idnits 2.17.1 draft-ietf-aqm-eval-guidelines-10.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 (February 2, 2016) is 3005 days in the past. Is this intentional? Checking references for intended status: Informational ---------------------------------------------------------------------------- == Missing Reference: 'Byte' is mentioned on line 310, but not defined == Outdated reference: A later version (-07) exists of draft-ietf-tcpm-cubic-01 ** Obsolete normative reference: RFC 793 (Obsoleted by RFC 9293) ** Obsolete normative reference: RFC 2679 (Obsoleted by RFC 7679) ** Obsolete normative reference: RFC 2680 (Obsoleted by RFC 7680) -- Obsolete informational reference (is this intentional?): RFC 2309 (Obsoleted by RFC 7567) Summary: 3 errors (**), 0 flaws (~~), 3 warnings (==), 2 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Internet Engineering Task Force N. Kuhn, Ed. 3 Internet-Draft CNES, Telecom Bretagne 4 Intended status: Informational P. Natarajan, Ed. 5 Expires: August 5, 2016 Cisco Systems 6 N. Khademi, Ed. 7 University of Oslo 8 D. Ros 9 Simula Research Laboratory AS 10 February 2, 2016 12 AQM Characterization Guidelines 13 draft-ietf-aqm-eval-guidelines-10 15 Abstract 17 Unmanaged large buffers in today's networks have given rise to a slew 18 of performance issues. These performance issues can be addressed by 19 some form of Active Queue Management (AQM) mechanism, optionally in 20 combination with a packet scheduling scheme such as fair queuing. 21 This document describes various criteria for performing precautionary 22 characterizations of AQM schemes. 24 Status of This Memo 26 This Internet-Draft is submitted in full conformance with the 27 provisions of BCP 78 and BCP 79. 29 Internet-Drafts are working documents of the Internet Engineering 30 Task Force (IETF). Note that other groups may also distribute 31 working documents as Internet-Drafts. The list of current Internet- 32 Drafts is at http://datatracker.ietf.org/drafts/current/. 34 Internet-Drafts are draft documents valid for a maximum of six months 35 and may be updated, replaced, or obsoleted by other documents at any 36 time. It is inappropriate to use Internet-Drafts as reference 37 material or to cite them other than as "work in progress." 39 This Internet-Draft will expire on August 5, 2016. 41 Copyright Notice 43 Copyright (c) 2016 IETF Trust and the persons identified as the 44 document authors. All rights reserved. 46 This document is subject to BCP 78 and the IETF Trust's Legal 47 Provisions Relating to IETF Documents 48 (http://trustee.ietf.org/license-info) in effect on the date of 49 publication of this document. Please review these documents 50 carefully, as they describe your rights and restrictions with respect 51 to this document. Code Components extracted from this document must 52 include Simplified BSD License text as described in Section 4.e of 53 the Trust Legal Provisions and are provided without warranty as 54 described in the Simplified BSD License. 56 Table of Contents 58 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3 59 1.1. Reducing the latency and maximizing the goodput . . . . . 5 60 1.2. Goals of this document . . . . . . . . . . . . . . . . . 5 61 1.3. Requirements Language . . . . . . . . . . . . . . . . . . 6 62 1.4. Glossary . . . . . . . . . . . . . . . . . . . . . . . . 6 63 2. End-to-end metrics . . . . . . . . . . . . . . . . . . . . . 6 64 2.1. Flow completion time . . . . . . . . . . . . . . . . . . 7 65 2.2. Flow start up time . . . . . . . . . . . . . . . . . . . 7 66 2.3. Packet loss . . . . . . . . . . . . . . . . . . . . . . . 7 67 2.4. Packet loss synchronization . . . . . . . . . . . . . . . 8 68 2.5. Goodput . . . . . . . . . . . . . . . . . . . . . . . . . 9 69 2.6. Latency and jitter . . . . . . . . . . . . . . . . . . . 9 70 2.7. Discussion on the trade-off between latency and goodput . 10 71 3. Generic setup for evaluations . . . . . . . . . . . . . . . . 10 72 3.1. Topology and notations . . . . . . . . . . . . . . . . . 11 73 3.2. Buffer size . . . . . . . . . . . . . . . . . . . . . . . 12 74 3.3. Congestion controls . . . . . . . . . . . . . . . . . . . 12 75 4. Methodology, Metrics, AQM Comparisons, Packet Sizes, 76 Scheduling and ECN . . . . . . . . . . . . . . . . . . . . . 13 77 4.1. Methodology . . . . . . . . . . . . . . . . . . . . . . . 13 78 4.2. Comments on metrics measurement . . . . . . . . . . . . . 13 79 4.3. Comparing AQM schemes . . . . . . . . . . . . . . . . . . 14 80 4.3.1. Performance comparison . . . . . . . . . . . . . . . 14 81 4.3.2. Deployment comparison . . . . . . . . . . . . . . . . 15 82 4.4. Packet sizes and congestion notification . . . . . . . . 15 83 4.5. Interaction with ECN . . . . . . . . . . . . . . . . . . 15 84 4.6. Interaction with Scheduling . . . . . . . . . . . . . . . 16 85 5. Transport Protocols . . . . . . . . . . . . . . . . . . . . . 16 86 5.1. TCP-friendly sender . . . . . . . . . . . . . . . . . . . 17 87 5.1.1. TCP-friendly sender with the same initial congestion 88 window . . . . . . . . . . . . . . . . . . . . . . . 17 89 5.1.2. TCP-friendly sender with different initial congestion 90 windows . . . . . . . . . . . . . . . . . . . . . . . 17 91 5.2. Aggressive transport sender . . . . . . . . . . . . . . . 18 92 5.3. Unresponsive transport sender . . . . . . . . . . . . . . 18 93 5.4. Less-than Best Effort transport sender . . . . . . . . . 19 94 6. Round Trip Time Fairness . . . . . . . . . . . . . . . . . . 19 95 6.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . 19 96 6.2. Recommended tests . . . . . . . . . . . . . . . . . . . . 20 97 6.3. Metrics to evaluate the RTT fairness . . . . . . . . . . 20 98 7. Burst Absorption . . . . . . . . . . . . . . . . . . . . . . 20 99 7.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . 20 100 7.2. Recommended tests . . . . . . . . . . . . . . . . . . . . 21 101 8. Stability . . . . . . . . . . . . . . . . . . . . . . . . . . 22 102 8.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . 22 103 8.2. Recommended tests . . . . . . . . . . . . . . . . . . . . 23 104 8.2.1. Definition of the congestion Level . . . . . . . . . 23 105 8.2.2. Mild congestion . . . . . . . . . . . . . . . . . . . 23 106 8.2.3. Medium congestion . . . . . . . . . . . . . . . . . . 23 107 8.2.4. Heavy congestion . . . . . . . . . . . . . . . . . . 24 108 8.2.5. Varying the congestion level . . . . . . . . . . . . 24 109 8.2.6. Varying available capacity . . . . . . . . . . . . . 24 110 8.3. Parameter sensitivity and stability analysis . . . . . . 25 111 9. Various Traffic Profiles . . . . . . . . . . . . . . . . . . 26 112 9.1. Traffic mix . . . . . . . . . . . . . . . . . . . . . . . 26 113 9.2. Bi-directional traffic . . . . . . . . . . . . . . . . . 26 114 10. Multi-AQM Scenario . . . . . . . . . . . . . . . . . . . . . 27 115 10.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . 27 116 10.2. Details on the evaluation scenario . . . . . . . . . . . 27 117 11. Implementation cost . . . . . . . . . . . . . . . . . . . . . 27 118 11.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . 27 119 11.2. Recommended discussion . . . . . . . . . . . . . . . . . 28 120 12. Operator Control and Auto-tuning . . . . . . . . . . . . . . 28 121 12.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . 28 122 12.2. Recommended discussion . . . . . . . . . . . . . . . . . 29 123 13. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . 29 124 14. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 30 125 15. Contributors . . . . . . . . . . . . . . . . . . . . . . . . 30 126 16. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 31 127 17. Security Considerations . . . . . . . . . . . . . . . . . . . 31 128 18. References . . . . . . . . . . . . . . . . . . . . . . . . . 31 129 18.1. Normative References . . . . . . . . . . . . . . . . . . 31 130 18.2. Informative References . . . . . . . . . . . . . . . . . 33 131 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 34 133 1. Introduction 135 Active Queue Management (AQM) [RFC7567] addresses the concerns 136 arising from using unnecessarily large and unmanaged buffers to 137 improve network and application performance. Several AQM algorithms 138 have been proposed in the past years, most notably Random Early 139 Detection (RED), BLUE, and Proportional Integral controller (PI), and 140 more recently CoDel [NICH2012] and PIE [PAN2013]. In general, these 141 algorithms actively interact with the Transmission Control Protocol 142 (TCP) and any other transport protocol that deploys a congestion 143 control scheme to manage the amount of data they keep in the network. 144 The available buffer space in the routers and switches should be 145 large enough to accommodate the short-term buffering requirements. 146 AQM schemes aim at reducing buffer occupancy, and therefore the end- 147 to-end delay. Some of these algorithms, notably RED, have also been 148 widely implemented in some network devices. However, the potential 149 benefits of the RED scheme have not been realized since RED is 150 reported to be usually turned off. The main reason of this 151 reluctance to use RED in today's deployments comes from its 152 sensitivity to the operating conditions in the network and the 153 difficulty of tuning its parameters. 155 A buffer is a physical volume of memory in which a queue or set of 156 queues are stored. When speaking of a specific queue in this 157 document, "buffer occupancy" refers to the amount of data (measured 158 in bytes or packets) that are in the queue, and the "maximum buffer 159 size" refers to the maximum buffer occupancy. In real 160 implementations of switches, a global memory is often shared between 161 the available devices, and thus, the maximum buffer size may vary 162 over the time. 164 Bufferbloat [BB2011] is the consequence of deploying large unmanaged 165 buffers on the Internet -- the buffering has often been measured to 166 be ten times or hundred times larger than needed. Large buffer sizes 167 in combination with TCP and/or unresponsive flows increases end-to- 168 end delay. This results in poor performance for latency-sensitive 169 applications such as real-time multimedia (e.g., voice, video, 170 gaming, etc). The degree to which this affects modern networking 171 equipment, especially consumer-grade equipment's, produces problems 172 even with commonly used web services. Active queue management is 173 thus essential to control queuing delay and decrease network latency. 175 The Active Queue Management and Packet Scheduling Working Group (AQM 176 WG) was chartered to address the problems with large unmanaged 177 buffers in the Internet. Specifically, the AQM WG is tasked with 178 standardizing AQM schemes that not only address concerns with such 179 buffers, but also are robust under a wide variety of operating 180 conditions. This document provides characterization guidelines that 181 can be used to assess the deployability of an AQM, whether it is 182 candidate for standardization at IETF or not. 184 [RFC7567] separately describes the AQM algorithm implemented in a 185 router from the scheduling of packets sent by the router. The rest 186 of this memo refers to the AQM as a dropping/marking policy as a 187 separate feature to any interface scheduling scheme. This document 188 may be complemented with another one on guidelines for assessing 189 combination of packet scheduling and AQM. We note that such a 190 document will inherit all the guidelines from this document plus any 191 additional scenarios relevant for packet scheduling such as flow 192 starvation evaluation or impact of the number of hash buckets. 194 1.1. Reducing the latency and maximizing the goodput 196 The trade-off between reducing the latency and maximizing the goodput 197 is intrinsically linked to each AQM scheme and is key to evaluating 198 its performance. This trade-off MUST be considered in a variety of 199 scenarios to ensure the safety of an AQM deployment. Whenever 200 possible, solutions ought to aim at both maximizing goodput and 201 minimizing latency. 203 1.2. Goals of this document 205 The guidelines help to quantify performance of AQM schemes in terms 206 of latency reduction, goodput maximization and the trade-off between 207 these two. The guidelines also discuss methods to understand the 208 various aspects associated with safely deploying and operating the 209 AQM scheme. These guidelines discuss methods to understand ease of 210 development, deployment and operational aspects of the AQM scheme 211 verses the potential gain in performance from the introduction of the 212 proposed scheme. 214 This memo details generic characterization scenarios against which 215 any AQM proposal should be evaluated, irrespective of whether or not 216 an AQM is standardized by the IETF. This documents recommends the 217 relevant scenarios and metrics to be considered. The document 218 presents central aspects of an AQM algorithm that should be 219 considered whatever the context, such as burst absorption capacity, 220 RTT fairness or resilience to fluctuating network conditions. 222 These guidelines do not cover every possible aspect of a particular 223 algorithm. In addition, it is worth noting that the proposed 224 criteria are not bound to a particular evaluation toolset. These 225 guidelines do not present context-dependent scenarios (such as 802.11 226 WLANs, data-centers or rural broadband networks). To keep the 227 guidelines generic, a number of potential router components and 228 algorithms (such as DiffServ) are omitted. 230 The goals of this document can thus be summarized as follows: 232 o The present characterization guidelines provide a non-exhaustive 233 list of scenario to help ascertain whether an AQM is not only 234 better than drop-tail (with BDP-sized buffer), but also safe to 235 deploy; 237 o The present characterization guidelines (1) are not bound to a 238 particular evaluation toolset and (2) can be used for various 239 deployment scenarios; 241 o The present characterization guidelines provide guidance for 242 better selecting an AQM for a specific environment; it is not 243 required that an AQM proposal is evaluated following these 244 guidelines for its standardization. 246 1.3. Requirements Language 248 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 249 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 250 document are to be interpreted as described in RFC 2119 [RFC2119]. 252 1.4. Glossary 254 o AQM: [RFC7567] separately describes the Active Queue Managment 255 (AQM) algorithm implemented in a router from the scheduling of 256 packets sent by the router. The rest of this memo refers to the 257 AQM as a dropping/marking policy as a separate feature to any 258 interface scheduling scheme. 260 o buffer: a physical volume of memory in which a queue or set of 261 queues are stored. 263 o buffer occupancy: amount of data that are stored in a buffer, 264 measured in bytes or packets. 266 o buffer size: maximum buffer occupancy, that is the maximum amount 267 of data that may be stored in a buffer, measured in bytes or 268 packets. 270 o goodput: goodput is defined as the number of bits per unit of time 271 forwarded to the correct destination minus any bits lost or 272 retransmitted [RFC2647]. 274 o SQRT: the square root function. 276 o ROUND: the round function. 278 2. End-to-end metrics 280 End-to-end delay is the result of propagation delay, serialization 281 delay, service delay in a switch, medium-access delay and queuing 282 delay, summed over the network elements along the path. AQM schemes 283 may reduce the queuing delay by providing signals to the sender on 284 the emergence of congestion, but any impact on the goodput must be 285 carefully considered. This section presents the metrics that could 286 be used to better quantify (1) the reduction of latency, (2) 287 maximization of goodput and (3) the trade-off between these two. 289 This section provides normative requirements for metrics that can be 290 used to assess the performance of an AQM scheme. 292 Some metrics listed in this section are not suited to every type of 293 traffic detailed in the rest of this document. It is therefore not 294 necessary to measure all of the following metrics: the chosen metric 295 may not be relevant to the context of the evaluation scenario (e.g., 296 latency vs. goodput trade-off in application-limited traffic 297 scenarios). Guidance is provided for each metric. 299 2.1. Flow completion time 301 The flow completion time is an important performance metric for the 302 end-user when the flow size is finite. Considering the fact that an 303 AQM scheme may drop/mark packets, the flow completion time is 304 directly linked to the dropping/marking policy of the AQM scheme. 305 This metric helps to better assess the performance of an AQM 306 depending on the flow size. The Flow Completion Time (FCT) is 307 related to the flow size (Fs) and the goodput for the flow (G) as 308 follows: 310 FCT [s] = Fs [Byte] / ( G [Bit/s] / 8 [Bit/Byte] ) 312 If this metric is used to evaluate the performance of web transfers, 313 it is suggested to rather consider the time needed to download all 314 the objects that compose the web page, as this makes more sense in 315 terms of user experience than assessing the time needed to download 316 each object. 318 2.2. Flow start up time 320 The flow start up time is the time between the request has been sent 321 from the client and the server starts to transmit data. The amount 322 of packets dropped by an AQM may seriously affect the waiting period 323 during which the data transfer has not started. This metric would 324 specifically focus on the operations such as DNS lookups, TCP opens 325 of SSL handshakes. 327 2.3. Packet loss 329 Packet loss can occur en-route, this can impact the end-to-end 330 performance measured at receiver. 332 The tester SHOULD evaluate loss experienced at the receiver using one 333 of the two metrics: 335 o the packet loss ratio: this metric is to be frequently measured 336 during the experiment. The long-term loss ratio is of interest 337 for steady-state scenarios only; 339 o the interval between consecutive losses: the time between two 340 losses is to be measured. 342 The packet loss ratio can be assessed by simply evaluating the loss 343 ratio as a function of the number of lost packets and the total 344 number of packets sent. This might not be easily done in laboratory 345 testing, for which these guidelines advice the tester: 347 o to check that for every packet, a corresponding packet was 348 received within a reasonable time, as explained in [RFC2680]. 350 o to keep a count of all packets sent, and a count of the non- 351 duplicate packets received, as explained in the section 10 of 352 [RFC2544]. 354 The interval between consecutive losses, which is also called a gap, 355 is a metric of interest for VoIP traffic and, as a result, has been 356 further specified in [RFC3611]. 358 2.4. Packet loss synchronization 360 One goal of an AQM algorithm is to help to avoid global 361 synchronization of flows sharing a bottleneck buffer on which the AQM 362 operates ([RFC2309],[RFC7567]). The "degree" of packet-loss 363 synchronization between flows SHOULD be assessed, with and without 364 the AQM under consideration. 366 As discussed e.g., in [HASS2008], loss synchronization among flows 367 may be quantified by several slightly different metrics that capture 368 different aspects of the same issue. However, in real-world 369 measurements the choice of metric could be imposed by practical 370 considerations -- e.g., whether fine-grained information on packet 371 losses in the bottleneck available or not. For the purpose of AQM 372 characterization, a good candidate metric is the global 373 synchronization ratio, measuring the proportion of flows losing 374 packets during a loss event. [JAY2006] used this metric in real- 375 world experiments to characterize synchronization along arbitrary 376 Internet paths; the full methodology is described in [JAY2006]. 378 If an AQM scheme is evaluated using real-life network environments, 379 it is worth pointing out that some network events, such as failed 380 link restoration may cause synchronized losses between active flows 381 and thus confuse the meaning of this metric. 383 2.5. Goodput 385 The goodput has been defined in section 3.17 of [RFC2647] as the 386 number of bits per unit of time forwarded to the correct destination 387 interface, minus any bits lost or retransmitted. This definition 388 induces that the test setup needs to be qualified to assure that it 389 is not generating losses on its own. 391 Measuring the end-to-end goodput provides an appreciation of how well 392 an AQM scheme improves transport and application performance. The 393 measured end-to-end goodput is linked to the dropping/marking policy 394 of the AQM scheme -- e.g., the fewer the number of packet drops, the 395 fewer packets need retransmission, minimizing the impact of AQM on 396 transport and application performance. Additionally, an AQM scheme 397 may resort to Explicit Congestion Notification (ECN) marking as an 398 initial means to control delay. Again, marking packets instead of 399 dropping them reduces the number of packet retransmissions and 400 increases goodput. End-to-end goodput values help to evaluate the 401 AQM scheme's effectiveness of an AQM scheme in minimizing packet 402 drops that impact application performance and to estimate how well 403 the AQM scheme works with ECN. 405 The measurement of the goodput allows the tester evaluate to which 406 extent an AQM is able to maintain a high bottleneck utilization. 407 This metric should be also obtained frequently during an experiment 408 as the long-term goodput is relevant for steady-state scenarios only 409 and may not necessarily reflect how the introduction of an AQM 410 actually impacts the link utilization during at a certain period of 411 time. Fluctuations in the values obtained from these measurements 412 may depend on other factors than the introduction of an AQM, such as 413 link layer losses due to external noise or corruption, fluctuating 414 bandwidths (802.11 WLANs), heavy congestion levels or transport 415 layer's rate reduction by congestion control mechanism. 417 2.6. Latency and jitter 419 The latency, or the one-way delay metric, is discussed in [RFC2679]. 420 There is a consensus on an adequate metric for the jitter, that 421 represents the one-way delay variations for packets from the same 422 flow: the Packet Delay Variation (PDV), detailed in [RFC5481], serves 423 well all use cases. 425 The end-to-end latency includes components other than just the 426 queuing delay, such as the signal processing delay, transmission 427 delay and the processing delay. Moreover, the jitter is caused by 428 variations in queuing and processing delay (e.g., scheduling 429 effects). The introduction of an AQM scheme would impact these 430 metrics (end-to-end latency and jitter) and therefore they should be 431 considered in the end-to-end evaluation of performance. 433 2.7. Discussion on the trade-off between latency and goodput 435 The metrics presented in this section may be considered as explained 436 in the rest of this document, in order to discuss and quantify the 437 trade-off between latency and goodput. 439 With regards to the goodput, and in addition to the long-term 440 stationary goodput value, it is RECOMMENDED to take measurements 441 every multiple of the minimum RTT (minRTT) between A and B. It is 442 suggested to take measurements at least every K x minRTT (to smooth 443 out the fluctuations), with K=10. Higher values for K are encouraged 444 whenever it is more appropriate for the presentation of the results. 445 The value for K may depend on the network's path characteristics. 446 The measurement period MUST be disclosed for each experiment and when 447 results/values are compared across different AQM schemes, the 448 comparisons SHOULD use exactly the same measurement periods. With 449 regards to latency, it is RECOMMENDED to take the samples on per- 450 packet basis whenever possible depending on the features provided by 451 hardware/software and the impact of sampling itself on the hardware 452 performance. It is generally RECOMMENDED to provide at least 10 453 samples per RTT. 455 From each of these sets of measurements, the cumulative density 456 function (CDF) of the considered metrics SHOULD be computed. If the 457 considered scenario introduces dynamically varying parameters, 458 temporal evolution of the metrics could also be generated. For each 459 scenario, the following graph may be generated: the x-axis shows 460 queuing delay (that is the average per-packet delay in excess of 461 minimum RTT), the y-axis the goodput. Ellipses are computed such as 462 detailed in [WINS2014]: "We take each individual [...] run [...] as 463 one point, and then compute the 1-epsilon elliptic contour of the 464 maximum-likelihood 2D Gaussian distribution that explains the points. 465 [...] we plot the median per-sender throughput and queueing delay as 466 a circle. [...] The orientation of an ellipse represents the 467 covariance between the throughput and delay measured for the 468 protocol." This graph provides part of a better understanding of (1) 469 the delay/goodput trade-off for a given congestion control mechanism 470 Section 5, and (2) how the goodput and average queue delay vary as a 471 function of the traffic load Section 8.2. 473 3. Generic setup for evaluations 475 This section presents the topology that can be used for each of the 476 following scenarios, the corresponding notations and discusses 477 various assumptions that have been made in the document. 479 3.1. Topology and notations 481 +---------+ +-----------+ 482 |senders A| |receivers B| 483 +---------+ +-----------+ 485 +--------------+ +--------------+ 486 |traffic class1| |traffic class1| 487 |--------------| |--------------| 488 | SEN.Flow1.1 +---------+ +-----------+ REC.Flow1.1 | 489 | + | | | | + | 490 | | | | | | | | 491 | + | | | | + | 492 | SEN.Flow1.X +-----+ | | +--------+ REC.Flow1.X | 493 +--------------+ | | | | +--------------+ 494 + +-+---+---+ +--+--+---+ + 495 | |Router L | |Router R | | 496 | |---------| |---------| | 497 | | AQM | | | | 498 | | BuffSize| | BuffSize| | 499 | | (Bsize) +-----+ (Bsize) | | 500 | +-----+--++ ++-+------+ | 501 + | | | | + 502 +--------------+ | | | | +--------------+ 503 |traffic classN| | | | | |traffic classN| 504 |--------------| | | | | |--------------| 505 | SEN.FlowN.1 +---------+ | | +-----------+ REC.FlowN.1 | 506 | + | | | | + | 507 | | | | | | | | 508 | + | | | | + | 509 | SEN.FlowN.Y +------------+ +-------------+ REC.FlowN.Y | 510 +--------------+ +--------------+ 512 Figure 1: Topology and notations 514 Figure 1 is a generic topology where: 516 o sender with different traffic characteristics (i.e., traffic 517 profiles) can be introduced; 519 o the timing of each flow could be different (i.e., when does each 520 flow start and stop); 522 o each traffic profile can comprise various number of flows; 524 o each link is characterized by a couple (one-way delay, capacity); 525 o flows are generated at A and sent to B, sharing a bottleneck (the 526 link between routers L and R); 528 o the tester SHOULD consider both scenarios of asymmetric and 529 symmetric bottleneck links in terms of bandwidth. In case of 530 asymmetric link, the capacity from senders to receivers is higher 531 than the one from receivers to senders; the symmetric link 532 scenario provides a basic understanding of the operation of the 533 AQM mechanism whereas the asymmetric link scenario evaluates an 534 AQM mechanism in a more realistic setup; 536 o in asymmetric link scenarios, the tester SHOULD study the bi- 537 directional traffic between A and B (downlink and uplink) with the 538 AQM mechanism deployed on one direction only. The tester MAY 539 additionally consider a scenario with AQM mechanism being deployed 540 on both directions. In each scenario, the tester SHOULD 541 investigate the impact of drop policy of the AQM on TCP ACK 542 packets and its impact on the performance. 544 Although this topology may not perfectly reflect actual topologies, 545 the simple topology is commonly used in the world of simulations and 546 small testbeds. It can be considered as adequate to evaluate AQM 547 proposals, similarly to the topology proposed in 548 [I-D.irtf-iccrg-tcpeval]. Testers ought to pay attention to the 549 topology that has been used to evaluate an AQM scheme when comparing 550 this scheme with a newly proposed AQM scheme. 552 3.2. Buffer size 554 The size of the buffers should be carefully chosen, and MAY be set to 555 the bandwidth-delay product; the bandwidth being the bottleneck 556 capacity and the delay the largest RTT in the considered network. 557 The size of the buffer can impact the AQM performance and is a 558 dimensioning parameter that will be considered when comparing AQM 559 proposals. 561 If a specific buffer size is required, the tester MUST justify and 562 detail the way the maximum queue size is set. Indeed, the maximum 563 size of the buffer may affect the AQM's performance and its choice 564 SHOULD be elaborated for a fair comparison between AQM proposals. 565 While comparing AQM schemes the buffer size SHOULD remain the same 566 across the tests. 568 3.3. Congestion controls 570 This document considers running three different congestion control 571 algorithms between A and B 572 o Standard TCP congestion control: the base-line congestion control 573 is TCP NewReno with SACK, as explained in [RFC5681]. 575 o Aggressive congestion controls: a base-line congestion control for 576 this category is TCP Cubic [I-D.ietf-tcpm-cubic]. 578 o Less-than Best Effort (LBE) congestion controls: an LBE congestion 579 control 'results in smaller bandwidth and/or delay impact on 580 standard TCP than standard TCP itself, when sharing a bottleneck 581 with it.' [RFC6297] 583 Other transport congestion controls can OPTIONALLY be evaluated in 584 addition. Recent transport layer protocols are not mentioned in the 585 following sections, for the sake of simplicity. 587 4. Methodology, Metrics, AQM Comparisons, Packet Sizes, Scheduling and 588 ECN 590 4.1. Methodology 592 A description of each test setup SHOULD be detailed to allow this 593 test to be compared with other tests. This also allows others to 594 replicate the tests if needed. This test setup SHOULD detail 595 software and hardware versions. The tester could make its data 596 available. 598 The proposals SHOULD be evaluated on real-life systems, or they MAY 599 be evaluated with event-driven simulations (such as ns-2, ns-3, 600 OMNET, etc). The proposed scenarios are not bound to a particular 601 evaluation toolset. 603 The tester is encouraged to make the detailed test setup and the 604 results publicly available. 606 4.2. Comments on metrics measurement 608 The document presents the end-to-end metrics that ought to be used to 609 evaluate the trade-off between latency and goodput in Section 2. In 610 addition to the end-to-end metrics, the queue-level metrics (normally 611 collected at the device operating the AQM) provide a better 612 understanding of the AQM behavior under study and the impact of its 613 internal parameters. Whenever it is possible (e.g., depending on the 614 features provided by the hardware/software), these guidelines advice 615 to consider queue-level metrics, such as link utilization, queuing 616 delay, queue size or packet drop/mark statistics in addition to the 617 AQM-specific parameters. However, the evaluation MUST be primarily 618 based on externally observed end-to-end metrics. 620 These guidelines do not aim to detail on the way these metrics can be 621 measured, since the way these metrics are measured is expected to 622 depend on the evaluation toolset. 624 4.3. Comparing AQM schemes 626 This document recognizes that these guidelines may be used for 627 comparing AQM schemes. 629 AQM schemes need to be compared against both performance and 630 deployment categories. In addition, this section details how best to 631 achieve a fair comparison of AQM schemes by avoiding certain 632 pitfalls. 634 4.3.1. Performance comparison 636 AQM schemes should be compared against the generic scenarios that are 637 summarized in Section 13. AQM schemes MAY be compared for specific 638 network environments such as data centers, home networks, etc. If an 639 AQM scheme has parameter(s) that were externally tuned for 640 optimization or other purposes, these values MUST be disclosed. 642 AQM schemes belong to different varieties such as queue-length based 643 schemes (ex. RED) or queueing-delay based scheme (ex. CoDel, PIE). 644 AQM schemes expose different control knobs associated with different 645 semantics. For example, while both PIE and CoDel are queueing-delay 646 based schemes and each expose a knob to control the queueing delay -- 647 PIE's "queueing delay reference" vs. CoDel's "queueing delay target", 648 the two tuning parameters of the two schemes have different 649 semantics, resulting in different control points. Such differences 650 in AQM schemes can be easily overlooked while making comparisons. 652 This document RECOMMENDS the following procedures for a fair 653 performance comparison between the AQM schemes: 655 1. comparable control parameters and comparable input values: 656 carefully identify the set of parameters that control similar 657 behavior between the two AQM schemes and ensure these parameters 658 have comparable input values. For example, to compare how well a 659 queue-length based AQM scheme controls queueing delay vs. a 660 queueing-delay based AQM scheme, a tester can identify the 661 parameters of the schemes that control queue delay and ensure 662 that their input values are comparable. Similarly, to compare 663 how well two AQM schemes accommodate packet bursts, the tester 664 can identify burst-related control parameters and ensure they are 665 configured with similar values. Additionally, it would be 666 preferable if an AQM proposal listed such parameters and 667 discussed how each relates to network characteristics such as 668 capacity, average RTT etc. 670 2. compare over a range of input configurations: there could be 671 situations when the set of control parameters that affect a 672 specific behavior have different semantics between the two AQM 673 schemes. As mentioned above, PIE has tuning parameters to 674 control queue delay that has a different semantics from those 675 used in CoDel. In such situations, these schemes need to be 676 compared over a range of input configurations. For example, 677 compare PIE vs. CoDel over the range of target delay input 678 configurations. 680 4.3.2. Deployment comparison 682 AQM schemes MUST be compared against deployment criteria such as the 683 parameter sensitivity (Section 8.3), auto-tuning (Section 12) or 684 implementation cost (Section 11). 686 4.4. Packet sizes and congestion notification 688 An AQM scheme may be considering packet sizes while generating 689 congestion signals. [RFC7141] discusses the motivations behind this. 690 For example, control packets such as DNS requests/responses, TCP 691 SYNs/ACKs are small, but their loss can severely impact the 692 application performance. An AQM scheme may therefore be biased 693 towards small packets by dropping them with smaller probability 694 compared to larger packets. However, such an AQM scheme is unfair to 695 data senders generating larger packets. Data senders, malicious or 696 otherwise, are motivated to take advantage of such AQM scheme by 697 transmitting smaller packets, and could result in unsafe deployments 698 and unhealthy transport and/or application designs. 700 An AQM scheme SHOULD adhere to the recommendations outlined in 701 [RFC7141], and SHOULD NOT provide undue advantage to flows with 702 smaller packets [RFC7567]. 704 4.5. Interaction with ECN 706 Deployed AQM algorithms SHOULD implement Explicit Congestion 707 Notification (ECN) as well as loss to signal congestion to endpoints 708 [RFC7567]. ECN [RFC3168] is an alternative that allows AQM schemes 709 to signal receivers about network congestion that does not use packet 710 drop. The benefits of providing ECN support for an AQM scheme are 711 described in [WELZ2015]. Section 3 of [WELZ2015] describes expected 712 operation of routers enabling ECN. AQM schemes SHOULD NOT drop or 713 remark packets solely because the ECT(0) or ECT(1) codepoints are 714 used, and when ECN-capable SHOULD set a CE-mark on ECN-capable 715 packets in the presence of incipient congestion. 717 If the tested AQM scheme can support ECN [RFC7567], the testers MUST 718 discuss and describe the support of ECN. Since these guidelines can 719 be used to evaluate the performance of the tested AQM with and 720 without ECN markings, they could also be used to quantify the 721 interest of enabling ECN. 723 4.6. Interaction with Scheduling 725 A network device may use per-flow or per-class queuing with a 726 scheduling algorithm to either prioritize certain applications or 727 classes of traffic, limit the rate of transmission, or to provide 728 isolation between different traffic flows within a common class 729 [RFC7567]. 731 The scheduling and the AQM conjointly impact on the end-to-end 732 performance. Therefore, the AQM proposal MUST discuss the 733 feasibility to add scheduling combined with the AQM algorithm. This 734 discussion as an instance, MAY explain whether the dropping policy is 735 applied when packets are being enqueued or dequeued. 737 These guidelines do not propose guidelines to assess the performance 738 of scheduling algorithms. Indeed, as opposed to characterizing AQM 739 schemes that is related to their capacity to control the queuing 740 delay in a queue, characterizing scheduling schemes is related to the 741 scheduling itself and its interaction with the AQM scheme. As one 742 example, the scheduler may create sub-queues and the AQM scheme may 743 be applied on each of the sub-queues, and/or the AQM could be applied 744 on the whole queue. Also, schedulers might, such as FQ-CoDel 745 [HOEI2015] or FavorQueue [ANEL2014], introduce flow prioritization. 746 In these cases, specific scenarios should be proposed to ascertain 747 that these scheduler schemes not only helps in tackling the 748 bufferbloat, but also are robust under a wide variety of operating 749 conditions. This is out of the scope of this document that focus on 750 dropping and/or marking AQM schemes. 752 5. Transport Protocols 754 Network and end-devices need to be configured with a reasonable 755 amount of buffer space to absorb transient bursts. In some 756 situations, network providers tend to configure devices with large 757 buffers to avoid packet drops triggered by a full buffer and to 758 maximize the link utilization for standard loss-based TCP traffic. 760 AQM algorithms are often evaluated by considering Transmission 761 Control Protocol (TCP) [RFC0793] with a limited number of 762 applications. TCP is a widely deployed transport. It fills up 763 available buffers until a sender transfering a bulk flow with TCP 764 receives a signal (packet drop) that reduces the sending rate. The 765 larger the buffer, the higher the buffer occupancy, and therefore the 766 queuing delay. An efficient AQM scheme sends out early congestion 767 signals to TCP to bring the queuing delay under control. 769 Not all endpoints (or applications) using TCP use the same flavor of 770 TCP. Variety of senders generate different classes of traffic which 771 may not react to congestion signals (aka non-responsive flows 772 [RFC7567]) or may not reduce their sending rate as expected (aka 773 Transport Flows that are less responsive than TCP[RFC7567], also 774 called "aggressive flows"). In these cases, AQM schemes seek to 775 control the queuing delay. 777 This section provides guidelines to assess the performance of an AQM 778 proposal for various traffic profiles -- different types of senders 779 (with different TCP congestion control variants, unresponsive, 780 aggressive). 782 5.1. TCP-friendly sender 784 5.1.1. TCP-friendly sender with the same initial congestion window 786 This scenario helps to evaluate how an AQM scheme reacts to a TCP- 787 friendly transport sender. A single long-lived, non application- 788 limited, TCP NewReno flow, with an Initial congestion Window (IW) set 789 to 3 packets, transfers data between sender A and receiver B. Other 790 TCP friendly congestion control schemes such as TCP-friendly rate 791 control [RFC5348] etc MAY also be considered. 793 For each TCP-friendly transport considered, the graph described in 794 Section 2.7 could be generated. 796 5.1.2. TCP-friendly sender with different initial congestion windows 798 This scenario can be used to evaluate how an AQM scheme adapts to a 799 traffic mix consisting of TCP flows with different values of the IW. 801 For this scenario, two types of flows MUST be generated between 802 sender A and receiver B: 804 o A single long-lived non application-limited TCP NewReno flow; 806 o A single application-limited TCP NewReno flow, with an IW set to 3 807 or 10 packets. The size of the data transferred must be strictly 808 higher than 10 packets and should be lower than 100 packets. 810 The transmission of the non application-limited flow must start 811 before the transmission of the application-limited flow and only 812 after the steady state has been reached by non application-limited 813 flow. 815 For each of these scenarios, the graph described in Section 2.7 could 816 be generated for each class of traffic (application-limited and non 817 application-limited). The completion time of the application-limited 818 TCP flow could be measured. 820 5.2. Aggressive transport sender 822 This scenario helps testers to evaluate how an AQM scheme reacts to a 823 transport sender that is more aggressive than a single TCP-friendly 824 sender. We define 'aggressiveness' as a higher increase factor than 825 standard upon a successful transmission and/or a lower than standard 826 decrease factor upon a unsuccessful transmission (e.g., in case of 827 congestion controls with Additive-Increase Multiplicative-Decrease 828 (AIMD) principle, a larger AI and/or MD factors). A single long- 829 lived, non application-limited, TCP Cubic flow transfers data between 830 sender A and receiver B. Other aggressive congestion control schemes 831 MAY also be considered. 833 For each flavor of aggressive transports, the graph described in 834 Section 2.7 could be generated. 836 5.3. Unresponsive transport sender 838 This scenario helps testers to evaluate how an AQM scheme reacts to a 839 transport sender that is less responsive than TCP. Note that faulty 840 transport implementations on an end host and/or faulty network 841 elements en-route that "hide" congestion signals in packet headers 842 [RFC7567] may also lead to a similar situation, such that the AQM 843 scheme needs to adapt to unresponsive traffic. To this end, these 844 guidelines propose the two following scenarios. 846 The first scenario can be used to evaluate queue build up. It 847 considers unresponsive flow(s) whose sending rate is greater than the 848 bottleneck link capacity between routers L and R. This scenario 849 consists of a long-lived non application limited UDP flow transmits 850 data between sender A and receiver B. Graphs described in 851 Section 2.7 could be generated. 853 The second scenario can be used to evaluate if the AQM scheme is able 854 to keep the responsive fraction under control. This scenario 855 considers a mixture of TCP-friendly and unresponsive traffics. It 856 consists of a long-lived UDP flow from unresponsive application and a 857 single long-lived, non application-limited (unlimited data available 858 to the transport sender from application layer), TCP New Reno flow 859 that transmit data between sender A and receiver B. As opposed to 860 the first scenario, the rate of the UDP traffic should not be greater 861 than the bottleneck capacity, and should be higher than half of the 862 bottleneck capacity. For each type of traffic, the graph described 863 in Section 2.7 could be generated. 865 5.4. Less-than Best Effort transport sender 867 This scenario helps to evaluate how an AQM scheme reacts to LBE 868 congestion controls that 'results in smaller bandwidth and/or delay 869 impact on standard TCP than standard TCP itself, when sharing a 870 bottleneck with it.' [RFC6297]. The potential fateful interaction 871 when AQM and LBE techniques are combined has been shown in 872 [GONG2014]; this scenario helps to evaluate whether the coexistence 873 of the proposed AQM and LBE techniques may be possible. 875 A single long-lived non application-limited TCP NewReno flow 876 transfers data between sender A and receiver B. Other TCP-friendly 877 congestion control schemes MAY also be considered. Single long-lived 878 non application-limited LEDBAT [RFC6817] flows transfer data between 879 sender A and receiver B. We recommend to set the target delay and 880 gain values of LEDBAT respectively to 5 ms and 10 [TRAN2014]. Other 881 LBE congestion control schemes, any of those listed in [RFC6297], MAY 882 also be considered. 884 For each of the TCP-friendly and LBE transports, the graph described 885 in Section 2.7 could be generated. 887 6. Round Trip Time Fairness 889 6.1. Motivation 891 An AQM scheme's congestion signals (via drops or ECN marks) must 892 reach the transport sender so that a responsive sender can initiate 893 its congestion control mechanism and adjust the sending rate. This 894 procedure is thus dependent on the end-to-end path RTT. When the RTT 895 varies, the onset of congestion control is impacted, and in turn 896 impacts the ability of an AQM scheme to control the queue. It is 897 therefore important to assess the AQM schemes for a set of RTTs 898 between A and B (e.g., from 5 ms to 200 ms). 900 The asymmetry in terms of difference in intrinsic RTT between various 901 paths sharing the same bottleneck SHOULD be considered so that the 902 fairness between the flows can be discussed since in this scenario, a 903 flow traversing on shorter RTT path may react faster to congestion 904 and recover faster from it compared to another flow on a longer RTT 905 path. The introduction of AQM schemes may potentially improve this 906 type of fairness. 908 Introducing an AQM scheme may cause the unfairness between the flows, 909 even if the RTTs are identical. This potential unfairness SHOULD be 910 investigated as well. 912 6.2. Recommended tests 914 The RECOMMENDED topology is detailed in Figure 1. 916 To evaluate the RTT fairness, for each run, two flows divided into 917 two categories. Category I whose RTT between sender A and receiver B 918 SHOULD be 100ms. Category II which RTT between sender A and receiver 919 B should be in the range [5ms;560ms] inclusive. The maximum value 920 for the RTT represents the RTT of a satellite link that, according to 921 section 2 of [RFC2488] should be at least 558ms. 923 A set of evaluated flows MUST use the same congestion control 924 algorithm: all the generated flows could be single long-lived non 925 application-limited TCP NewReno flows. 927 6.3. Metrics to evaluate the RTT fairness 929 The outputs that MUST be measured are: (1) the cumulative average 930 goodput of the flow from Category I, goodput_Cat_I (Section 2.5); (2) 931 the cumulative average goodput of the flow from Category II, 932 goodput_Cat_II (Section 2.5); (3) the ratio goodput_Cat_II/ 933 goodput_Cat_I; (4) the average packet drop rate for each category 934 (Section 2.3). 936 7. Burst Absorption 938 "AQM mechanisms need to control the overall queue sizes, to ensure 939 that arriving bursts can be accommodated without dropping packets" 940 [RFC7567]. 942 7.1. Motivation 944 An AQM scheme can face bursts of packet arrivals due to various 945 reasons. Dropping one or more packets from a burst can result in 946 performance penalties for the corresponding flows, since dropped 947 packets have to be retransmitted. Performance penalties can result 948 in failing to meet SLAs and be a disincentive to AQM adoption. 950 The ability to accommodate bursts translates to larger queue length 951 and hence more queuing delay. On the one hand, it is important that 952 an AQM scheme quickly brings bursty traffic under control. On the 953 other hand, a peak in the packet drop rates to bring a packet burst 954 quickly under control could result in multiple drops per flow and 955 severely impact transport and application performance. Therefore, an 956 AQM scheme ought to bring bursts under control by balancing both 957 aspects -- (1) queuing delay spikes are minimized and (2) performance 958 penalties for ongoing flows in terms of packet drops are minimized. 960 An AQM scheme that maintains short queues allows some remaining space 961 in the buffer for bursts of arriving packets. The tolerance to 962 bursts of packets depends upon the number of packets in the queue, 963 which is directly linked to the AQM algorithm. Moreover, an AQM 964 scheme may implement a feature controlling the maximum size of 965 accepted bursts, that can depend on the buffer occupancy or the 966 currently estimated queuing delay. The impact of the buffer size on 967 the burst allowance may be evaluated. 969 7.2. Recommended tests 971 For this scenario, tester MUST evaluate how the AQM performs with the 972 following traffic generated from sender A to receiver B: 974 o Web traffic with IW10; 976 o Bursty video frames; 978 o Constant Bit Rate (CBR) UDP traffic. 980 o A single non application-limited bulk TCP flow as background 981 traffic. 983 Figure 2 presents the various cases for the traffic that MUST be 984 generated between sender A and receiver B. 986 +-------------------------------------------------+ 987 |Case| Traffic Type | 988 | +-----+------------+----+--------------------+ 989 | |Video|Web (IW 10)| CBR| Bulk TCP Traffic | 990 +----|-----|------------|----|--------------------| 991 |I | 0 | 1 | 1 | 0 | 992 +----|-----|------------|----|--------------------| 993 |II | 0 | 1 | 1 | 1 | 994 |----|-----|------------|----|--------------------| 995 |III | 1 | 1 | 1 | 0 | 996 +----|-----|------------|----|--------------------| 997 |IV | 1 | 1 | 1 | 1 | 998 +----+-----+------------+----+--------------------+ 1000 Figure 2: Bursty traffic scenarios 1002 A new web page download could start after the previous web page 1003 download is finished. Each web page could be composed by at least 50 1004 objects and the size of each object should be at least 1kB. 6 TCP 1005 parallel connections SHOULD be generated to download the objects, 1006 each parallel connections having an initial congestion window set to 1007 10 packets. 1009 For each of these scenarios, the graph described in Section 2.7 could 1010 be generated for each application. Metrics such as end-to-end 1011 latency, jitter, flow completion time MAY be generated. For the 1012 cases of frame generation of bursty video traffic as well as the 1013 choice of web traffic pattern, these details and their presentation 1014 are left to the testers. 1016 8. Stability 1018 8.1. Motivation 1020 The safety of an AQM scheme is directly related to its stability 1021 under varying operating conditions such as varying traffic profiles 1022 and fluctuating network conditions. Since operating conditions can 1023 vary often the AQM needs to remain stable under these conditions 1024 without the need for additional external tuning. 1026 Network devices can experience varying operating conditions depending 1027 on factors such as time of the day, deployment scenario, etc. For 1028 example: 1030 o Traffic and congestion levels are higher during peak hours than 1031 off-peak hours. 1033 o In the presence of a scheduler, the draining rate of a queue can 1034 vary depending on the occupancy of other queues: a low load on a 1035 high priority queue implies a higher draining rate for the lower 1036 priority queues. 1038 o The capacity available can vary over time (e.g., a lossy channel, 1039 a link supporting traffic in a higher diffserv class). 1041 Whether the target context is a not stable environment, the ability 1042 of an AQM scheme to maintain its control over the queuing delay and 1043 buffer occupancy can be challenged. This document proposes 1044 guidelines to assess the behavior of AQM schemes under varying 1045 congestion levels and varying draining rates. 1047 8.2. Recommended tests 1049 Note that the traffic profiles explained below comprises non 1050 application-limited TCP flows. For each of the below scenarios, the 1051 graphs described in Section 2.7 SHOULD be generated, and the goodput 1052 of the various flows should be cumulated. For Section 8.2.5 and 1053 Section 8.2.6 they SHOULD incorporate the results in per-phase basis 1054 as well. 1056 Wherever the notion of time has explicitly mentioned in this 1057 subsection, time 0 starts from the moment all TCP flows have already 1058 reached their congestion avoidance phase. 1060 8.2.1. Definition of the congestion Level 1062 In these guidelines, the congestion levels are represented by the 1063 projected packet drop rate, had a drop-tail queue was chosen instead 1064 of an AQM scheme. When the bottleneck is shared among non 1065 application-limited TCP flows. l_r, the loss rate projection can be 1066 expressed as a function of N, the number of bulk TCP flows, and S, 1067 the sum of the bandwidth-delay product and the maximum buffer size, 1068 both expressed in packets, based on Eq. 3 of [MORR2000]: 1070 l_r = 0.76 * N^2 / S^2 1072 N = S * SQRT(1/0.76) * SQRT (l_r) 1074 These guidelines use the loss rate to define the different congestion 1075 levels, but they do not stipulate that in other circumstances, 1076 measuring the congestion level gives you an accurate estimation of 1077 the loss rate or vice-versa. 1079 8.2.2. Mild congestion 1081 This scenario can be used to evaluate how an AQM scheme reacts to a 1082 light load of incoming traffic resulting in mild congestion -- packet 1083 drop rates around 0.1%. The number of bulk flows required to achieve 1084 this congestion level, N_mild, is then: 1086 N_mild = ROUND (0.036*S) 1088 8.2.3. Medium congestion 1090 This scenario can be used to evaluate how an AQM scheme reacts to 1091 incoming traffic resulting in medium congestion -- packet drop rates 1092 around 0.5%. The number of bulk flows required to achieve this 1093 congestion level, N_med, is then: 1095 N_med = ROUND (0.081*S) 1097 8.2.4. Heavy congestion 1099 This scenario can be used to evaluate how an AQM scheme reacts to 1100 incoming traffic resulting in heavy congestion -- packet drop rates 1101 around 1%. The number of bulk flows required to achieve this 1102 congestion level, N_heavy, is then: 1104 N_heavy = ROUND (0.114*S) 1106 8.2.5. Varying the congestion level 1108 This scenario can be used to evaluate how an AQM scheme reacts to 1109 incoming traffic resulting in various levels of congestion during the 1110 experiment. In this scenario, the congestion level varies within a 1111 large time-scale. The following phases may be considered: phase I - 1112 mild congestion during 0-20s; phase II - medium congestion during 1113 20-40s; phase III - heavy congestion during 40-60s; phase I again, 1114 and so on. 1116 8.2.6. Varying available capacity 1118 This scenario can be used to help characterize how the AQM behaves 1119 and adapts to bandwidth changes. The experiments are not meant to 1120 reflect the exact conditions of Wi-Fi environments since it is hard 1121 to design repetitive experiments or accurate simulations for such 1122 scenarios. 1124 To emulate varying draining rates, the bottleneck capacity between 1125 nodes 'Router L' and 'Router R' varies over the course of the 1126 experiment as follows: 1128 o Experiment 1: the capacity varies between two values within a 1129 large time-scale. As an example, the following phases may be 1130 considered: phase I - 100Mbps during 0-20s; phase II - 10Mbps 1131 during 20-40s; phase I again, and so on. 1133 o Experiment 2: the capacity varies between two values within a 1134 short time-scale. As an example, the following phases may be 1135 considered: phase I - 100Mbps during 0-100ms; phase II - 10Mbps 1136 during 100-200ms; phase I again, and so on. 1138 The tester MAY choose a phase time-interval value different than what 1139 is stated above, if the network's path conditions (such as bandwidth- 1140 delay product) necessitate. In this case the choice of such time- 1141 interval value SHOULD be stated and elaborated. 1143 The tester MAY additionally evaluate the two mentioned scenarios 1144 (short-term and long-term capacity variations), during and/or 1145 including TCP slow-start phase. 1147 More realistic fluctuating capacity patterns MAY be considered. The 1148 tester MAY choose to incorporate realistic scenarios with regards to 1149 common fluctuation of bandwidth in state-of-the-art technologies. 1151 The scenario consists of TCP NewReno flows between sender A and 1152 receiver B. To better assess the impact of draining rates on the AQM 1153 behavior, the tester MUST compare its performance with those of drop- 1154 tail and SHOULD provide a reference document for their proposal 1155 discussing performance and deployment compared to those of drop-tail. 1156 Burst traffic, such as presented in Section 7.2, could also be 1157 considered to assess the impact of varying available capacity on the 1158 burst absorption of the AQM. 1160 8.3. Parameter sensitivity and stability analysis 1162 The control law used by an AQM is the primary means by which the 1163 queuing delay is controlled. Hence understanding the control law is 1164 critical to understanding the behavior of the AQM scheme. The 1165 control law could include several input parameters whose values 1166 affect the AQM scheme's output behavior and its stability. 1167 Additionally, AQM schemes may auto-tune parameter values in order to 1168 maintain stability under different network conditions (such as 1169 different congestion levels, draining rates or network environments). 1170 The stability of these auto-tuning techniques is also important to 1171 understand. 1173 Transports operating under the control of AQM experience the effect 1174 of multiple control loops that react over different timescales. It 1175 is therefore important that proposed AQM schemes are seen to be 1176 stable when they are deployed at multiple points of potential 1177 congestion along an Internet path. The pattern of congestion signals 1178 (loss or ECN-marking) arising from AQM methods also need to not 1179 adversely interact with the dynamics of the transport protocols that 1180 they control. 1182 AQM proposals SHOULD provide background material showing control 1183 theoretic analysis of the AQM control law and the input parameter 1184 space within which the control law operates as expected; or could use 1185 another way to discuss the stability of the control law. For 1186 parameters that are auto-tuned, the material SHOULD include stability 1187 analysis of the auto-tuning mechanism(s) as well. Such analysis 1188 helps to understand an AQM control law better and the network 1189 conditions/deployments under which the AQM is stable. 1191 9. Various Traffic Profiles 1193 This section provides guidelines to assess the performance of an AQM 1194 proposal for various traffic profiles such as traffic with different 1195 applications or bi-directional traffic. 1197 9.1. Traffic mix 1199 This scenario can be used to evaluate how an AQM scheme reacts to a 1200 traffic mix consisting of different applications such as: 1202 o Bulk TCP transfer 1204 o Web traffic 1206 o VoIP 1208 o Constant Bit Rate (CBR) UDP traffic 1210 o Adaptive video streaming 1212 Various traffic mixes can be considered. These guidelines RECOMMEND 1213 to examine at least the following example: 1 bi-directional VoIP; 6 1214 Web pages download (such as detailed in Section 7.2); 1 CBR; 1 1215 Adaptive Video; 5 bulk TCP. Any other combinations could be 1216 considered and should be carefully documented. 1218 For each scenario, the graph described in Section 2.7 could be 1219 generated for each class of traffic. Metrics such as end-to-end 1220 latency, jitter and flow completion time MAY be reported. 1222 9.2. Bi-directional traffic 1224 Control packets such as DNS requests/responses, TCP SYNs/ACKs are 1225 small, but their loss can severely impact the application 1226 performance. The scenario proposed in this section will help in 1227 assessing whether the introduction of an AQM scheme increases the 1228 loss probability of these important packets. 1230 For this scenario, traffic MUST be generated in both downlink and 1231 uplink, such as defined in Section 3.1. These guidelines RECOMMEND 1232 to consider a mild congestion level and the traffic presented in 1233 Section 8.2.2 in both directions. In this case, the metrics reported 1234 MUST be the same as in Section 8.2 for each direction. 1236 The traffic mix presented in Section 9.1 MAY also be generated in 1237 both directions. 1239 10. Multi-AQM Scenario 1241 10.1. Motivation 1243 Transports operating under the control of AQM experience the effect 1244 of multiple control loops that react over different timescales. It 1245 is therefore important that proposed AQM schemes are seen to be 1246 stable when they are deployed at multiple points of potential 1247 congestion along an Internet path. The pattern of congestion signals 1248 (loss or ECN-marking) arising from AQM methods also need to not 1249 adversely interact with the dynamics of the transport protocols that 1250 they control. 1252 10.2. Details on the evaluation scenario 1254 +---------+ +-----------+ 1255 |senders A|---+ +---|receivers A| 1256 +---------+ | | +-----------+ 1257 +-----+---+ +---------+ +--+-----+ 1258 |Router L |--|Router M |--|Router R| 1259 |AQM | |AQM | |No AQM | 1260 +---------+ +--+------+ +--+-----+ 1261 +---------+ | | +-----------+ 1262 |senders B|-------------+ +---|receivers B| 1263 +---------+ +-----------+ 1265 Figure 3: Topology for the Multi-AQM scenario 1267 This scenario can be used to evaluate how having AQM schemes in 1268 sequence impact the induced latency reduction, the induced goodput 1269 maximization and the trade-off between these two. The topology 1270 presented in Figure 3 could be used. AQM schemes introduced in 1271 Router L and Router M should be the same; any other configurations 1272 could be considered. For this scenario, it is recommended to 1273 consider a mild congestion level, the number of flows specified in 1274 Section 8.2.2 being equally shared among senders A and B. Any other 1275 relevant combination of congestion levels could be considered. We 1276 recommend to measure the metrics presented in Section 8.2. 1278 11. Implementation cost 1280 11.1. Motivation 1282 Successful deployment of AQM is directly related to its cost of 1283 implementation. Network devices can need hardware or software 1284 implementations of the AQM mechanism. Depending on a device's 1285 capabilities and limitations, the device may or may not be able to 1286 implement some or all parts of their AQM logic. 1288 AQM proposals SHOULD provide pseudo-code for the complete AQM scheme, 1289 highlighting generic implementation-specific aspects of the scheme 1290 such as "drop-tail" vs. "drop-head", inputs (e.g., current queuing 1291 delay, queue length), computations involved, need for timers, etc. 1292 This helps to identify costs associated with implementing the AQM 1293 scheme on a particular hardware or software device. This also 1294 facilitates discsusions around which kind of devices can easily 1295 support the AQM and which cannot. 1297 11.2. Recommended discussion 1299 AQM proposals SHOULD highlight parts of their AQM logic that are 1300 device dependent and discuss if and how AQM behavior could be 1301 impacted by the device. For example, a queueing-delay based AQM 1302 scheme requires current queuing delay as input from the device. If 1303 the device already maintains this value, then it can be trivial to 1304 implement the their AQM logic on the device. If the device provides 1305 indirect means to estimate the queuing delay (for example: 1306 timestamps, dequeuing rate), then the AQM behavior is sensitive to 1307 the precision of the queuing delay estimations are for that device. 1308 Highlighting the sensitivity of an AQM scheme to queuing delay 1309 estimations helps implementers to identify appropriate means of 1310 implementing the mechanism on a device. 1312 12. Operator Control and Auto-tuning 1314 12.1. Motivation 1316 One of the biggest hurdles of RED deployment was/is its parameter 1317 sensitivity to operating conditions -- how difficult it is to tune 1318 RED parameters for a deployment to achieve acceptable benefit from 1319 using RED. Fluctuating congestion levels and network conditions add 1320 to the complexity. Incorrect parameter values lead to poor 1321 performance. 1323 Any AQM scheme is likely to have parameters whose values affect the 1324 control law and behaviour of an AQM. Exposing all these parameters 1325 as control parameters to a network operator (or user) can easily 1326 result in a unsafe AQM deployment. Unexpected AQM behavior ensues 1327 when parameter values are set improperly. A minimal number of 1328 control parameters minimizes the number of ways a user can break a 1329 system where an AQM scheme is deployed at. Fewer control parameters 1330 make the AQM scheme more user-friendly and easier to deploy and 1331 debug. 1333 [RFC7567] states "AQM algorithms SHOULD NOT require tuning of initial 1334 or configuration parameters in common use cases." A scheme ought to 1335 expose only those parameters that control the macroscopic AQM 1336 behavior such as queue delay threshold, queue length threshold, etc. 1338 Additionally, the safety of an AQM scheme is directly related to its 1339 stability under varying operating conditions such as varying traffic 1340 profiles and fluctuating network conditions, as described in 1341 Section 8. Operating conditions vary often and hence the AQM needs 1342 to remain stable under these conditions without the need for 1343 additional external tuning. If AQM parameters require tuning under 1344 these conditions, then the AQM must self-adapt necessary parameter 1345 values by employing auto-tuning techniques. 1347 12.2. Recommended discussion 1349 In order to understand an AQM's deployment considerations and 1350 performance under a specific environment, AQM proposals SHOULD 1351 describe the parameters that control the macroscopic AQM behavior, 1352 and identify any parameters that require tuning to operational 1353 conditions. It could be interesting to also discuss that even if an 1354 AQM scheme may not adequately auto-tune its parameters, the resulting 1355 performance may not be optimal, but close to something reasonable. 1357 If there are any fixed parameters within the AQM, their setting 1358 SHOULD be discussed and justified, to help understand whether a fixed 1359 parameter value is applicable for a particular environment. 1361 If an AQM scheme is evaluated with parameter(s) that were externally 1362 tuned for optimization or other purposes, these values MUST be 1363 disclosed. 1365 13. Conclusion 1367 Figure 4 lists the scenarios and their requirements. 1369 +------------------------------------------------------------------+ 1370 |Scenario |Sec. |Requirement | 1371 +------------------------------------------------------------------+ 1372 +------------------------------------------------------------------+ 1373 |Interaction with ECN | 4.5 |MUST be discussed if supported | 1374 +------------------------------------------------------------------+ 1375 |Interaction with Scheduling| 4.6 |Feasibility MUST be discussed | 1376 +------------------------------------------------------------------+ 1377 |Transport Protocols |5. | | 1378 | TCP-friendly sender | 5.1 |Scenario MUST be considered | 1379 | Aggressive sender | 5.2 |Scenario MUST be considered | 1380 | Unresponsive sender | 5.3 |Scenario MUST be considered | 1381 | LBE sender | 5.4 |Scenario MAY be considered | 1382 +------------------------------------------------------------------+ 1383 |Round Trip Time Fairness | 6.2 |Scenario MUST be considered | 1384 +------------------------------------------------------------------+ 1385 |Burst Absorption | 7.2 |Scenario MUST be considered | 1386 +------------------------------------------------------------------+ 1387 |Stability |8. | | 1388 | Varying congestion levels | 8.2.5|Scenario MUST be considered | 1389 | Varying available capacity| 8.2.6|Scenario MUST be considered | 1390 | Parameters and stability | 8.3 |This SHOULD be discussed | 1391 +------------------------------------------------------------------+ 1392 |Various Traffic Profiles |9. | | 1393 | Traffic mix | 9.1 |Scenario is RECOMMENDED | 1394 | Bi-directional traffic | 9.2 |Scenario MAY be considered | 1395 +------------------------------------------------------------------+ 1396 |Multi-AQM | 10.2 |Scenario MAY be considered | 1397 +------------------------------------------------------------------+ 1398 |Implementation Cost | 11.2 |Pseudo-code SHOULD be provided | 1399 +------------------------------------------------------------------+ 1400 |Operator Control | 12.2 |Tuning SHOULD NOT be required | 1401 +------------------------------------------------------------------+ 1403 Figure 4: Summary of the scenarios and their requirements 1405 14. Acknowledgements 1407 This work has been partially supported by the European Community 1408 under its Seventh Framework Programme through the Reducing Internet 1409 Transport Latency (RITE) project (ICT-317700). 1411 15. Contributors 1413 Many thanks to S. Akhtar, A.B. Bagayoko, F. Baker, R. Bless, D. 1414 Collier-Brown, G. Fairhurst, J. Gettys, T. Hoiland-Jorgensen, K. 1415 Kilkki, C. Kulatunga, W. Lautenschlager, A.C. Morton, R. Pan, G. 1417 Skinner, D. Taht and M. Welzl for detailed and wise feedback on 1418 this document. 1420 16. IANA Considerations 1422 This memo includes no request to IANA. 1424 17. Security Considerations 1426 Some security considerations for AQM are identified in [RFC7567].This 1427 document, by itself, presents no new privacy nor security issues. 1429 18. References 1431 18.1. Normative References 1433 [I-D.ietf-tcpm-cubic] 1434 Rhee, I., Xu, L., Ha, S., Zimmermann, A., Eggert, L., and 1435 R. Scheffenegger, "CUBIC for Fast Long-Distance Networks", 1436 draft-ietf-tcpm-cubic-01 (work in progress), January 2016. 1438 [I-D.irtf-iccrg-tcpeval] 1439 Hayes, D., Ros, D., Andrew, L., and S. Floyd, "Common TCP 1440 Evaluation Suite", draft-irtf-iccrg-tcpeval-01 (work in 1441 progress), July 2014. 1443 [RFC0793] Postel, J., "Transmission Control Protocol", STD 7, 1444 RFC 793, DOI 10.17487/RFC0793, September 1981, 1445 . 1447 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 1448 Requirement Levels", RFC 2119, 1997. 1450 [RFC2488] Allman, M., Glover, D., and L. Sanchez, "Enhancing TCP 1451 Over Satellite Channels using Standard Mechanisms", 1452 BCP 28, RFC 2488, DOI 10.17487/RFC2488, January 1999, 1453 . 1455 [RFC2544] Bradner, S. and J. McQuaid, "Benchmarking Methodology for 1456 Network Interconnect Devices", RFC 2544, 1457 DOI 10.17487/RFC2544, March 1999, 1458 . 1460 [RFC2647] Newman, D., "Benchmarking Terminology for Firewall 1461 Performance", RFC 2647, DOI 10.17487/RFC2647, August 1999, 1462 . 1464 [RFC2679] Almes, G., Kalidindi, S., and M. Zekauskas, "A One-way 1465 Delay Metric for IPPM", RFC 2679, DOI 10.17487/RFC2679, 1466 September 1999, . 1468 [RFC2680] Almes, G., Kalidindi, S., and M. Zekauskas, "A One-way 1469 Packet Loss Metric for IPPM", RFC 2680, 1470 DOI 10.17487/RFC2680, September 1999, 1471 . 1473 [RFC3168] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition 1474 of Explicit Congestion Notification (ECN) to IP", 1475 RFC 3168, DOI 10.17487/RFC3168, September 2001, 1476 . 1478 [RFC3611] Friedman, T., Ed., Caceres, R., Ed., and A. Clark, Ed., 1479 "RTP Control Protocol Extended Reports (RTCP XR)", 1480 RFC 3611, DOI 10.17487/RFC3611, November 2003, 1481 . 1483 [RFC5348] Floyd, S., Handley, M., Padhye, J., and J. Widmer, "TCP 1484 Friendly Rate Control (TFRC): Protocol Specification", 1485 RFC 5348, DOI 10.17487/RFC5348, September 2008, 1486 . 1488 [RFC5481] Morton, A. and B. Claise, "Packet Delay Variation 1489 Applicability Statement", RFC 5481, DOI 10.17487/RFC5481, 1490 March 2009, . 1492 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion 1493 Control", RFC 5681, DOI 10.17487/RFC5681, September 2009, 1494 . 1496 [RFC6297] Welzl, M. and D. Ros, "A Survey of Lower-than-Best-Effort 1497 Transport Protocols", RFC 6297, DOI 10.17487/RFC6297, June 1498 2011, . 1500 [RFC6817] Shalunov, S., Hazel, G., Iyengar, J., and M. Kuehlewind, 1501 "Low Extra Delay Background Transport (LEDBAT)", RFC 6817, 1502 DOI 10.17487/RFC6817, December 2012, 1503 . 1505 [RFC7141] Briscoe, B. and J. Manner, "Byte and Packet Congestion 1506 Notification", RFC 7141, 2014. 1508 [RFC7567] Baker, F., Ed. and G. Fairhurst, Ed., "IETF 1509 Recommendations Regarding Active Queue Management", 1510 BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015, 1511 . 1513 18.2. Informative References 1515 [ANEL2014] 1516 Anelli, P., Diana, R., and E. Lochin, "FavorQueue: a 1517 Parameterless Active Queue Management to Improve TCP 1518 Traffic Performance", Computer Networks vol. 60, 2014. 1520 [BB2011] "BufferBloat: what's wrong with the internet?", ACM 1521 Queue vol. 9, 2011. 1523 [GONG2014] 1524 Gong, Y., Rossi, D., Testa, C., Valenti, S., and D. Taht, 1525 "Fighting the bufferbloat: on the coexistence of AQM and 1526 low priority congestion control", Computer Networks, 1527 Elsevier, 2014, 60, pp.115 - 128 , 2014. 1529 [HASS2008] 1530 Hassayoun, S. and D. Ros, "Loss Synchronization and Router 1531 Buffer Sizing with High-Speed Versions of TCP", IEEE 1532 INFOCOM Workshops , 2008. 1534 [HOEI2015] 1535 Hoeiland-Joergensen, T., McKenney, P., Taht, D., Gettys, 1536 J., and E. Dumazet, "FlowQueue-Codel", IETF (Work-in- 1537 Progress) , January 2015. 1539 [JAY2006] Jay, P., Fu, Q., and G. Armitage, "A preliminary analysis 1540 of loss synchronisation between concurrent TCP flows", 1541 Australian Telecommunication Networks and Application 1542 Conference (ATNAC) , 2006. 1544 [MORR2000] 1545 Morris, R., "Scalable TCP congestion control", IEEE 1546 INFOCOM , 2000. 1548 [NICH2012] 1549 Nichols, K. and V. Jacobson, "Controlling Queue Delay", 1550 ACM Queue , 2012. 1552 [PAN2013] Pan, R., Natarajan, P., Piglione, C., Prabhu, MS., 1553 Subramanian, V., Baker, F., and B. VerSteeg, "PIE: A 1554 lightweight control scheme to address the bufferbloat 1555 problem", IEEE HPSR , 2013. 1557 [RFC2309] Braden, B., Clark, D., Crowcroft, J., Davie, B., Deering, 1558 S., Estrin, D., Floyd, S., Jacobson, V., Minshall, G., 1559 Partridge, C., Peterson, L., Ramakrishnan, K., Shenker, 1560 S., Wroclawski, J., and L. Zhang, "Recommendations on 1561 Queue Management and Congestion Avoidance in the 1562 Internet", RFC 2309, April 1998. 1564 [TRAN2014] 1565 Trang, S., Kuhn, N., Lochin, E., Baudoin, C., Dubois, E., 1566 and P. Gelard, "On The Existence Of Optimal LEDBAT 1567 Parameters", IEEE ICC 2014 - Communication QoS, 1568 Reliability and Modeling Symposium , 2014. 1570 [WELZ2015] 1571 Welzl, M. and G. Fairhurst, "The Benefits to Applications 1572 of using Explicit Congestion Notification (ECN)", IETF 1573 (Work-in-Progress) , June 2015. 1575 [WINS2014] 1576 Winstein, K., "Transport Architectures for an Evolving 1577 Internet", PhD thesis, Massachusetts Institute of 1578 Technology , 2014. 1580 Authors' Addresses 1582 Nicolas Kuhn (editor) 1583 CNES, Telecom Bretagne 1584 18 avenue Edouard Belin 1585 Toulouse 31400 1586 France 1588 Phone: +33 5 61 27 32 13 1589 Email: nicolas.kuhn@cnes.fr 1591 Preethi Natarajan (editor) 1592 Cisco Systems 1593 510 McCarthy Blvd 1594 Milpitas, California 1595 United States 1597 Email: prenatar@cisco.com 1598 Naeem Khademi (editor) 1599 University of Oslo 1600 Department of Informatics, PO Box 1080 Blindern 1601 N-0316 Oslo 1602 Norway 1604 Phone: +47 2285 24 93 1605 Email: naeemk@ifi.uio.no 1607 David Ros 1608 Simula Research Laboratory AS 1609 P.O. Box 134 1610 Lysaker, 1325 1611 Norway 1613 Phone: +33 299 25 21 21 1614 Email: dros@simula.no