will introduce the pseudocode in two passes. The first pass 894 explains the core concepts, deferring handling of overload to the 895 second pass. To aid comparison, line numbers are kept in step 896 between the two passes by using letter suffixes where the longer code 897 needs extra lines. 899 A full open source implementation for Linux is available at: 900 https://github.com/olgabo/dualpi2. 902 A.1. Pass #1: Core Concepts 904 The pseudocode manipulates three main structures of variables: the 905 packet (pkt), the L4S queue (lq) and the Classic queue (cq). The 906 pseudocode consists of the following four functions: 908 o initialization code (Figure 2) that sets parameter defaults (the 909 API for setting non-default values is omitted for brevity) 911 o enqueue code (Figure 3) 913 o dequeue code (Figure 4) 915 o code to regularly update the base probability (p) used in the 916 dequeue code (Figure 5). 918 It also uses the following functions that are not shown in full here: 920 o scheduler(), which selects between the head packets of the two 921 queues; the choice of scheduler technology is discussed later; 923 o cq.len() or lq.len() returns the current length (aka. backlog) of 924 the relevant queue in bytes; 926 o cq.time() or lq.time() returns the current queuing delay (aka. 927 sojourn time or service time) of the relevant queue in units of 928 time; 930 Queuing delay could be measured directly by storing a per-packet 931 time-stamp as each packet is enqueued, and subtracting this from the 932 system time when the packet is dequeued. If time-stamping is not 933 easy to introduce with certain hardware, queuing delay could be 934 predicted indirectly by dividing the size of the queue by the 935 predicted departure rate, which might be known precisely for some 936 link technologies (see for example [RFC8034]). 938 In our experiments so far (building on experiments with PIE) on 939 broadband access links ranging from 4 Mb/s to 200 Mb/s with base RTTs 940 from 5 ms to 100 ms, DualPI2 achieves good results with the default 941 parameters in Figure 2. The parameters are categorised by whether 942 they relate to the Base PI2 AQM, the L4S AQM or the framework 943 coupling them together. Variables derived from these parameters are 944 also included at the end of each category. Each parameter is 945 explained as it is encountered in the walk-through of the pseudocode 946 below. 948 1: dualpi2_params_init(...) { % Set input parameter defaults 949 2: % PI2 AQM parameters 950 3: target = 15 ms % PI AQM Classic queue delay target 951 4: Tupdate = 16 ms % PI Classic queue sampling interval 952 5: alpha = 10 Hz^2 % PI integral gain 953 6: beta = 100 Hz^2 % PI proportional gain 954 7: p_Cmax = 1/4 % Max Classic drop/mark prob 955 8: % Derived PI2 AQM variables 956 9: alpha_U = alpha *Tupdate % PI integral gain per update interval 957 10: beta_U = beta * Tupdate % PI prop'nal gain per update interval 958 11: 959 12: % DualQ Coupled framework parameters 960 13: k = 2 % Coupling factor 961 14: % scheduler weight or equival't parameter (scheduler-dependent) 962 15: limit = MAX_LINK_RATE * 250 ms % Dual buffer size 963 16: 964 17: % L4S AQM parameters 965 18: T_time = 1 ms % L4S marking threshold in time 966 19: T_len = 2 * MTU % Min L4S marking threshold in bytes 967 20: % Derived L4S AQM variables 968 21: p_Lmax = min(k*sqrt(p_Cmax), 1) % Max L4S marking prob 969 22: } 971 Figure 2: Example Header Pseudocode for DualQ Coupled PI2 AQM 973 The overall goal of the code is to maintain the base probability (p), 974 which is an internal variable from which the marking and dropping 975 probabilities for L4S and Classic traffic (p_L and p_C) are derived. 976 The variable named p in the pseudocode and in this walk-through is 977 the same as p' (p-prime) in Section 2.4. The probabilities p_L and 978 p_C are derived in lines 3, 4 and 5 of the dualpi2_update() function 979 (Figure 5) then used in the dualpi2_dequeue() function (Figure 4). 981 The code walk-through below builds up to explaining that part of the 982 code eventually, but it starts from packet arrival. 984 1: dualpi2_enqueue(lq, cq, pkt) { % Test limit and classify lq or cq 985 2: if ( lq.len() + cq.len() > limit ) 986 3: drop(pkt) % drop packet if buffer is full 987 4: else { % Packet classifier 988 5: if ( ecn(pkt) modulo 2 == 1 ) % ECN bits = ECT(1) or CE 989 6: lq.enqueue(pkt) 990 7: else % ECN bits = not-ECT or ECT(0) 991 8: cq.enqueue(pkt) 992 9: } 993 10: } 995 Figure 3: Example Enqueue Pseudocode for DualQ Coupled PI2 AQM 997 1: dualpi2_dequeue(lq, cq, pkt) { % Couples L4S & Classic queues 998 2: while ( lq.len() + cq.len() > 0 ) 999 3: if ( scheduler() == lq ) { 1000 4: lq.dequeue(pkt) % Scheduler chooses lq 1001 5: if ( ((lq.time() > T_time) % step marking ... 1002 6: AND (lq.len() > T_len)) 1003 7: OR (p_CL > rand()) ) % ...or linear marking 1004 8: mark(pkt) 1005 9: } else { 1006 10: cq.dequeue(pkt) % Scheduler chooses cq 1007 11: if ( p_C > rand() ) { % probability p_C = p^2 1008 12: if ( ecn(pkt) == 0 ) { % if ECN field = not-ECT 1009 13: drop(pkt) % squared drop 1010 14: continue % continue to the top of the while loop 1011 15: } 1012 16: mark(pkt) % squared mark 1013 17: } 1014 18: } 1015 19: return(pkt) % return the packet and stop 1016 20: } 1017 21: return(NULL) % no packet to dequeue 1018 22: } 1020 Figure 4: Example Dequeue Pseudocode for DualQ Coupled PI2 AQM 1022 When packets arrive, first a common queue limit is checked as shown 1023 in line 2 of the enqueuing pseudocode in Figure 3. Note that the 1024 limit is deliberately tested before enqueue to avoid any bias against 1025 larger packets (so the actual buffer has to be one MTU larger than 1026 limit). If limit is not exceeded, the packet will be classified and 1027 enqueued to the Classic or L4S queue dependent on the least 1028 significant bit of the ECN field in the IP header (line 5). Packets 1029 with a codepoint having an LSB of 0 (Not-ECT and ECT(0)) will be 1030 enqueued in the Classic queue. Otherwise, ECT(1) and CE packets will 1031 be enqueued in the L4S queue. Optional additional packet 1032 classification flexibility is omitted for brevity. 1034 The dequeue pseudocode (Figure 4) schedules one packet for dequeuing 1035 (or zero if the queue is empty). It also makes all the AQM decisions 1036 on dropping and marking. The alternative of applying the AQMs at 1037 enqueue would shift some processing from the critical time when each 1038 packet is dequeued. However, it would also add a whole queue of 1039 delay to the control signals, making the control loop very sloppy. 1041 All the dequeue code is contained within a large while loop so that 1042 if it decides to drop a packet, it will continue until it selects a 1043 packet to schedule. Line 3 of the dequeue pseudocode is where the 1044 scheduler chooses between the L4S queue (lq) and the Classic queue 1045 (cq). Detailed implementation of the scheduler is not shown (see 1046 discussion later). 1048 o If an L4S packet is scheduled, lines 5 to 8 mark the packet if 1049 either the L4S threshold (T_time) is exceeded, or if a random 1050 marking decision is drawn according to p_CL (maintained by the 1051 dualpi2_update() function discussed below). This logical 'OR' on 1052 a per-packet basis implements the max() function shown in Figure 1 1053 to couple the outputs of the two AQMs together. The L4S threshold 1054 is usually in units of time (default T_time = 1 ms). However, on 1055 slow links the packet serialization time can approach the 1056 threshold T_time, so line 6 sets a floor of T_len (=2 MTU) to the 1057 threshold, otherwise marking is always too frequent on slow links. 1059 o If a Classic packet is scheduled, lines 10 to 17 drop or mark the 1060 packet based on the squared probability p_C. 1062 There is some concern that using a step function for the Native L4S 1063 AQM requires end-systems to smooth the signal for a lot longer - 1064 until its fidelity is sufficient. The latency benefits of a ramp are 1065 being investigated as a simple alternative to the step. This ramp 1066 would be similar to the RED algorithm, with the following 1067 differences: 1069 o The min and max of the ramp are defined in units of queuing delay, 1070 not bytes, so that configuration remains invariant as the queue 1071 departure rate varies. 1073 o It uses instantaneous queueing delay without smoothing (smoothing 1074 is done in the end-systems). 1076 o Determinism is being experimented with instead of randomness; to 1077 reduce the delay necessary to smooth out the noise of randomness 1078 from the signal. For each packet, the algorithm would accumulate 1079 p'_L in a counter and mark the packet that took the counter over 1080 1, then subtract 1 from the counter and continue. 1082 o The ramp rises linearly directly from 0 to 1, not to a an 1083 intermediate value of p'_L as RED would, because there is no need 1084 to keep ECN marking probability low. 1086 This ramp algorithm would require two configuration parameters (min 1087 and max threshold in units of queuing time), in contrast to the 1088 single parameter of a step. 1090 1: dualpi2_update(lq, cq, target) { % Update p every Tupdate 1091 2: curq = cq.time() % use queuing time of first-in Classic packet 1092 3: p = p + alpha_U * (curq - target) + beta_U * (curq - prevq) 1093 4: p_CL = p * k % Coupled L4S prob = base prob * coupling factor 1094 5: p_C = p^2 % Classic prob = (base prob)^2 1095 6: prevq = curq 1096 7: } 1098 Figure 5: Example PI-Update Pseudocode for DualQ Coupled PI2 AQM 1100 The base probability (p) is kept up to date by the core PI algorithm 1101 in Figure 5, which is executed every Tupdate. 1103 Note that p solely depends on the queuing time in the Classic queue. 1104 In line 2, the current queuing delay (curq) is evaluated from how 1105 long the head packet was in the Classic queue (cq). The function 1106 cq.time() (not shown) subtracts the time stamped at enqueue from the 1107 current time and implicitly takes the current queuing delay as 0 if 1108 the queue is empty. 1110 The algorithm centres on line 3, which is a classical Proportional- 1111 Integral (PI) controller that alters p dependent on: a) the error 1112 between the current queuing delay (curq) and the target queuing delay 1113 ('target' - see [RFC8033]); and b) the change in queuing delay since 1114 the last sample. The name 'PI' represents the fact that the second 1115 factor (how fast the queue is growing) is _P_roportional to load 1116 while the first is the _I_ntegral of the load (so it removes any 1117 standing queue in excess of the target). 1119 The two 'gain factors' in line 3, alpha_U and beta_U, respectively 1120 weight how strongly each of these elements ((a) and (b)) alters p. 1121 They are in units of 'per second of delay' or Hz, because they 1122 transform differences in queueing delay into changes in probability. 1124 alpha_U and beta_U are derived from the input parameters alpha and 1125 beta (see lines 5 and 6 of Figure 2). These recommended values of 1126 alpha and beta come from the stability analysis in [PI2] so that the 1127 AQM can change p as fast as possible in response to changes in load 1128 without over-compensating and therefore causing oscillations in the 1129 queue. 1131 alpha and beta determine how much p ought to change if it was updated 1132 every second. It is best to update p as frequently as possible, but 1133 the update interval (Tupdate) will probably be constrained by 1134 hardware performance. For link rates from 4 - 200 Mb/s, we found 1135 Tupdate=16ms (as recommended in [RFC8033]) is sufficient. However 1136 small the chosen value of Tupdate, p should change by the same amount 1137 per second, but in finer more frequent steps. So the gain factors 1138 used for updating p in Figure 5 need to be scaled by (Tupdate/1s), 1139 which is done in lines 9 and 10 of Figure 2). The suffix '_U' 1140 represents 'per update time' (Tupdate). 1142 In corner cases, p can overflow the range [0,1] so the resulting 1143 value of p has to be bounded (omitted from the pseudocode). Then, as 1144 already explained, the coupled and Classic probabilities are derived 1145 from the new p in lines 4 and 5 as p_CL = k*p and p_C = p^2. 1147 Because the coupled L4S marking probability (p_CL) is factored up by 1148 k, the dynamic gain parameters alpha and beta are also inherently 1149 factored up by k for the L4S queue, which is necessary to ensure that 1150 Classic TCP and DCTCP controls have the same stability. So, if alpha 1151 is 10 Hz^2, the effective gain factor for the L4S queue is k*alpha, 1152 which is 20 Hz^2 with the default coupling factor of k=2. 1154 Unlike in PIE [RFC8033], alpha_U and beta_U do not need to be tuned 1155 every Tupdate dependent on p. Instead, in PI2, alpha_U and beta_U 1156 are independent of p because the squaring applied to Classic traffic 1157 tunes them inherently. This is explained in [PI2], which also 1158 explains why this more principled approach removes the need for most 1159 of the heuristics that had to be added to PIE. 1161 {ToDo: Scaling beta with Tupdate and scaling both alpha & beta with 1162 RTT} 1164 A.2. Pass #2: Overload Details 1166 Figure 6 repeats the dequeue function of Figure 4, but with overload 1167 details added. Similarly Figure 7 repeats the core PI algorithm of 1168 Figure 5 with overload details added. The initialization and enqueue 1169 functions are unchanged. 1171 In line 7 of the initialization function (Figure 2), the default 1172 maximum Classic drop probability p_Cmax = 1/4 or 25%. This is the 1173 point at which it is deemed that the Classic queue has become 1174 persistently overloaded, so it switches to using solely drop, even 1175 for ECN-capable packets. This protects the queue against any 1176 unresponsive traffic that falsely claims that it is responsive to ECN 1177 marking, as required by [RFC3168] and [RFC7567]. 1179 Line 21 of the initialization function translates this into a maximum 1180 L4S marking probability (p_Lmax) by rearranging Equation (1). With a 1181 coupling factor of k=2 (the default) or greater, this translates to a 1182 maximum L4S marking probability of 1 (or 100%). This is intended to 1183 ensure that the L4S queue starts to introduce dropping once marking 1184 saturates and can rise no further. The 'TCP Prague' requirements 1185 [I-D.ietf-tsvwg-ecn-l4s-id] state that, when an L4S congestion 1186 control detects a drop, it falls back to a response that coexists 1187 with 'Classic' TCP. So it is correct that the L4S queue drops 1188 packets proportional to p^2, as if they are Classic packets. 1190 Both these switch-overs are triggered by the tests for overload 1191 introduced in lines 4b and 12b of the dequeue function (Figure 6). 1192 Lines 8c to 8g drop L4S packets with probability p^2. Lines 8h to 8i 1193 mark the remaining packets with probability p_CL. 1195 Lines 2c to 2d in the core PI algorithm (Figure 7) deal with overload 1196 of the L4S queue when there is no Classic traffic. This is 1197 necessary, because the core PI algorithm maintains the appropriate 1198 drop probability to regulate overload, but it depends on the length 1199 of the Classic queue. If there is no Classic queue the naive 1200 algorithm in Figure 5 drops nothing, even if the L4S queue is 1201 overloaded - so tail drop would have to take over (lines 3 and 4 of 1202 Figure 3). 1204 If the test at line 2a finds that the Classic queue is empty, line 2d 1205 measures the current queue delay using the L4S queue instead. While 1206 the L4S queue is not overloaded, its delay will always be tiny 1207 compared to the target Classic queue delay. So p_L will be driven to 1208 zero, and the L4S queue will naturally be governed solely by 1209 threshold marking (lines 5 and 6 of the dequeue algorithm in 1210 Figure 6). But, if unresponsive L4S source(s) cause overload, the 1211 DualQ transitions smoothly to L4S marking based on the PI algorithm. 1212 And as overload increases, it naturally transitions from marking to 1213 dropping by the switch-over mechanism already described. 1215 1: dualpi2_dequeue(lq, cq) { % Couples L4S & Classic queues, lq & cq 1216 2: while ( lq.len() + cq.len() > 0 ) 1217 3: if ( scheduler() == lq ) { 1218 4a: lq.dequeue(pkt) 1219 4b: if ( p_CL < p_Lmax ) { % Check for overload saturation 1220 5: if ( ((lq.time() > T_time) % step marking ... 1221 6: AND (lq.len > T_len)) 1222 7: OR (p_CL > rand()) ) % ...or linear marking 1223 8a: mark(pkt) 1224 8b: } else { % overload saturation 1225 8c: if ( p_C > rand() ) { % probability p_C = p^2 1226 8e: drop(pkt) % revert to Classic drop due to overload 1227 8f: continue % continue to the top of the while loop 1228 8g: } 1229 8h: if ( p_CL > rand() ) % probability p_CL = k * p 1230 8i: mark(pkt) % linear marking of remaining packets 1231 8j: } 1232 9: } else { 1233 10: cq.dequeue(pkt) 1234 11: if ( p_C > rand() ) { % probability p_C = p^2 1235 12a: if ( (ecn(pkt) == 0) % ECN field = not-ECT 1236 12b: OR (p_C >= p_Cmax) ) { % Overload disables ECN 1237 13: drop(pkt) % squared drop, redo loop 1238 14: continue % continue to the top of the while loop 1239 15: } 1240 16: mark(pkt) % squared mark 1241 17: } 1242 18: } 1243 19: return(pkt) % return the packet and stop 1244 20: } 1245 21: return(NULL) % no packet to dequeue 1246 22: } 1248 Figure 6: Example Dequeue Pseudocode for DualQ Coupled PI2 AQM 1249 (Including Integer Arithmetic and Overload Code) 1251 1: dualpi2_update(lq, cq, target) { % Update p every Tupdate 1252 2a: if ( cq.len() > 0 ) 1253 2b: curq = cq.time() %use queuing time of first-in Classic packet 1254 2c: else % Classic queue empty 1255 2d: curq = lq.time() % use queuing time of first-in L4S packet 1256 3: p = p + alpha_U * (curq - target) + beta_U * (curq - prevq) 1257 4: p_CL = p * k % L4S prob = base prob * coupling factor 1258 5: p_C = p^2 % Classic prob = (base prob)^2 1259 6: prevq = curq 1260 7: } 1262 Figure 7: Example PI-Update Pseudocode for DualQ Coupled PI2 AQM 1263 (Including Overload Code) 1265 The choice of scheduler technology is critical to overload protection 1266 (see Section 4.1). 1268 o A well-understood weighted scheduler such as weighted round robin 1269 (WRR) is recommended. The scheduler weight for Classic should be 1270 low, e.g. 1/16. 1272 o Alternatively, a time-shifted FIFO could be used. This is a very 1273 simple scheduler, but it does not fully isolate latency in the L4S 1274 queue from uncontrolled bursts in the Classic queue. It works by 1275 selecting the head packet that has waited the longest, biased 1276 against the Classic traffic by a time-shift of tshift. To 1277 implement time-shifted FIFO, the "if (scheduler() == lq )" test in 1278 line 3 of the dequeue code would simply be replaced by "if ( 1279 lq.time() + tshift >= cq.time() )". For the public Internet a 1280 good value for tshift is 50ms. For private networks with smaller 1281 diameter, about 4*target would be reasonable. 1283 o A strict priority scheduler would be inappropriate, because it 1284 would starve Classic if L4S was overloaded. 1286 Appendix B. Example DualQ Coupled Curvy RED Algorithm 1288 As another example of a DualQ Coupled AQM algorithm, the pseudocode 1289 below gives the Curvy RED based algorithm we used and tested. 1290 Although we designed the AQM to be efficient in integer arithmetic, 1291 to aid understanding it is first given using real-number arithmetic. 1292 Then, one possible optimization for integer arithmetic is given, also 1293 in pseudocode. To aid comparison, the line numbers are kept in step 1294 between the two by using letter suffixes where the longer code needs 1295 extra lines. 1297 1: dualq_dequeue(lq, cq) { % Couples L4S & Classic queues, lq & cq 1298 2: if ( lq.dequeue(pkt) ) { 1299 3a: p_L = cq.sec() / 2^S_L 1300 3b: if ( lq.byt() > T ) 1301 3c: mark(pkt) 1302 3d: elif ( p_L > maxrand(U) ) 1303 4: mark(pkt) 1304 5: return(pkt) % return the packet and stop here 1305 6: } 1306 7: while ( cq.dequeue(pkt) ) { 1307 8a: alpha = 2^(-f_C) 1308 8b: Q_C = alpha * pkt.sec() + (1-alpha)* Q_C % Classic Q EWMA 1309 9a: sqrt_p_C = Q_C / 2^S_C 1310 9b: if ( sqrt_p_C > maxrand(2*U) ) 1311 10: drop(pkt) % Squared drop, redo loop 1312 11: else 1313 12: return(pkt) % return the packet and stop here 1314 13: } 1315 14: return(NULL) % no packet to dequeue 1316 15: } 1318 16: maxrand(u) { % return the max of u random numbers 1319 17: maxr=0 1320 18: while (u-- > 0) 1321 19: maxr = max(maxr, rand()) % 0 <= rand() < 1 1322 20: return(maxr) 1323 21: } 1325 Figure 8: Example Dequeue Pseudocode for DualQ Coupled Curvy RED AQM 1327 Packet classification code is not shown, as it is no different from 1328 Figure 3. Potential classification schemes are discussed in 1329 Section 2. The Curvy RED algorithm has not been maintained to the 1330 same degree as the DualPI2 algorithm. Some ideas used in DualPI2 1331 would need to be translated into Curvy RED, such as i) the 1332 conditional priority scheduler instead of strict priority ii) the 1333 time-based L4S threshold; iii) turning off ECN as overload 1334 protection; iv) Classic ECN support. These are not shown in the 1335 Curvy RED pseudocode, but would need to be implemented for 1336 production. {ToDo} 1338 At the outer level, the structure of dualq_dequeue() implements 1339 strict priority scheduling. The code is written assuming the AQM is 1340 applied on dequeue (Note 1) . Every time dualq_dequeue() is called, 1341 the if-block in lines 2-6 determines whether there is an L4S packet 1342 to dequeue by calling lq.dequeue(pkt), and otherwise the while-block 1343 in lines 7-13 determines whether there is a Classic packet to 1344 dequeue, by calling cq.dequeue(pkt). (Note 2) 1345 In the lower priority Classic queue, a while loop is used so that, if 1346 the AQM determines that a classic packet should be dropped, it 1347 continues to test for classic packets deciding whether to drop each 1348 until it actually forwards one. Thus, every call to dualq_dequeue() 1349 returns one packet if at least one is present in either queue, 1350 otherwise it returns NULL at line 14. (Note 3) 1352 Within each queue, the decision whether to drop or mark is taken as 1353 follows (to simplify the explanation, it is assumed that U=1): 1355 L4S: If the test at line 2 determines there is an L4S packet to 1356 dequeue, the tests at lines 3a and 3c determine whether to mark 1357 it. The first is a simple test of whether the L4S queue (lq.byt() 1358 in bytes) is greater than a step threshold T in bytes (Note 4). 1359 The second test is similar to the random ECN marking in RED, but 1360 with the following differences: i) the marking function does not 1361 start with a plateau of zero marking until a minimum threshold, 1362 rather the marking probability starts to increase as soon as the 1363 queue is positive; ii) marking depends on queuing time, not bytes, 1364 in order to scale for any link rate without being reconfigured; 1365 iii) marking of the L4S queue does not depend on itself, it 1366 depends on the queuing time of the _other_ (Classic) queue, where 1367 cq.sec() is the queuing time of the packet at the head of the 1368 Classic queue (zero if empty); iv) marking depends on the 1369 instantaneous queuing time (of the other Classic queue), not a 1370 smoothed average; v) the queue is compared with the maximum of U 1371 random numbers (but if U=1, this is the same as the single random 1372 number used in RED). 1374 Specifically, in line 3a the marking probability p_L is set to the 1375 Classic queueing time qc.sec() in seconds divided by the L4S 1376 scaling parameter 2^S_L, which represents the queuing time (in 1377 seconds) at which marking probability would hit 100%. Then in line 1378 3d (if U=1) the result is compared with a uniformly distributed 1379 random number between 0 and 1, which ensures that marking 1380 probability will linearly increase with queueing time. The 1381 scaling parameter is expressed as a power of 2 so that division 1382 can be implemented as a right bit-shift (>>) in line 3 of the 1383 integer variant of the pseudocode (Figure 9). 1385 Classic: If the test at line 7 determines that there is at least one 1386 Classic packet to dequeue, the test at line 9b determines whether 1387 to drop it. But before that, line 8b updates Q_C, which is an 1388 exponentially weighted moving average (Note 5) of the queuing time 1389 in the Classic queue, where pkt.sec() is the instantaneous 1390 queueing time of the current Classic packet and alpha is the EWMA 1391 constant for the classic queue. In line 8a, alpha is represented 1392 as an integer power of 2, so that in line 8 of the integer code 1393 the division needed to weight the moving average can be 1394 implemented by a right bit-shift (>> f_C). 1396 Lines 9a and 9b implement the drop function. In line 9a the 1397 averaged queuing time Q_C is divided by the Classic scaling 1398 parameter 2^S_C, in the same way that queuing time was scaled for 1399 L4S marking. This scaled queuing time is given the variable name 1400 sqrt_p_C because it will be squared to compute Classic drop 1401 probability, so before it is squared it is effectively the square 1402 root of the drop probability. The squaring is done by comparing 1403 it with the maximum out of two random numbers (assuming U=1). 1404 Comparing it with the maximum out of two is the same as the 1405 logical `AND' of two tests, which ensures drop probability rises 1406 with the square of queuing time (Note 6). Again, the scaling 1407 parameter is expressed as a power of 2 so that division can be 1408 implemented as a right bit-shift in line 9 of the integer 1409 pseudocode. 1411 The marking/dropping functions in each queue (lines 3 & 9) are two 1412 cases of a new generalization of RED called Curvy RED, motivated as 1413 follows. When we compared the performance of our AQM with fq_CoDel 1414 and PIE, we came to the conclusion that their goal of holding queuing 1415 delay to a fixed target is misguided [CRED_Insights]. As the number 1416 of flows increases, if the AQM does not allow TCP to increase queuing 1417 delay, it has to introduce abnormally high levels of loss. Then loss 1418 rather than queuing becomes the dominant cause of delay for short 1419 flows, due to timeouts and tail losses. 1421 Curvy RED constrains delay with a softened target that allows some 1422 increase in delay as load increases. This is achieved by increasing 1423 drop probability on a convex curve relative to queue growth (the 1424 square curve in the Classic queue, if U=1). Like RED, the curve hugs 1425 the zero axis while the queue is shallow. Then, as load increases, 1426 it introduces a growing barrier to higher delay. But, unlike RED, it 1427 requires only one parameter, the scaling, not three. The diadvantage 1428 of Curvy RED is that it is not adapted to a wide range of RTTs. 1429 Curvy RED can be used as is when the RTT range to support is limited 1430 otherwise an adaptation mechanism is required. 1432 There follows a summary listing of the two parameters used for each 1433 of the two queues: 1435 Classic: 1437 S_C : The scaling factor of the dropping function scales Classic 1438 queuing times in the range [0, 2^(S_C)] seconds into a dropping 1439 probability in the range [0,1]. To make division efficient, it 1440 is constrained to be an integer power of two; 1442 f_C : To smooth the queuing time of the Classic queue and make 1443 multiplication efficient, we use a negative integer power of 1444 two for the dimensionless EWMA constant, which we define as 1445 alpha = 2^(-f_C). 1447 L4S : 1449 S_L (and k'): As for the Classic queue, the scaling factor of 1450 the L4S marking function scales Classic queueing times in the 1451 range [0, 2^(S_L)] seconds into a probability in the range 1452 [0,1]. Note that S_L = S_C + k', where k' is the coupling 1453 between the queues. So S_L and k' count as only one parameter; 1454 k' is related to k in Equation (1) (Section 2.1) by k=2^k', 1455 where both k and k' are constants. Then implementations can 1456 avoid costly division by shifting p_L by k' bits to the right. 1458 T : The queue size in bytes at which step threshold marking 1459 starts in the L4S queue. 1461 {ToDo: These are the raw parameters used within the algorithm. A 1462 configuration front-end could accept more meaningful parameters and 1463 convert them into these raw parameters.} 1465 From our experiments so far, recommended values for these parameters 1466 are: S_C = -1; f_C = 5; T = 5 * MTU for the range of base RTTs 1467 typical on the public Internet. [CRED_Insights] explains why these 1468 parameters are applicable whatever rate link this AQM implementation 1469 is deployed on and how the parameters would need to be adjusted for a 1470 scenario with a different range of RTTs (e.g. a data centre) {ToDo 1471 incorporate a summary of that report into this draft}. The setting of 1472 k depends on policy (see Section 2.5 and Appendix C respectively for 1473 its recommended setting and guidance on alternatives). 1475 There is also a cUrviness parameter, U, which is a small positive 1476 integer. It is likely to take the same hard-coded value for all 1477 implementations, once experiments have determined a good value. We 1478 have solely used U=1 in our experiments so far, but results might be 1479 even better with U=2 or higher. 1481 Note that the dropping function at line 9 calls maxrand(2*U), which 1482 gives twice as much curviness as the call to maxrand(U) in the 1483 marking function at line 3. This is the trick that implements the 1484 square rule in equation (1) (Section 2.1). This is based on the fact 1485 that, given a number X from 1 to 6, the probability that two dice 1486 throws will both be less than X is the square of the probability that 1487 one throw will be less than X. So, when U=1, the L4S marking 1488 function is linear and the Classic dropping function is squared. If 1489 U=2, L4S would be a square function and Classic would be quartic. 1490 And so on. 1492 The maxrand(u) function in lines 16-21 simply generates u random 1493 numbers and returns the maximum (Note 7). Typically, maxrand(u) 1494 could be run in parallel out of band. For instance, if U=1, the 1495 Classic queue would require the maximum of two random numbers. So, 1496 instead of calling maxrand(2*U) in-band, the maximum of every pair of 1497 values from a pseudorandom number generator could be generated out- 1498 of-band, and held in a buffer ready for the Classic queue to consume. 1500 1: dualq_dequeue(lq, cq) { % Couples L4S & Classic queues, lq & cq 1501 2: if ( lq.dequeue(pkt) ) { 1502 3: if ((lq.byt() > T) || ((cq.ns() >> (S_L-2)) > maxrand(U))) 1503 4: mark(pkt) 1504 5: return(pkt) % return the packet and stop here 1505 6: } 1506 7: while ( cq.dequeue(pkt) ) { 1507 8: Q_C += (pkt.ns() - Q_C) >> f_C % Classic Q EWMA 1508 9: if ( (Q_C >> (S_C-2) ) > maxrand(2*U) ) 1509 10: drop(pkt) % Squared drop, redo loop 1510 11: else 1511 12: return(pkt) % return the packet and stop here 1512 13: } 1513 14: return(NULL) % no packet to dequeue 1514 15: } 1516 Figure 9: Optimised Example Dequeue Pseudocode for Coupled DualQ AQM 1517 using Integer Arithmetic 1519 Notes: 1521 1. The drain rate of the queue can vary if it is scheduled relative 1522 to other queues, or to cater for fluctuations in a wireless 1523 medium. To auto-adjust to changes in drain rate, the queue must 1524 be measured in time, not bytes or packets [CoDel]. In our Linux 1525 implementation, it was easiest to measure queuing time at 1526 dequeue. Queuing time can be estimated when a packet is enqueued 1527 by measuring the queue length in bytes and dividing by the recent 1528 drain rate. 1530 2. An implementation has to use priority queueing, but it need not 1531 implement strict priority. 1533 3. If packets can be enqueued while processing dequeue code, an 1534 implementer might prefer to place the while loop around both 1535 queues so that it goes back to test again whether any L4S packets 1536 arrived while it was dropping a Classic packet. 1538 4. In order not to change too many factors at once, for now, we keep 1539 the marking function for DCTCP-only traffic as similar as 1540 possible to DCTCP. However, unlike DCTCP, all processing is at 1541 dequeue, so we determine whether to mark a packet at the head of 1542 the queue by the byte-length of the queue _behind_ it. We plan 1543 to test whether using queuing time will work in all 1544 circumstances, and if we find that the step can cause 1545 oscillations, we will investigate replacing it with a steep 1546 random marking curve. 1548 5. An EWMA is only one possible way to filter bursts; other more 1549 adaptive smoothing methods could be valid and it might be 1550 appropriate to decrease the EWMA faster than it increases. 1552 6. In practice at line 10 the Classic queue would probably test for 1553 ECN capability on the packet to determine whether to drop or mark 1554 the packet. However, for brevity such detail is omitted. All 1555 packets classified into the L4S queue have to be ECN-capable, so 1556 no dropping logic is necessary at line 3. Nonetheless, L4S 1557 packets could be dropped by overload code (see Section 4.1). 1559 7. In the integer variant of the pseudocode (Figure 9) real numbers 1560 are all represented as integers scaled up by 2^32. In lines 3 & 1561 9 the function maxrand() is arranged to return an integer in the 1562 range 0 <= maxrand() < 2^32. Queuing times are also scaled up by 1563 2^32, but in two stages: i) In lines 3 and 8 queuing times 1564 cq.ns() and pkt.ns() are returned in integer nanoseconds, making 1565 the values about 2^30 times larger than when the units were 1566 seconds, ii) then in lines 3 and 9 an adjustment of -2 to the 1567 right bit-shift multiplies the result by 2^2, to complete the 1568 scaling by 2^32. 1570 Appendix C. Guidance on Controlling Throughput Equivalence 1572 +---------------+------+-------+ 1573 | RTT_C / RTT_L | Reno | Cubic | 1574 +---------------+------+-------+ 1575 | 1 | k'=1 | k'=0 | 1576 | 2 | k'=2 | k'=1 | 1577 | 3 | k'=2 | k'=2 | 1578 | 4 | k'=3 | k'=2 | 1579 | 5 | k'=3 | k'=3 | 1580 +---------------+------+-------+ 1582 Table 1: Value of k' for which DCTCP throughput is roughly the same 1583 as Reno or Cubic, for some example RTT ratios 1585 k' is related to k in Equation (1) (Section 2.1) by k=2^k'. 1587 To determine the appropriate policy, the operator first has to judge 1588 whether it wants DCTCP flows to have roughly equal throughput with 1589 Reno or with Cubic (because, even in its Reno-compatibility mode, 1590 Cubic is about 1.4 times more aggressive than Reno). Then the 1591 operator needs to decide at what ratio of RTTs it wants DCTCP and 1592 Classic flows to have roughly equal throughput. For example choosing 1593 k'=0 (equivalent to k=1) will make DCTCP throughput roughly the same 1594 as Cubic, _if their RTTs are the same_. 1596 However, even if the base RTTs are the same, the actual RTTs are 1597 unlikely to be the same, because Classic (Cubic or Reno) traffic 1598 needs a large queue to avoid under-utilization and excess drop, 1599 whereas L4S (DCTCP) does not. The operator might still choose this 1600 policy if it judges that DCTCP throughput should be rewarded for 1601 keeping its own queue short. 1603 On the other hand, the operator will choose one of the higher values 1604 for k', if it wants to slow DCTCP down to roughly the same throughput 1605 as Classic flows, to compensate for Classic flows slowing themselves 1606 down by causing themselves extra queuing delay. 1608 The values for k' in the table are derived from the formulae, which 1609 was developed in [DCttH15]: 1611 2^k' = 1.64 (RTT_reno / RTT_dc) (2) 1612 2^k' = 1.19 (RTT_cubic / RTT_dc ) (3) 1614 For localized traffic from a particular ISP's data centre, we used 1615 the measured RTTs to calculate that a value of k'=3 (equivalant to 1616 k=8) would achieve throughput equivalence, and our experiments 1617 verified the formula very closely. 1619 For a typical mix of RTTs from local data centres and across the 1620 general Internet, a value of k'=1 (equivalent to k=2) is recommended 1621 as a good workable compromise. 1623 Appendix D. Open Issues 1625 Most of the following open issues are also tagged '{ToDo}' at the 1626 appropriate point in the document: 1628 Operational guidance to monitor L4S experiment 1630 Interaction between Diffserv & L4S 1632 Define additional classifier flexibility more clearly 1633 PI2 appendix: scaling of alpha & beta, esp. dependence of beta_U 1634 on Tupdate 1636 Curvy RED appendix: complete the unfinished parts 1638 Authors' Addresses 1640 Koen De Schepper 1641 Nokia Bell Labs 1642 Antwerp 1643 Belgium 1645 Email: koen.de_schepper@nokia.com 1646 URI: https://www.bell-labs.com/usr/koen.de_schepper 1648 Bob Briscoe (editor) 1649 CableLabs 1650 UK 1652 Email: ietf@bobbriscoe.net 1653 URI: http://bobbriscoe.net/ 1655 Olga Bondarenko 1656 Simula Research Lab 1657 Lysaker 1658 Norway 1660 Email: olgabnd@gmail.com 1661 URI: https://www.simula.no/people/olgabo 1663 Ing-jyh Tsang 1664 Nokia 1665 Antwerp 1666 Belgium 1668 Email: ing-jyh.tsang@nokia.com