idnits 2.17.1 draft-ietf-ippm-model-based-metrics-09.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- No issues found here. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year == Line 1514 has weird spacing: '... and n = h1...' -- The document date (February 27, 2017) is 2615 days in the past. Is this intentional? Checking references for intended status: Experimental ---------------------------------------------------------------------------- -- Obsolete informational reference (is this intentional?): RFC 2861 (Obsoleted by RFC 7661) Summary: 0 errors (**), 0 flaws (~~), 2 warnings (==), 2 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 IP Performance Working Group M. Mathis 3 Internet-Draft Google, Inc 4 Intended status: Experimental A. Morton 5 Expires: August 31, 2017 AT&T Labs 6 February 27, 2017 8 Model Based Metrics for Bulk Transport Capacity 9 draft-ietf-ippm-model-based-metrics-09.txt 11 Abstract 13 We introduce a new class of Model Based Metrics designed to assess if 14 a complete Internet path can be expected to meet a predefined Target 15 Transport Performance by applying a suite of IP diagnostic tests to 16 successive subpaths. The subpath-at-a-time tests can be robustly 17 applied to critical infrastructure, such as network interconnections 18 or even individual devices, to accurately detect if any part of the 19 infrastructure will prevent paths traversing it from meeting the 20 Target Transport Performance. 22 Model Based Metrics rely on peer-reviewed mathematical models to 23 specify a Targeted Suite of IP Diagnostic tests, designed to assess 24 whether common transport protocols can be expected to meet a 25 predetermined Target Transport Performance over an Internet path. 27 For Bulk Transport Capacity, the IP diagnostics are built on test 28 streams that mimic TCP over the complete path and statistical 29 criteria for evaluating the packet transfer statistics of those 30 streams. The temporal structure of the test stream (bursts, etc) 31 mimic TCP or other transport protocol carrying bulk data over a long 32 path. However they are constructed to be independent of the details 33 of the subpath under test, end systems or applications. Likewise the 34 success criteria evaluates the packet transfer statistics of the 35 subpath against criteria determined by protocol performance models 36 applied to the Target Transport Performance of the complete path. 37 The success criteria also does not depend on the details of the 38 subpath, end systems or application. 40 Model Based Metrics exhibit several important new properties not 41 present in other Bulk Transport Capacity Metrics, including the 42 ability to reason about concatenated or overlapping subpaths. The 43 results are vantage independent which is critical for supporting 44 independent validation of tests by comparing results from multiple 45 measurement points. 47 This document does not define the IP diagnostic tests, but provides a 48 framework for designing suites of IP diagnostic tests that are 49 tailored to confirming that infrastructure can meet the predetermined 50 Target Transport Performance. 52 Status of This Memo 54 This Internet-Draft is submitted in full conformance with the 55 provisions of BCP 78 and BCP 79. 57 Internet-Drafts are working documents of the Internet Engineering 58 Task Force (IETF). Note that other groups may also distribute 59 working documents as Internet-Drafts. The list of current Internet- 60 Drafts is at http://datatracker.ietf.org/drafts/current/. 62 Internet-Drafts are draft documents valid for a maximum of six months 63 and may be updated, replaced, or obsoleted by other documents at any 64 time. It is inappropriate to use Internet-Drafts as reference 65 material or to cite them other than as "work in progress." 67 This Internet-Draft will expire on August 31, 2017. 69 Copyright Notice 71 Copyright (c) 2017 IETF Trust and the persons identified as the 72 document authors. All rights reserved. 74 This document is subject to BCP 78 and the IETF Trust's Legal 75 Provisions Relating to IETF Documents 76 (http://trustee.ietf.org/license-info) in effect on the date of 77 publication of this document. Please review these documents 78 carefully, as they describe your rights and restrictions with respect 79 to this document. Code Components extracted from this document must 80 include Simplified BSD License text as described in Section 4.e of 81 the Trust Legal Provisions and are provided without warranty as 82 described in the Simplified BSD License. 84 Table of Contents 86 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3 87 1.1. Version Control . . . . . . . . . . . . . . . . . . . . . 5 88 2. Overview . . . . . . . . . . . . . . . . . . . . . . . . . . 7 89 3. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 10 90 4. Background . . . . . . . . . . . . . . . . . . . . . . . . . 16 91 4.1. TCP properties . . . . . . . . . . . . . . . . . . . . . 18 92 4.2. Diagnostic Approach . . . . . . . . . . . . . . . . . . . 19 93 4.3. New requirements relative to RFC 2330 . . . . . . . . . . 20 94 5. Common Models and Parameters . . . . . . . . . . . . . . . . 21 95 5.1. Target End-to-end parameters . . . . . . . . . . . . . . 21 96 5.2. Common Model Calculations . . . . . . . . . . . . . . . . 22 97 5.3. Parameter Derating . . . . . . . . . . . . . . . . . . . 23 98 5.4. Test Preconditions . . . . . . . . . . . . . . . . . . . 23 99 6. Generating test streams . . . . . . . . . . . . . . . . . . . 24 100 6.1. Mimicking slowstart . . . . . . . . . . . . . . . . . . . 25 101 6.2. Constant window pseudo CBR . . . . . . . . . . . . . . . 27 102 6.3. Scanned window pseudo CBR . . . . . . . . . . . . . . . . 27 103 6.4. Concurrent or channelized testing . . . . . . . . . . . . 28 104 7. Interpreting the Results . . . . . . . . . . . . . . . . . . 29 105 7.1. Test outcomes . . . . . . . . . . . . . . . . . . . . . . 29 106 7.2. Statistical criteria for estimating run_length . . . . . 30 107 7.3. Reordering Tolerance . . . . . . . . . . . . . . . . . . 32 108 8. IP Diagnostic Tests . . . . . . . . . . . . . . . . . . . . . 33 109 8.1. Basic Data Rate and Packet Transfer Tests . . . . . . . . 33 110 8.1.1. Delivery Statistics at Paced Full Data Rate . . . . . 34 111 8.1.2. Delivery Statistics at Full Data Windowed Rate . . . 34 112 8.1.3. Background Packet Transfer Statistics Tests . . . . . 35 113 8.2. Standing Queue Tests . . . . . . . . . . . . . . . . . . 35 114 8.2.1. Congestion Avoidance . . . . . . . . . . . . . . . . 36 115 8.2.2. Bufferbloat . . . . . . . . . . . . . . . . . . . . . 37 116 8.2.3. Non excessive loss . . . . . . . . . . . . . . . . . 37 117 8.2.4. Duplex Self Interference . . . . . . . . . . . . . . 38 118 8.3. Slowstart tests . . . . . . . . . . . . . . . . . . . . . 38 119 8.3.1. Full Window slowstart test . . . . . . . . . . . . . 38 120 8.3.2. Slowstart AQM test . . . . . . . . . . . . . . . . . 39 121 8.4. Sender Rate Burst tests . . . . . . . . . . . . . . . . . 39 122 8.5. Combined and Implicit Tests . . . . . . . . . . . . . . . 40 123 8.5.1. Sustained Bursts Test . . . . . . . . . . . . . . . . 40 124 8.5.2. Passive Measurements . . . . . . . . . . . . . . . . 41 125 9. An Example . . . . . . . . . . . . . . . . . . . . . . . . . 42 126 9.1. Observations about applicability . . . . . . . . . . . . 43 127 10. Validation . . . . . . . . . . . . . . . . . . . . . . . . . 44 128 11. Security Considerations . . . . . . . . . . . . . . . . . . . 45 129 12. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . 45 130 13. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 46 131 14. References . . . . . . . . . . . . . . . . . . . . . . . . . 46 132 14.1. Normative References . . . . . . . . . . . . . . . . . . 46 133 14.2. Informative References . . . . . . . . . . . . . . . . . 46 134 Appendix A. Model Derivations . . . . . . . . . . . . . . . . . 49 135 A.1. Queueless Reno . . . . . . . . . . . . . . . . . . . . . 50 136 Appendix B. The effects of ACK scheduling . . . . . . . . . . . 51 137 Appendix C. Version Control . . . . . . . . . . . . . . . . . . 52 138 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 52 140 1. Introduction 142 Model Based Metrics (MBM) rely on peer-reviewed mathematical models 143 to specify a Targeted Suite of IP Diagnostic tests, designed to 144 assess whether common transport protocols can be expected to meet a 145 predetermined Target Transport Performance over an Internet path. 146 This note describes the modeling framework to derive the test 147 parameters for assessing an Internet path's ability to support a 148 predetermined Bulk Transport Capacity. 150 Each test in the Targeted IP Diagnostic Suite (TIDS) measures some 151 aspect of IP packet transfer needed to meet the Target Transport 152 Performance. For Bulk Transport Capacity the TIDS includes IP 153 diagnostic tests to verify that there is: sufficient IP capacity 154 (data rate); sufficient queue space at bottlenecks to absorb and 155 deliver typical transport bursts; and that the background packet loss 156 ratio is low enough not to interfere with congestion control; and 157 other properties described below. Unlike typical IPPM metrics which 158 yield measures of network properties, Model Based Metrics nominally 159 yield pass/fail evaluations of the ability of standard transport 160 protocols to meet the specific performance objective over some 161 network path. 163 In most cases, the IP diagnostic tests can be implemented by 164 combining existing IPPM metrics with additional controls for 165 generating test streams having a specified temporal structure (bursts 166 or standing queues caused by constant bit rate streams, etc.) and 167 statistical criteria for evaluating packet transfer. The temporal 168 structure of the test streams mimic transport protocol behavior over 169 the complete path; the statistical criteria models the transport 170 protocol's response to less than ideal IP packet transfer. 172 This note addresses Bulk Transport Capacity. It describes an 173 alternative to the approach presented in "A Framework for Defining 174 Empirical Bulk Transfer Capacity Metrics" [RFC3148]. Other Model 175 Based Metrics may cover other applications and transports, such as 176 VoIP over UDP and RTP, and new transport protocols. 178 The MBM approach, mapping Target Transport Performance to a Targeted 179 IP Diagnostic Suite (TIDS) of IP tests, solves some intrinsic 180 problems with using TCP or other throughput maximizing protocols for 181 measurement. In particular all throughput maximizing protocols (and 182 TCP congestion control in particular) cause some level of congestion 183 in order to detect when they have reached the available capacity 184 limitation of the network. This self inflicted congestion obscures 185 the network properties of interest and introduces non-linear dynamic 186 equilibrium behaviors that make any resulting measurements useless as 187 metrics because they have no predictive value for conditions or paths 188 different than that of the measurement itself. In order to prevent 189 these effects it is necessary to suppress the effects of TCP 190 congestion control in the measurement method. These issues are 191 discussed at length in Section 4. Readers whom are unfamiliar with 192 basic properties of TCP and TCP-like congestion control may find it 193 easier to start at Section 4 or Section 4.1. 195 A Targeted IP Diagnostic Suite does not have such difficulties. IP 196 diagnostics can be constructed such that they make strong statistical 197 statements about path properties that are independent of the 198 measurement details, such as vantage and choice of measurement 199 points. Model Based Metrics are designed to bridge the gap between 200 empirical IP measurements and expected TCP performance for multiple 201 standardized versions of TCP. 203 1.1. Version Control 205 RFC Editor: Please remove this entire subsection prior to 206 publication. 208 Please send comments about this draft to ippm@ietf.org. See 209 http://goo.gl/02tkD for more information including: interim drafts, 210 an up to date todo list and information on contributing. 212 Formatted: Mon Feb 27 13:49:06 PST 2017 214 Changes since -08 draft: 216 o Language, spelling and usage nits. 217 o Expanded the abstract describe the models. 218 o Remove superfluous standards like language 219 o Remove superfluous "future technology" language. 220 o Interconnects -> network interconnections. 221 o Added more labels to Figure 1. 222 o Defined Bulk Transport. 223 o Clarified "implied bottleneck IP capacity" 224 o Clarified the history of the BTC metrics. 225 o Clarified stochastic vs non-stochastic test traffic generation. 226 o Reworked Fig 2 and 6.1 "Mimicking slowstart" 227 o Described the unsynchronized parallel stream failure case. 228 o Discussed how to measure devices that use virtual queues. 229 o Changed section 8.5.2 (Streaming Media) to be Passive 230 Measurements. 232 Changes since -07 draft: 234 o Sharpened the use of "statistical criteria" 235 o Sharpened the definition of test_window, and removed related 236 redundant text in several places 237 o Clarified "equilibrium" as "dynamic equilibrium, similar to 238 processes observed in chemistry" 239 o Properly explained "Heisenberg" as "observer effect" 240 o Added the observation from RFC 6576 that HW and SW congestion 241 control implementations do not generally give the same results. 242 o Noted that IP and application metrics differ as to how overhead is 243 handled. MBM is explicit about how it handles overhead. 244 o Clarified the language and added a new reference about the 245 problems caused by token bucket policers. 246 o Added an subsection in the example that comments on some of issues 247 that need to be mentioned in a future usage or applicability doc. 248 o Updated ippm-2680-bis to RFC7680 249 o Many terminology, punctuation and spelling nits. 251 Changes since -06 draft: 253 o More language nits: 255 * "Targeted IP Diagnostic Suite (TIDS)" replaces "Targeted 256 Diagnostic Suite (TDS)". 257 * "implied bottleneck IP capacity" replaces "implied bottleneck 258 IP rate". 259 * Updated to ECN CE Marks. 260 * Added "specified temporal structure" 261 * "test stream" replaces "test traffic" 262 * "packet transfer" replaces "packet delivery" 263 * Reworked discussion of slowstart, bursts and pacing. 264 * RFC 7567 replaces RFC 2309. 266 Changes since -05 draft: 268 o Wordsmithing on sections overhauled in -05 draft. 269 o Reorganized the document: 271 * Relocated subsection "Preconditions". 272 * Relocated subsection "New Requirements relative to RFC 2330". 273 o Addressed nits and not so nits by Ruediger Geib. (Thanks!) 274 o Substantially tightened the entire definitions section. 275 o Many terminology changes, to better conform to other docs : 277 * IP rate and IP capacity (following RFC 5136) replaces various 278 forms of link data rate. 279 * subpath replaces link. 280 * target_window_size replaces target_pipe_size. 281 * implied bottleneck IP rate replaces effective bottleneck link 282 rate. 283 * Packet delivery statistics replaces delivery statistics. 285 Changes since -04 draft: 287 o The introduction was heavily overhauled: split into a separate 288 introduction and overview. 289 o The new shorter introduction: 291 * Is a problem statement; 292 * This document provides a framework; 293 * That it replaces TCP measurement by IP tests; 294 * That the results are pass/fail. 295 o Added a diagram of the framework to the overview 296 o and introduces all of the elements of the framework. 297 o Renumbered sections, reducing the depth of some section numbers. 298 o Updated definitions to better agree with other documents: 300 * Reordered section 2 301 * Bulk [data] performance -> Bulk Transport Capacity, everywhere 302 including the title. 303 * loss rate and loss probability -> packet loss ratio 304 * end-to-end path -> complete path 305 * [end-to-end][target] performance -> Target Transport 306 Performance 307 * load test -> capacity test 309 2. Overview 311 This document describes a modeling framework for deriving a Targeted 312 IP Diagnostic Suite from a predetermined Target Transport 313 Performance. It is not a complete specification, and relies on other 314 standards documents to define important details such as packet Type-P 315 selection, sampling techniques, vantage selection, etc. We imagine 316 Fully Specified - Targeted IP Diagnostic Suites (FS-TIDS), that 317 define all of these details. We use Targeted IP Diagnostic Suite 318 (TIDS) to refer to the subset of such a specification that is in 319 scope for this document. This terminology is defined in Section 3. 321 Section 4 describes some key aspects of TCP behavior and what they 322 imply about the requirements for IP packet transfer. Most of the IP 323 diagnostic tests needed to confirm that the path meets these 324 properties can be built on existing IPPM metrics, with the addition 325 of statistical criteria for evaluating packet transfer and in a few 326 cases, new mechanisms to implement the required temporal structure. 327 (One group of tests, the standing queue tests described in 328 Section 8.2, don't correspond to existing IPPM metrics, but suitable 329 new IPPM metrics can be patterned after the existing definitions.) 331 Figure 1 shows the MBM modeling and measurement framework. The 332 Target Transport Performance, at the top of the figure, is determined 333 by the needs of the user or application, outside the scope of this 334 document. For Bulk Transport Capacity, the main performance 335 parameter of interest is the Target Data Rate. However, since TCP's 336 ability to compensate for less than ideal network conditions is 337 fundamentally affected by the Round Trip Time (RTT) and the Maximum 338 Transmission Unit (MTU) of the complete path, these parameters must 339 also be specified in advance based on knowledge about the intended 340 application setting. They may reflect a specific application over a 341 real path through the Internet or an idealized application and 342 hypothetical path representing a typical user community. Section 5 343 describes the common parameters and models derived from the Target 344 Transport Performance. 346 Target Transport Performance 347 (Target Data Rate, Target RTT and Target MTU) 348 | 349 ________V_________ 350 | mathematical | 351 | models | 352 | | 353 ------------------ 354 Traffic parameters | | Statistical criteria 355 | | 356 _______V____________V____Targeted_______ 357 | | * * * | Diagnostic Suite | 358 _____|_______V____________V________________ | 359 __|____________V____________V______________ | | 360 | IP diagnostic tests | | | 361 | | | | | | 362 | _____________V__ __V____________ | | | 363 | | traffic | | Delivery | | | | 364 | | pattern | | Evaluation | | | | 365 | | generation | | | | | | 366 | -------v-------- ------^-------- | | | 367 | | v test stream via ^ | | |-- 368 | | -->======================>-- | | | 369 | | subpath under test | |- 370 ----V----------------------------------V--- | 371 | | | | | | 372 V V V V V V 373 fail/inconclusive pass/fail/inconclusive 374 (traffic generation status) (test result) 376 Overall Modeling Framework 378 Figure 1 380 The mathematical models are used to determine Traffic parameters and 381 subsequently to design traffic patterns that mimic TCP or other 382 transport protocol delivering bulk data and operating at the Target 383 Data Rate, MTU and RTT over a full range of conditions, including 384 flows that are bursty at multiple time scales. The traffic patterns 385 are generated based on the three Target parameters of complete path 386 and independent of the properties of individual subpaths using the 387 techniques described in Section 6. As much as possible the test 388 streams are generated deterministically (precomputed) to minimize the 389 extent to which test methodology, measurement points, measurement 390 vantage or path partitioning affect the details of the measurement 391 traffic. 393 Section 7 describes packet transfer statistics and methods to test 394 them against the statistical criteria provided by the mathematical 395 models. Since the statistical criteria typically apply to the 396 complete path (a composition of subpaths) [RFC6049], in situ testing 397 requires that the end-to-end statistical criteria be apportioned as 398 separate criteria for each subpath. Subpaths that are expected to be 399 bottlenecks would then be permitted to contribute a larger fraction 400 of the end-to-end packet loss budget. In compensation, subpaths that 401 to not exhibit bottlenecks have must be constrained to contribute 402 less packet loss. Thus the statistical criteria for each subpath in 403 each test of a TIDS is an apportioned share of the end-to-end 404 statistical criteria for the complete path which was determined by 405 the mathematical model. 407 Section 8 describes the suite of individual tests needed to verify 408 all of required IP delivery properties. A subpath passes if and only 409 if all of the individual IP diagnostic tests pass. Any subpath that 410 fails any test indicates that some users are likely to fail to attain 411 their Target Transport Performance under some conditions. In 412 addition to passing or failing, a test can be deemed to be 413 inconclusive for a number of reasons including: the precomputed 414 traffic pattern was not accurately generated; the measurement results 415 were not statistically significant; and others such as failing to 416 meet some required test preconditions. If all tests pass but some 417 are inconclusive, then the entire suite is deemed to be inconclusive. 419 In Section 9 we present an example TIDS that might be representative 420 of High Definition (HD) video, and illustrate how Model Based Metrics 421 can be used to address difficult measurement situations, such as 422 confirming that inter-carrier exchanges have sufficient performance 423 and capacity to deliver HD video between ISPs. 425 Since there is some uncertainty in the modeling process, Section 10 426 describes a validation procedure to diagnose and minimize false 427 positive and false negative results. 429 3. Terminology 431 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 432 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 433 document are to be interpreted as described in [RFC2119]. 435 Terms containing underscores (rather than spaces) appear in equations 436 and typically have algorithmic definitions. 438 General Terminology: 440 Target: A general term for any parameter specified by or derived 441 from the user's application or transport performance requirements. 442 Target Transport Performance: Application or transport performance 443 target values for the complete path. For Bulk Transport Capacity 444 defined in this note the Target Transport Performance includes the 445 Target Data Rate, Target RTT and Target MTU as described below. 446 Target Data Rate: The specified application data rate required for 447 an application's proper operation. Conventional Bulk Transport 448 Capacity (BTC) metrics are focused on the Target Data Rate, 449 however these metrics had little or no predictive value because 450 they do not consider the effects of the other two parameters of 451 the Target Transport Performance, the RTT and MTU of the complete 452 paths. 453 Target RTT (Round Trip Time): The specified baseline (minimum) RTT 454 of the longest complete path over which the user expects to be 455 able to meet the target performance. TCP and other transport 456 protocol's ability to compensate for path problems is generally 457 proportional to the number of round trips per second. The Target 458 RTT determines both key parameters of the traffic patterns (e.g. 459 burst sizes) and the thresholds on acceptable IP packet transfer 460 statistics. The Target RTT must be specified considering 461 appropriate packets sizes: MTU sized packets on the forward path, 462 ACK sized packets (typically header_overhead) on the return path. 463 Note that Target RTT is specified and not measured, MBM 464 measurements derived for a given target_RTT will be applicable to 465 any path with a smaller RTTs. 466 Target MTU (Maximum Transmission Unit): The specified maximum MTU 467 supported by the complete path the over which the application 468 expects to meet the target performance. In this document assume a 469 1500 Byte MTU unless otherwise specified. If some subpath has a 470 smaller MTU, then it becomes the Target MTU for the complete path, 471 and all model calculations and subpath tests must use the same 472 smaller MTU. 473 Targeted IP Diagnostic Suite (TIDS): A set of IP diagnostic tests 474 designed to determine if an otherwise ideal complete path 475 containing the subpath under test can sustain flows at a specific 476 target_data_rate using target_MTU sized packets when the RTT of 477 the complete path is target_RTT. 478 Fully Specified Targeted IP Diagnostic Suite (FS-TIDS): A TIDS 479 together with additional specification such as measurement packet 480 type ("type-p" [RFC2330]), etc. which are out of scope for this 481 document, but need to be drawn from other standards documents. 482 Bulk Transport Capacity: Bulk Transport Capacity Metrics evaluate an 483 Internet path's ability to carry bulk data, such as large files, 484 streaming (non-real time) video, and under some conditions, web 485 images and other content. Prior efforts to define BTC metrics 486 have been based on [RFC3148], which predates our understanding of 487 TCP and the requirements described in Section 4. In general "Bulk 488 Transport" indicates that performance is determined by the 489 interplay between the network, cross traffic and congestion 490 control in the transport protocol. It excludes situations where 491 performance is dominated by the RTT alone (e.g. transactions) or 492 bottlenecks elsewhere, such as in the application itself. 493 IP diagnostic tests: Measurements or diagnostics to determine if 494 packet transfer statistics meet some precomputed target. 495 traffic patterns: The temporal patterns or burstiness of traffic 496 generated by applications over transport protocols such as TCP. 497 There are several mechanisms that cause bursts at various time 498 scales as described in Section 4.1. Our goal here is to mimic the 499 range of common patterns (burst sizes and rates, etc), without 500 tying our applicability to specific applications, implementations 501 or technologies, which are sure to become stale. 502 packet transfer statistics: Raw, detailed or summary statistics 503 about packet transfer properties of the IP layer including packet 504 losses, ECN Congestion Experienced (CE) marks, reordering, or any 505 other properties that may be germane to transport performance. 506 packet loss ratio: As defined in [RFC7680]. 507 apportioned: To divide and allocate, for example budgeting packet 508 loss across multiple subpaths such that the losses will accumulate 509 to less than a specified end-to-end loss ratio. Apportioning 510 metrics is essentially the inverse of the process described in 511 [RFC5835]. 512 open loop: A control theory term used to describe a class of 513 techniques where systems that naturally exhibit circular 514 dependencies can be analyzed by suppressing some of the 515 dependencies, such that the resulting dependency graph is acyclic. 517 Terminology about paths, etc. See [RFC2330] and [RFC7398] for 518 existing terms and definitions. 520 data sender: Host sending data and receiving ACKs. 521 data receiver: Host receiving data and sending ACKs. 522 complete path: The end-to-end path from the data sender to the data 523 receiver. 525 subpath: A portion of the complete path. Note that there is no 526 requirement that subpaths be non-overlapping. A subpath can be a 527 small as a single device, link or interface. 528 measurement point: Measurement points as described in [RFC7398]. 529 test path: A path between two measurement points that includes a 530 subpath of the complete path under test. If the measurement 531 points are off path, the test path may include "test leads" 532 between the measurement points and the subpath. 533 dominant bottleneck: The bottleneck that generally determines most 534 of packet transfer statistics for the entire path. It typically 535 determines a flow's self clock timing, packet loss and ECN 536 Congestion Experienced (CE) marking rate, with other potential 537 bottlenecks having less effect on the packet transfer statistics. 538 See Section 4.1 on TCP properties. 539 front path: The subpath from the data sender to the dominant 540 bottleneck. 541 back path: The subpath from the dominant bottleneck to the receiver. 542 return path: The path taken by the ACKs from the data receiver to 543 the data sender. 544 cross traffic: Other, potentially interfering, traffic competing for 545 network resources (bandwidth and/or queue capacity). 547 Properties determined by the complete path and application. These 548 are described in more detail in Section 5.1. 550 Application Data Rate: General term for the data rate as seen by the 551 application above the transport layer in bytes per second. This 552 is the payload data rate, and explicitly excludes transport and 553 lower level headers (TCP/IP or other protocols), retransmissions 554 and other overhead that is not part to the total quantity of data 555 delivered to the application. 556 IP rate: The actual number of IP-layer bytes delivered through a 557 subpath, per unit time, including TCP and IP headers, retransmits 558 and other TCP/IP overhead. Follows from IP-type-P Link Usage 559 [RFC5136]. 560 IP capacity: The maximum number of IP-layer bytes that can be 561 transmitted through a subpath, per unit time, including TCP and IP 562 headers, retransmits and other TCP/IP overhead. Follows from IP- 563 type-P Link Capacity [RFC5136]. 564 bottleneck IP capacity: The IP capacity of the dominant bottleneck 565 in the forward path. All throughput maximizing protocols estimate 566 this capacity by observing the IP rate delivered through the 567 bottleneck. Most protocols derive their self clocks from the 568 timing of this data. See Section 4.1 and Appendix B for more 569 details. 570 implied bottleneck IP capacity: This is the bottleneck IP capacity 571 implied by the ACKs returning from the receiver. It is determined 572 by looking at how much application data the ACK stream at the 573 sender reports delivered to the data receiver per unit time at 574 various time scales. If the return path is thinning, batching or 575 otherwise altering the ACK timing the implied bottleneck IP 576 capacity over short time scales might be substantially larger than 577 the bottleneck IP capacity averaged over a full RTT. Since TCP 578 derives its clock from the data delivered through the bottleneck, 579 the front path must have sufficient buffering to absorb any data 580 bursts at the dimensions (size and IP rate) implied by the ACK 581 stream, which are potentially doubled during slowstart. If the 582 return path is not altering the ACK stream, then the implied 583 bottleneck IP capacity will be the same as the bottleneck IP 584 capacity. See Section 4.1 and Appendix B for more details. 585 sender interface rate: The IP rate which corresponds to the IP 586 capacity of the data sender's interface. Due to sender efficiency 587 algorithms including technologies such as TCP segmentation offload 588 (TSO), nearly all modern servers deliver data in bursts at full 589 interface link rate. Today 1 or 10 Gb/s are typical. 590 Header_overhead: The IP and TCP header sizes, which are the portion 591 of each MTU not available for carrying application payload. 592 Without loss of generality this is assumed to be the size for 593 returning acknowledgments (ACKs). For TCP, the Maximum Segment 594 Size (MSS) is the Target MTU minus the header_overhead. 596 Basic parameters common to models and subpath tests are defined here 597 are described in more detail in Section 5.2. Note that these are 598 mixed between application transport performance (excludes headers) 599 and IP performance (which include TCP headers and retransmissions as 600 part of the IP payload). 602 Window [size]: The total quantity of data carried by packets in- 603 flight plus the data represented by ACKs circulating in the 604 network is referred to as the window. See Section 4.1. Sometimes 605 used with other qualifiers (congestion window, cwnd or receiver 606 window) to indicate which mechanism is controlling the window. 607 pipe size: A general term for number of packets needed in flight 608 (the window size) to exactly fill some network path or subpath. 609 It corresponds to the window size which maximizes network power, 610 the observed data rate divided by the observed RTT. Often used 611 with additional qualifiers to specify which path, or under what 612 conditions, etc. 613 target_window_size: The average number of packets in flight (the 614 window size) needed to meet the Target Data Rate, for the 615 specified Target RTT, and MTU. It implies the scale of the bursts 616 that the network might experience. 617 run length: A general term for the observed, measured, or specified 618 number of packets that are (expected to be) delivered between 619 losses or ECN Congestion Experienced (CE) marks. Nominally one 620 over the sum of the loss and ECN CE marking probabilities, if 621 there are independently and identically distributed. 622 target_run_length: The target_run_length is an estimate of the 623 minimum number of non-congestion marked packets needed between 624 losses or ECN Congestion Experienced (CE) marks necessary to 625 attain the target_data_rate over a path with the specified 626 target_RTT and target_MTU, as computed by a mathematical model of 627 TCP congestion control. A reference calculation is shown in 628 Section 5.2 and alternatives in Appendix A 629 reference target_run_length: target_run_length computed precisely by 630 the method in Section 5.2. This is likely to be slightly more 631 conservative than required by modern TCP implementations. 633 Ancillary parameters used for some tests: 635 derating: Under some conditions the standard models are too 636 conservative. The modeling framework permits some latitude in 637 relaxing or "derating" some test parameters as described in 638 Section 5.3 in exchange for a more stringent TIDS validation 639 procedures, described in Section 10. Models can be derated by 640 including a multiplicative derating factor to make tests less 641 stringent. 642 subpath_IP_capacity: The IP capacity of a specific subpath. 643 test path: A subpath of a complete path under test. 644 test_path_RTT: The RTT observed between two measurement points using 645 packet sizes that are consistent with the transport protocol. 646 This is generally MTU sized packets of the forward path, 647 header_overhead sized packets on the return path. 648 test_path_pipe: The pipe size of a test path. Nominally the 649 test_path_RTT times the test path IP_capacity. 650 test_window: The smallest window sufficient to meet or exceed the 651 target_rate when operating with a pure self clock over a test 652 path. The test_window is typically given by 653 ceiling(target_data_rate*test_path_RTT/(target_MTU- 654 header_overhead)) but see the discussion in Appendix B about the 655 effects of channel scheduling on RTT. On some test paths the 656 test_window may need to be adjusted slightly to compensate for the 657 RTT being inflated by the devices that schedule packets. 659 The terminology below is used to define temporal patterns for test 660 stream. These patterns are designed to mimic TCP behavior, as 661 described in Section 4.1. 663 packet headway: Time interval between packets, specified from the 664 start of one to the start of the next. e.g. If packets are sent 665 with a 1 mS headway, there will be exactly 1000 packets per 666 second. 668 burst headway: Time interval between bursts, specified from the 669 start of the first packet one burst to the start of the first 670 packet of the next burst. e.g. If 4 packet bursts are sent with a 671 1 mS burst headway, there will be exactly 4000 packets per second. 672 paced single packets: Send individual packets at the specified rate 673 or packet headway. 674 paced bursts: Send bursts on a timer. Specify any 3 of: average 675 data rate, packet size, burst size (number of packets) and burst 676 headway (burst start to start). By default the bursts are assumed 677 to occur at full sender interface rate, such that the packet 678 headway within each burst is the minimum supported by the sender's 679 interface. Under some conditions it is useful to explicitly 680 specify the packet headway within each burst. 681 slowstart rate: Mimic TCP slowstart by sending 4 packet paced bursts 682 at an average data rate equal to twice the implied bottleneck IP 683 capacity (but not more than the sender interface rate). This is a 684 two level burst pattern described in more detail in Section 6.1. 685 If the implied bottleneck IP capacity is more than half of the 686 sender interface rate, slowstart rate becomes sender interface 687 rate. 688 slowstart burst: Mimic one round of TCP slowstart by sending a 689 specified number of packets packets in a two level burst pattern 690 that resembles slowstart. 691 repeated slowstart bursts: Repeat Slowstart bursts once per 692 target_RTT. For TCP each burst would be twice as large as the 693 prior burst, and the sequence would end at the first ECN CE mark 694 or lost packet. For measurement, all slowstart bursts would be 695 the same size (nominally target_window_size but other sizes might 696 be specified), and the ECN CE marks and lost packets are counted. 698 The tests described in this note can be grouped according to their 699 applicability. 701 Capacity tests: Capacity tests determine if a network subpath has 702 sufficient capacity to deliver the Target Transport Performance. 703 As long as the test stream is within the proper envelope for the 704 Target Transport Performance, the average packet losses or ECN 705 Congestion Experienced (CE) marks must be below the statistical 706 criteria computed by the model. As such, capacity tests reflect 707 parameters that can transition from passing to failing as a 708 consequence of cross traffic, additional presented load or the 709 actions of other network users. By definition, capacity tests 710 also consume significant network resources (data capacity and/or 711 queue buffer space), and the test schedules must be balanced by 712 their cost. 713 Monitoring tests: Monitoring tests are designed to capture the most 714 important aspects of a capacity test, but without presenting 715 excessive ongoing load themselves. As such they may miss some 716 details of the network's performance, but can serve as a useful 717 reduced-cost proxy for a capacity test, for example to support 718 continuous production network monitoring. 719 Engineering tests: Engineering tests evaluate how network algorithms 720 (such as AQM and channel allocation) interact with TCP-style self 721 clocked protocols and adaptive congestion control based on packet 722 loss and ECN Congestion Experienced (CE) marks. These tests are 723 likely to have complicated interactions with cross traffic and 724 under some conditions can be inversely sensitive to load. For 725 example a test to verify that an AQM algorithm causes ECN CE marks 726 or packet drops early enough to limit queue occupancy may 727 experience a false pass result in the presence of cross traffic. 728 It is important that engineering tests be performed under a wide 729 range of conditions, including both in situ and bench testing, and 730 over a wide variety of load conditions. Ongoing monitoring is 731 less likely to be useful for engineering tests, although sparse in 732 situ testing might be appropriate. 734 4. Background 736 At the time the "Framework for IP Performance Metrics" [RFC2330] was 737 published (1998), sound Bulk Transport Capacity (BTC) measurement was 738 known to be well beyond our capabilities. Even when Framework for 739 Empirical BTC Metrics [RFC3148] was published, we knew that we didn't 740 really understand the problem. Now, by hindsight we understand why 741 assessing BTC is such a hard problem: 743 o TCP is a control system with circular dependencies - everything 744 affects performance, including components that are explicitly not 745 part of the test (for example, the host processing power is not 746 in-scope of path performance tests). 747 o Congestion control is a dynamic equilibrium process, similar to 748 processes observed in chemistry and other fields. The network and 749 transport protocols find an operating point which balances between 750 opposing forces: the transport protocol pushing harder (raising 751 the data rate and/or window) while the network pushes back 752 (raising packet loss ratio, RTT and/or ECN CE marks). By design 753 TCP congestion control keeps raising the data rate until the 754 network gives some indication that its capacity has been exceeded 755 by dropping packets or adding ECN CE marks. If a TCP sender 756 accurately fills a path to its IP capacity, (e.g. the bottleneck 757 is 100% utilized), then packet losses and ECN CE marks are mostly 758 determined by the TCP sender and how aggressively it seeks 759 additional capacity, and not the network itself, since the network 760 must send exactly the signals that TCP needs to set its rate. 761 o TCP's ability to compensate for network impairments (such as loss, 762 delay and delay variation, outside of those caused by TCP itself) 763 is directly proportional to the number of send-ACK round trip 764 exchanges per second (i.e. inversely proportional to the RTT). As 765 a consequence an impaired subpath may pass a short RTT local test 766 even though it fails when the subpath is extended by an 767 effectively perfect network to some larger RTT. 768 o TCP has an extreme form of the Observer Effect (colloquially know 769 as the Heisenberg effect). Measurement and cross traffic interact 770 in unknown and ill defined ways. The situation is actually worse 771 than the traditional physics problem where you can at least 772 estimate bounds on the relative momentum of the measurement and 773 measured particles. For network measurement you can not in 774 general determine even the order of magnitude of the effect. It 775 is possible to construct measurement scenarios where the 776 measurement traffic starves real user traffic, yielding an overly 777 inflated measurement. The inverse is also possible: the user 778 traffic can fill the network, such that the measurement traffic 779 detects only minimal available capacity. You can not in general 780 determine which scenario might be in effect, so you can not gauge 781 the relative magnitude of the uncertainty introduced by 782 interactions with other network traffic. 783 o As a consequence of the properties listed above it is difficult, 784 if not impossible, for two independent implementations (HW or SW) 785 of TCP congestion control to produce equivalent performance 786 results [RFC6576] under the same network conditions, 788 These properties are a consequence of the dynamic equilibrium 789 behavior intrinsic to how all throughput maximizing protocols 790 interact with the Internet. These protocols rely on control systems 791 based on estimated network metrics to regulate the quantity of data 792 to send into the network. The packet sending characteristics in turn 793 alter the network properties estimated by the control system metrics, 794 such that there are circular dependencies between every transmission 795 characteristic and every estimated metric. Since some of these 796 dependencies are nonlinear, the entire system is nonlinear, and any 797 change anywhere causes a difficult to predict response in network 798 metrics. As a consequence Bulk Transport Capacity metrics have 799 entirely thwarted the analytic framework envisioned in [RFC2330] 801 Model Based Metrics overcome these problems by making the measurement 802 system open loop: the packet transfer statistics (akin to the network 803 estimators) do not affect the traffic or traffic patterns (bursts), 804 which are computed on the basis of the Target Transport Performance. 805 A path or subpath meeting the Target Transfer Performance 806 requirements would exhibit packet transfer statistics and estimated 807 metrics that would not cause the control system to slow the traffic 808 below the Target Data Rate. 810 4.1. TCP properties 812 TCP and other self clocked protocols (e.g. SCTP) carry the vast 813 majority of all Internet data. Their dominant bulk data transport 814 behavior is to have an approximately fixed quantity of data and 815 acknowledgments (ACKs) circulating in the network. The data receiver 816 reports arriving data by returning ACKs to the data sender, the data 817 sender typically responds by sending approximately the same quantity 818 of data back into the network. The total quantity of data plus the 819 data represented by ACKs circulating in the network is referred to as 820 the window. The mandatory congestion control algorithms 821 incrementally adjust the window by sending slightly more or less data 822 in response to each ACK. The fundamentally important property of 823 this system is that it is self clocked: The data transmissions are a 824 reflection of the ACKs that were delivered by the network, the ACKs 825 are a reflection of the data arriving from the network. 827 A number of protocol features cause bursts of data, even in idealized 828 networks that can be modeled as simple queuing systems. 830 During slowstart the IP rate is doubled on each RTT by sending twice 831 as much data as was delivered to the receiver during the prior RTT. 832 Each returning ACK causes the sender to transmit twice the data the 833 ACK reported arriving at the receiver. For slowstart to be able to 834 fill the pipe, the network must be able to tolerate slowstart bursts 835 up to the full pipe size inflated by the anticipated window reduction 836 on the first loss or ECN CE mark. For example, with classic Reno 837 congestion control, an optimal slowstart has to end with a burst that 838 is twice the bottleneck rate for one RTT in duration. This burst 839 causes a queue which is equal to the pipe size (i.e. the window is 840 twice the pipe size) so when the window is halved in response to the 841 first packet loss, the new window will be the pipe size. 843 Note that if the bottleneck IP rate is less that half of the capacity 844 of the front path (which is almost always the case), the slowstart 845 bursts will not by themselves cause significant queues anywhere else 846 along the front path; they primarily exercise the queue at the 847 dominant bottleneck. 849 Several common efficiency algorithms also cause bursts. The self 850 clock is typically applied to groups of packets: the receiver's 851 delayed ACK algorithm generally sends only one ACK per two data 852 segments. Furthermore the modern senders use TCP segmentation 853 offload (TSO) to reduce CPU overhead. The sender's software stack 854 builds super sized TCP segments that the TSO hardware splits into MTU 855 sized segments on the wire. The net effect of TSO, delayed ACK and 856 other efficiency algorithms is to send bursts of segments at full 857 sender interface rate. 859 Note that these efficiency algorithms are almost always in effect, 860 including during slowstart, such that slowstart typically has a two 861 level burst structure. Section 6.1 describes slowstart in more 862 detail. 864 Additional sources of bursts include TCP's initial window [RFC6928], 865 application pauses, channel allocation mechanisms and network devices 866 that schedule ACKs. Appendix B describes these last two items. If 867 the application pauses (stops reading or writing data) for some 868 fraction of an RTT, many TCP implementations catch up to their 869 earlier window size by sending a burst of data at the full sender 870 interface rate. To fill a network with a realistic application, the 871 network has to be able to tolerate sender interface rate bursts large 872 enough to restore the prior window following application pauses. 874 Although the sender interface rate bursts are typically smaller than 875 the last burst of a slowstart, they are at a higher IP rate so they 876 potentially exercise queues at arbitrary points along the front path 877 from the data sender up to and including the queue at the dominant 878 bottleneck. There is no model for how frequent or what sizes of 879 sender rate bursts the network should tolerate. 881 In conclusion, to verify that a path can meet a Target Transport 882 Performance, it is necessary to independently confirm that the path 883 can tolerate bursts at the scales that can be caused by the above 884 mechanisms. Three cases are believed to be sufficient: 886 o Two level slowstart bursts sufficient to get connections started 887 properly. 888 o Ubiquitous sender interface rate bursts caused by efficiency 889 algorithms. We assume 4 packet bursts to be the most common case, 890 since it matches the effects of delayed ACK during slowstart. 891 These bursts should be assumed not to significantly affect packet 892 transfer statistics. 893 o Infrequent sender interface rate bursts that are full 894 target_window_size. Target_run_length may be derated for these 895 large fast bursts. 897 If a subpath can meet the required packet loss ratio for bursts at 898 all of these scales then it has sufficient buffering at all potential 899 bottlenecks to tolerate any of the bursts that are likely introduced 900 by TCP or other transport protocols. 902 4.2. Diagnostic Approach 904 A complete path of a given RTT and MTU, which are equal to or smaller 905 than the Target RTT and equal to or larger than the Target MTU 906 respectively, is expected to be able to attain a specified Bulk 907 Transport Capacity when all of the following conditions are met: 909 1. The IP capacity is above the Target Data Rate by sufficient 910 margin to cover all TCP/IP overheads. This can be confirmed by 911 the tests described in Section 8.1 or any number of IP capacity 912 tests adapted to implement MBM. 913 2. The observed packet transfer statistics are better than required 914 by a suitable TCP performance model (e.g. fewer packet losses or 915 ECN CE marks). See Section 8.1 or any number of low or fixed 916 rate packet loss tests outside of MBM. 917 3. There is sufficient buffering at the dominant bottleneck to 918 absorb a slowstart bursts large enough to get the flow out of 919 slowstart at a suitable window size. See Section 8.3. 920 4. There is sufficient buffering in the front path to absorb and 921 smooth sender interface rate bursts at all scales that are likely 922 to be generated by the application, any channel arbitration in 923 the ACK path or any other mechanisms. See Section 8.4. 924 5. When there is a slowly rising standing queue at the bottleneck 925 the onset of packet loss has to be at an appropriate point (time 926 or queue depth) and progressive [RFC7567]. See Section 8.2. 927 6. When there is a standing queue at a bottleneck for a shared media 928 subpath (e.g. half duplex), there must be a suitable bounds on 929 the interaction between ACKs and data, for example due to the 930 channel arbitration mechanism. See Section 8.2.4. 932 Note that conditions 1 through 4 require capacity tests for 933 validation, and thus may need to be monitored on an ongoing basis. 934 Conditions 5 and 6 require engineering tests, which are best 935 performed in controlled environments such as a bench test. They 936 won't generally fail due to load, but may fail in the field due to 937 configuration errors, etc. and should be spot checked. 939 A tool that can perform many of the tests is available from 940 [MBMSource]. 942 4.3. New requirements relative to RFC 2330 944 Model Based Metrics are designed to fulfill some additional 945 requirements that were not recognized at the time RFC 2330 was 946 written [RFC2330]. These missing requirements may have significantly 947 contributed to policy difficulties in the IP measurement space. Some 948 additional requirements are: 950 o IP metrics must be actionable by the ISP - they have to be 951 interpreted in terms of behaviors or properties at the IP or lower 952 layers, that an ISP can test, repair and verify. 954 o Metrics should be spatially composable, such that measures of 955 concatenated paths should be predictable from subpaths. 956 o Metrics must be vantage point invariant over a significant range 957 of measurement point choices, including off path measurement 958 points. The only requirements on MP selection should be that the 959 RTT between the MPs is below some reasonable bound, and that the 960 effects of the "test leads" connecting MPs to the subpath under 961 test can be can be calibrated out of the measurements. The latter 962 might be be accomplished if the test leads are effectively ideal 963 or their properties can be deducted from the measurements between 964 the MPs. While many of tests require that the test leads have at 965 least as much IP capacity as the subpath under test, some do not, 966 for example Background Packet Transfer Tests described in 967 Section 8.1.3. 968 o Metric measurements should be repeatable by multiple parties with 969 no specialized access to MPs or diagnostic infrastructure. It 970 should be possible for different parties to make the same 971 measurement and observe the same results. In particular it is 972 specifically important that both a consumer (or their delegate) 973 and ISP be able to perform the same measurement and get the same 974 result. Note that vantage independence is key to meeting this 975 requirement. 977 5. Common Models and Parameters 979 5.1. Target End-to-end parameters 981 The target end-to-end parameters are the Target Data Rate, Target RTT 982 and Target MTU as defined in Section 3. These parameters are 983 determined by the needs of the application or the ultimate end user 984 and the complete Internet path over which the application is expected 985 to operate. The target parameters are in units that make sense to 986 upper layers: payload bytes delivered to the application, above TCP. 987 They exclude overheads associated with TCP and IP headers, 988 retransmits and other protocols (e.g. DNS). Note that IP-based 989 network services include TCP headers and retransmissions as part of 990 delivered payload, and this difference is recognized in calculations 991 below (header_overhead). 993 Other end-to-end parameters defined in Section 3 include the 994 effective bottleneck data rate, the sender interface data rate and 995 the TCP and IP header sizes. 997 The target_data_rate must be smaller than all subpath IP capacities 998 by enough headroom to carry the transport protocol overhead, 999 explicitly including retransmissions and an allowance for 1000 fluctuations in TCP's actual data rate. Specifying a 1001 target_data_rate with insufficient headroom is likely to result in 1002 brittle measurements having little predictive value. 1004 Note that the target parameters can be specified for a hypothetical 1005 path, for example to construct TIDS designed for bench testing in the 1006 absence of a real application; or for a live in situ test of 1007 production infrastructure. 1009 The number of concurrent connections is explicitly not a parameter to 1010 this model. If a subpath requires multiple connections in order to 1011 meet the specified performance, that must be stated explicitly and 1012 the procedure described in Section 6.4 applies. 1014 5.2. Common Model Calculations 1016 The Target Transport Performance is used to derive the 1017 target_window_size and the reference target_run_length. 1019 The target_window_size, is the average window size in packets needed 1020 to meet the target_rate, for the specified target_RTT and target_MTU. 1021 It is given by: 1023 target_window_size = ceiling( target_rate * target_RTT / ( target_MTU 1024 - header_overhead ) ) 1026 Target_run_length is an estimate of the minimum required number of 1027 unmarked packets that must be delivered between losses or ECN 1028 Congestion Experienced (CE) marks, as computed by a mathematical 1029 model of TCP congestion control. The derivation here follows 1030 [MSMO97], and by design is quite conservative. 1032 Reference target_run_length is derived as follows: assume the 1033 subpath_IP_capacity is infinitesimally larger than the 1034 target_data_rate plus the required header_overhead. Then 1035 target_window_size also predicts the onset of queuing. A larger 1036 window will cause a standing queue at the bottleneck. 1038 Assume the transport protocol is using standard Reno style Additive 1039 Increase, Multiplicative Decrease (AIMD) congestion control [RFC5681] 1040 (but not Appropriate Byte Counting [RFC3465]) and the receiver is 1041 using standard delayed ACKs. Reno increases the window by one packet 1042 every pipe_size worth of ACKs. With delayed ACKs this takes 2 Round 1043 Trip Times per increase. To exactly fill the pipe, the spacing of 1044 losses must be no closer than when the peak of the AIMD sawtooth 1045 reached exactly twice the target_window_size. Otherwise, the 1046 multiplicative window reduction triggered by the loss would cause the 1047 network to be under-filled. Following [MSMO97] the number of packets 1048 between losses must be the area under the AIMD sawtooth. They must 1049 be no more frequent than every 1 in 1050 ((3/2)*target_window_size)*(2*target_window_size) packets, which 1051 simplifies to: 1053 target_run_length = 3*(target_window_size^2) 1055 Note that this calculation is very conservative and is based on a 1056 number of assumptions that may not apply. Appendix A discusses these 1057 assumptions and provides some alternative models. If a different 1058 model is used, a FS-TIDS must document the actual method for 1059 computing target_run_length and ratio between alternate 1060 target_run_length and the reference target_run_length calculated 1061 above, along with a discussion of the rationale for the underlying 1062 assumptions. 1064 These two parameters, target_window_size and target_run_length, 1065 directly imply most of the individual parameters for the tests in 1066 Section 8. 1068 5.3. Parameter Derating 1070 Since some aspects of the models are very conservative, the MBM 1071 framework permits some latitude in derating test parameters. Rather 1072 than trying to formalize more complicated models we permit some test 1073 parameters to be relaxed as long as they meet some additional 1074 procedural constraints: 1076 o The FS-TIDS must document and justify the actual method used to 1077 compute the derated metric parameters. 1078 o The validation procedures described in Section 10 must be used to 1079 demonstrate the feasibility of meeting the Target Transport 1080 Performance with infrastructure that infinitesimally passes the 1081 derated tests. 1082 o The validation process for a FS-TIDS itself must be documented is 1083 such a way that other researchers can duplicate the validation 1084 experiments. 1086 Except as noted, all tests below assume no derating. Tests where 1087 there is not currently a well established model for the required 1088 parameters explicitly include derating as a way to indicate 1089 flexibility in the parameters. 1091 5.4. Test Preconditions 1093 Many tests have preconditions which are required to assure their 1094 validity. Examples include: the presence or non-presence of cross 1095 traffic on specific subpaths; negotiating ECN; and appropriate 1096 preamble packet stream to testing to put reactive network elements 1097 into the proper states [RFC7312]. If preconditions are not properly 1098 satisfied for some reason, the tests should be considered to be 1099 inconclusive. In general it is useful to preserve diagnostic 1100 information as to why the preconditions were not met, and any test 1101 data that was collected even if it is not useful for the intended 1102 test. Such diagnostic information and partial test data may be 1103 useful for improving the test or test procedures themselves. 1105 It is important to preserve the record that a test was scheduled, 1106 because otherwise precondition enforcement mechanisms can introduce 1107 sampling bias. For example, canceling tests due to cross traffic on 1108 subscriber access links might introduce sampling bias in tests of the 1109 rest of the network by reducing the number of tests during peak 1110 network load. 1112 Test preconditions and failure actions must be specified in a FS- 1113 TIDS. 1115 6. Generating test streams 1117 Many important properties of Model Based Metrics, such as vantage 1118 independence, are a consequence of using test streams that have 1119 temporal structures that mimic TCP or other transport protocols 1120 running over a complete path. As described in Section 4.1, self 1121 clocked protocols naturally have burst structures related to the RTT 1122 and pipe size of the complete path. These bursts naturally get 1123 larger (contain more packets) as either the Target RTT or Target Data 1124 Rate get larger, or the Target MTU gets smaller. An implication of 1125 these relationships is that test streams generated by running self 1126 clocked protocols over short subpaths may not adequately exercise the 1127 queuing at any bottleneck to determine if the subpath can support the 1128 full Target Transport Performance over the complete path. 1130 Failing to authentically mimic TCP's temporal structure is part of 1131 the reason why simple performance tools such as iPerf, netperf, nc, 1132 etc have the reputation of yielding false pass results over short 1133 test paths, even when some subpath has a flaw. 1135 The definitions in Section 3 are sufficient for most test streams. 1136 We describe the slowstart and standing queue test streams in more 1137 detail. 1139 In conventional measurement practice stochastic processes are used to 1140 eliminate many unintended correlations and sample biases. However 1141 MBM tests are designed to explicitly mimic temporal correlations 1142 caused by network or protocol elements themselves. Some portions of 1143 these system, such as traffic arrival (test scheduling) are naturally 1144 stochastic. Other behaviors, such as back-to-back packet 1145 transmissions, are dominated by implementation specific deterministic 1146 effects. Although these behaviors always contain non-deterministic 1147 elements and might be modeled stochastically, these details typically 1148 do not contribute significantly to the overall system behavior. 1149 Furthermore, it is known that real protocols are subject to failures 1150 caused by network property estimators suffering from bias due to 1151 correlation in their own traffic. For example TCP's RTT estimator 1152 used to determine the Retransmit Time Out (RTO), can be fooled by 1153 periodic cross traffic or start-stop applications. For these reasons 1154 many details of the test streams are specified deterministically. 1156 It may prove useful to introduce fine grained noise sources into the 1157 models used for generating test streams in an update of Model Based 1158 Metrics, but the complexity is not warranted at the time this 1159 document was written. 1161 6.1. Mimicking slowstart 1163 TCP slowstart has a two level burst structure as shown in Figure 2. 1164 The fine time structure is caused by efficiency algorithms that 1165 deliberately batch work (CPU, channel allocation, etc) to better 1166 amortize certain network and host overheads. ACKs passing through 1167 the return path typically cause the sender to transmit small bursts 1168 of data at full sender interface rate. For example TCP Segmentation 1169 Offload (TSO) and Delayed Acknowledgment both contribute to this 1170 effect. During slowstart these bursts are at the same headway as the 1171 returning ACKs, but are typically twice as large (e.g. having twice 1172 as much data) as the ACK reported was delivered to the receiver. Due 1173 to variations in delayed ACK and algorithms such as Appropriate Byte 1174 Counting [RFC3465], different pairs of senders and receivers produce 1175 slightly different burst patterns. Without loss of generality, we 1176 assume each ACK causes 4 packet sender interface rate bursts at an 1177 average headway equal to the ACK headway, and corresponding to 1178 sending at an average rate equal to twice the effective bottleneck IP 1179 rate. Each slowstart burst consists of a series of 4 packet sender 1180 interface rate bursts such that the total number of packets is the 1181 current window size (as of the last packet in the burst). 1183 The coarse time structure is due to each RTT being a reflection of 1184 the prior RTT. For real transport protocols, each slowstart burst is 1185 twice as large (twice the window) as the previous burst but is spread 1186 out in time by the network bottleneck, such that each successive RTT 1187 exhibits the same effective bottleneck IP rate. The slowstart phase 1188 ends on the first lost packet or ECN mark, which is intended to 1189 happen after successive slowstart bursts merge in time: the next 1190 burst starts before the bottleneck queue is fully drained and the 1191 prior burst is complete. 1193 For diagnostic tests described below we preserve the fine time 1194 structure but manipulate the coarse structure of the slowstart bursts 1195 (burst size and headway) to measure the ability of the dominant 1196 bottleneck to absorb and smooth slowstart bursts. 1198 Note that a stream of repeated slowstart bursts has three different 1199 average rates, depending on the averaging time interval. At the 1200 finest time scale (a few packet times at the sender interface) the 1201 peak of the average IP rate is the same as the sender interface rate; 1202 at a medium timescale (a few ACK times at the dominant bottleneck) 1203 the peak of the average IP rate is twice the implied bottleneck IP 1204 capacity; and at time scales longer than the target_RTT and when the 1205 burst size is equal to the target_window_size, the average rate is 1206 equal to the target_data_rate. This pattern corresponds to repeating 1207 the last RTT of TCP slowstart when delayed ACK and sender side byte 1208 counting are present but without the limits specified in Appropriate 1209 Byte Counting [RFC3465]. 1211 time ==> ( - equals one packet) 1213 Fine time structure of the packet stream: 1215 ---- ---- ---- ---- ---- 1217 |<>| sender interface rate bursts (typically 3 or 4 packets) 1218 |<===>| burst headway (from the ACK headway) 1220 \____repeating sender______/ 1221 rate bursts 1223 Coarse (RTT level) time structure of the packet stream: 1225 ---- ---- ---- ---- ---- ---- ---- ... 1227 |<========================>| slowstart burst size (from the window) 1228 |<==============================================>| slowstart headway 1229 (from the RTT) 1230 \__________________________/ \_________ ... 1231 one slowstart burst Repeated slowstart bursts 1233 Multiple levels of Slowstart Bursts 1235 Figure 2 1237 6.2. Constant window pseudo CBR 1239 Implement pseudo constant bit rate by running a standard self clocked 1240 protocol such as TCP with a fixed window size. If that window size 1241 is test_window, the data rate will be slightly above the target_rate. 1243 Since the test_window is constrained to be an integer number of 1244 packets, for small RTTs or low data rates there may not be 1245 sufficiently precise control over the data rate. Rounding the 1246 test_window up (the default) is likely to result in data rates that 1247 are higher than the target rate, but reducing the window by one 1248 packet may result in data rates that are too small. Also cross 1249 traffic potentially raises the RTT, implicitly reducing the rate. 1250 Cross traffic that raises the RTT nearly always makes the test more 1251 strenuous (more demanding for the network path). A FS-TIDS 1252 specifying a constant window CBR test must explicitly indicate under 1253 what conditions errors in the data rate cause tests to inconclusive. 1255 Since constant window pseudo CBR testing is sensitive to RTT 1256 fluctuations it will be less accurate at controlling the data rate in 1257 environments with fluctuating delays. Conventional paced measurement 1258 traffic may be more appropriate for these environments. 1260 6.3. Scanned window pseudo CBR 1262 Scanned window pseudo CBR is similar to the constant window CBR 1263 described above, except the window is scanned across a range of sizes 1264 designed to include two key events, the onset of queuing and the 1265 onset of packet loss or ECN CE marks. The window is scanned by 1266 incrementing it by one packet every 2*target_window_size delivered 1267 packets. This mimics the additive increase phase of standard Reno 1268 TCP congestion avoidance when delayed ACKs are in effect. Normally 1269 the window increases separated by intervals slightly longer than 1270 twice the target_RTT. 1272 There are two ways to implement this test: one built by applying a 1273 window clamp to standard congestion control in a standard protocol 1274 such as TCP and the other built by stiffening a non-standard 1275 transport protocol. When standard congestion control is in effect, 1276 any losses or ECN CE marks cause the transport to revert to a window 1277 smaller than the clamp such that the scanning clamp loses control the 1278 window size. The NPAD pathdiag tool is an example of this class of 1279 algorithms [Pathdiag]. 1281 Alternatively a non-standard congestion control algorithm can respond 1282 to losses by transmitting extra data, such that it maintains the 1283 specified window size independent of losses or ECN CE marks. Such a 1284 stiffened transport explicitly violates mandatory Internet congestion 1285 control [RFC5681] and is not suitable for in situ testing. It is 1286 only appropriate for engineering testing under laboratory conditions. 1287 The Windowed Ping tool implements such a test [WPING]. The tool 1288 described in the paper has been updated.[mpingSource] 1290 The test procedures in Section 8.2 describe how to the partition the 1291 scans into regions and how to interpret the results. 1293 6.4. Concurrent or channelized testing 1295 The procedures described in this document are only directly 1296 applicable to single stream measurement, e.g. one TCP connection or 1297 measurement stream. In an ideal world, we would disallow all 1298 performance claims based multiple concurrent streams, but this is not 1299 practical due to at least two issues. First, many very high rate 1300 link technologies are channelized and at last partially pin the flow 1301 to channel mapping to minimize packet reordering within flows. 1302 Second, TCP itself has scaling limits. Although the former problem 1303 might be overcome through different design decisions, the later 1304 problem is more deeply rooted. 1306 All congestion control algorithms that are philosophically aligned 1307 with the standard [RFC5681] (e.g. claim some level of TCP 1308 compatibility, friendliness or fairness) have scaling limits, in the 1309 sense that as a long fast network (LFN) with a fixed RTT and MTU gets 1310 faster, these congestion control algorithms get less accurate and as 1311 a consequence have difficulty filling the network [CCscaling]. These 1312 properties are a consequence of the original Reno AIMD congestion 1313 control design and the requirement in [RFC5681] that all transport 1314 protocols have similar responses to congestion. 1316 There are a number of reasons to want to specify performance in terms 1317 of multiple concurrent flows, however this approach is not 1318 recommended for data rates below several megabits per second, which 1319 can be attained with run lengths under 10000 packets on many paths. 1320 Since the required run length goes as the square of the data rate, at 1321 higher rates the run lengths can be unreasonably large, and multiple 1322 flows might be the only feasible approach. 1324 If multiple flows are deemed necessary to meet aggregate performance 1325 targets then this MUST be stated in both the design of the TIDS and 1326 in any claims about network performance. The IP diagnostic tests 1327 MUST be performed concurrently with the specified number of 1328 connections. For the tests that use bursty test streams, the bursts 1329 should be synchronized across streams unless there is a priori 1330 knowledge that the applications have some explicit mechanism to 1331 stagger their own bursts. In the absences of an explicit mechanism 1332 to stagger bursts many network and application artifacts will 1333 sometimes implicitly synchronize bursts. A test that does not 1334 control burst synchronization may be prone to false pass results for 1335 some applications. 1337 7. Interpreting the Results 1339 7.1. Test outcomes 1341 To perform an exhaustive test of a complete network path, each test 1342 of the TIDS is applied to each subpath of the complete path. If any 1343 subpath fails any test then a standard transport protocol running 1344 over the complete path can also be expected to fail to attain the 1345 Target Transport Performance under some conditions. 1347 In addition to passing or failing, a test can be deemed to be 1348 inconclusive for a number of reasons. Proper instrumentation and 1349 treatment of inconclusive outcomes is critical to the accuracy and 1350 robustness of Model Based Metrics. Tests can be inconclusive if the 1351 precomputed traffic pattern or data rates were not accurately 1352 generated; the measurement results were not statistically 1353 significant; and others causes such as failing to meet some required 1354 preconditions for the test. See Section 5.4 1356 For example consider a test that implements Constant Window Pseudo 1357 CBR (Section 6.2) by adding rate controls and detailed IP packet 1358 transfer instrumentation to TCP (e.g. [RFC4898]). TCP includes 1359 built in control systems which might interfere with the sending data 1360 rate. If such a test meets the required packet transfer statistics 1361 (e.g. run length) while failing to attain the specified data rate it 1362 must be treated as an inconclusive result, because we can not a 1363 priori determine if the reduced data rate was caused by a TCP problem 1364 or a network problem, or if the reduced data rate had a material 1365 effect on the observed packet transfer statistics. 1367 Note that for capacity tests, if the observed packet transfer 1368 statistics meet the statistical criteria for failing (accepting 1369 hypnosis H1 in Section 7.2), the test can can be considered to have 1370 failed because it doesn't really matter that the test didn't attain 1371 the required data rate. 1373 The really important new properties of MBM, such as vantage 1374 independence, are a direct consequence of opening the control loops 1375 in the protocols, such that the test stream does not depend on 1376 network conditions or IP packets received. Any mechanism that 1377 introduces feedback between the path's measurements and the test 1378 stream generation is at risk of introducing nonlinearities that spoil 1379 these properties. Any exceptional event that indicates that such 1380 feedback has happened should cause the test to be considered 1381 inconclusive. 1383 One way to view inconclusive tests is that they reflect situations 1384 where a test outcome is ambiguous between limitations of the network 1385 and some unknown limitation of the IP diagnostic test itself, which 1386 may have been caused by some uncontrolled feedback from the network. 1388 Note that procedures that attempt to search the target parameter 1389 space to find the limits on some parameter such as target_data_rate 1390 are at risk of breaking the location independent properties of Model 1391 Based Metrics, if any part of the boundary between passing and 1392 inconclusive or failing results is sensitive to RTT (which is 1393 normally the case). For example the maximum data rate for a marginal 1394 link (e.g. exhibiting excess errors) is likely to be sensitive to 1395 the test_path_RTT. The maximum observed data rate over the test path 1396 has very little predictive value for the maximum rate over a 1397 different path. 1399 One of the goals for evolving TIDS designs will be to keep sharpening 1400 distinction between inconclusive, passing and failing tests. The 1401 criteria for for passing, failing and inconclusive tests MUST be 1402 explicitly stated for every test in the TIDS or FS-TIDS. 1404 One of the goals of evolving the testing process, procedures, tools 1405 and measurement point selection should be to minimize the number of 1406 inconclusive tests. 1408 It may be useful to keep raw packet transfer statistics and ancillary 1409 metrics [RFC3148] for deeper study of the behavior of the network 1410 path and to measure the tools themselves. Raw packet transfer 1411 statistics can help to drive tool evolution. Under some conditions 1412 it might be possible to re-evaluate the raw data for satisfying 1413 alternate Target Transport Performance. However it is important to 1414 guard against sampling bias and other implicit feedback which can 1415 cause false results and exhibit measurement point vantage 1416 sensitivity. Simply applying different delivery criteria based on a 1417 different Target Transport Performance is insufficient if the test 1418 traffic patterns (bursts, etc.) does not match the alternate Target 1419 Transport Performance. 1421 7.2. Statistical criteria for estimating run_length 1423 When evaluating the observed run_length, we need to determine 1424 appropriate packet stream sizes and acceptable error levels for 1425 efficient measurement. In practice, can we compare the empirically 1426 estimated packet loss and ECN Congestion Experienced (CE) marking 1427 ratios with the targets as the sample size grows? How large a sample 1428 is needed to say that the measurements of packet transfer indicate a 1429 particular run length is present? 1431 The generalized measurement can be described as recursive testing: 1432 send packets (individually or in patterns) and observe the packet 1433 transfer performance (packet loss ratio or other metric, any marking 1434 we define). 1436 As each packet is sent and measured, we have an ongoing estimate of 1437 the performance in terms of the ratio of packet loss or ECN CE mark 1438 to total packets (i.e. an empirical probability). We continue to 1439 send until conditions support a conclusion or a maximum sending limit 1440 has been reached. 1442 We have a target_mark_probability, 1 mark per target_run_length, 1443 where a "mark" is defined as a lost packet, a packet with ECN CE 1444 mark, or other signal. This constitutes the null Hypothesis: 1446 H0: no more than one mark in target_run_length = 1447 3*(target_window_size)^2 packets 1449 and we can stop sending packets if on-going measurements support 1450 accepting H0 with the specified Type I error = alpha (= 0.05 for 1451 example). 1453 We also have an alternative Hypothesis to evaluate: if performance is 1454 significantly lower than the target_mark_probability. Based on 1455 analysis of typical values and practical limits on measurement 1456 duration, we choose four times the H0 probability: 1458 H1: one or more marks in (target_run_length/4) packets 1460 and we can stop sending packets if measurements support rejecting H0 1461 with the specified Type II error = beta (= 0.05 for example), thus 1462 preferring the alternate hypothesis H1. 1464 H0 and H1 constitute the Success and Failure outcomes described 1465 elsewhere in the memo, and while the ongoing measurements do not 1466 support either hypothesis the current status of measurements is 1467 inconclusive. 1469 The problem above is formulated to match the Sequential Probability 1470 Ratio Test (SPRT) [StatQC]. Note that as originally framed the 1471 events under consideration were all manufacturing defects. In 1472 networking, ECN CE marks and lost packets are not defects but 1473 signals, indicating that the transport protocol should slow down. 1475 The Sequential Probability Ratio Test also starts with a pair of 1476 hypothesis specified as above: 1478 H0: p0 = one defect in target_run_length 1479 H1: p1 = one defect in target_run_length/4 1481 As packets are sent and measurements collected, the tester evaluates 1482 the cumulative defect count against two boundaries representing H0 1483 Acceptance or Rejection (and acceptance of H1): 1485 Acceptance line: Xa = -h1 + s*n 1486 Rejection line: Xr = h2 + s*n 1488 where n increases linearly for each packet sent and 1490 h1 = { log((1-alpha)/beta) }/k 1491 h2 = { log((1-beta)/alpha) }/k 1492 k = log{ (p1(1-p0)) / (p0(1-p1)) } 1493 s = [ log{ (1-p0)/(1-p1) } ]/k 1495 for p0 and p1 as defined in the null and alternative Hypotheses 1496 statements above, and alpha and beta as the Type I and Type II 1497 errors. 1499 The SPRT specifies simple stopping rules: 1501 o Xa < defect_count(n) < Xb: continue testing 1502 o defect_count(n) <= Xa: Accept H0 1503 o defect_count(n) >= Xb: Accept H1 1505 The calculations above are implemented in the R-tool for Statistical 1506 Analysis [Rtool] , in the add-on package for Cross-Validation via 1507 Sequential Testing (CVST) [CVST]. 1509 Using the equations above, we can calculate the minimum number of 1510 packets (n) needed to accept H0 when x defects are observed. For 1511 example, when x = 0: 1513 Xa = 0 = -h1 + s*n 1514 and n = h1 / s 1516 7.3. Reordering Tolerance 1518 All tests MUST be instrumented for packet level reordering [RFC4737]. 1519 However, there is no consensus for how much reordering should be 1520 acceptable. Over the last two decades the general trend has been to 1521 make protocols and applications more tolerant to reordering (see for 1522 example [RFC4015]), in response to the gradual increase in reordering 1523 in the network. This increase has been due to the deployment of 1524 technologies such as multi threaded routing lookups and Equal Cost 1525 MultiPath (ECMP) routing. These techniques increase parallelism in 1526 network and are critical to enabling overall Internet growth to 1527 exceed Moore's Law. 1529 Note that transport retransmission strategies can trade off 1530 reordering tolerance vs how quickly they can repair losses vs 1531 overhead from spurious retransmissions. In advance of new 1532 retransmission strategies we propose the following strawman: 1533 Transport protocols should be able to adapt to reordering as long as 1534 the reordering extent is not more than the maximum of one quarter 1535 window or 1 mS, whichever is larger. Within this limit on reorder 1536 extent, there should be no bound on reordering density. 1538 By implication, recording which is less than these bounds should not 1539 be treated as a network impairment. However [RFC4737] still applies: 1540 reordering should be instrumented and the maximum reordering that can 1541 be properly characterized by the test (because of the bound on 1542 history buffers) should be recorded with the measurement results. 1544 Reordering tolerance and diagnostic limitations, such as the size of 1545 the history buffer used to diagnose packets that are way out-of- 1546 order, MUST be specified in a FSTIDS. 1548 8. IP Diagnostic Tests 1550 The IP diagnostic tests below are organized by traffic pattern: basic 1551 data rate and packet transfer statistics, standing queues, slowstart 1552 bursts, and sender rate bursts. We also introduce some combined 1553 tests which are more efficient when networks are expected to pass, 1554 but conflate diagnostic signatures when they fail. 1556 There are a number of test details which are not fully defined here. 1557 They must be fully specified in a FS-TIDS. From a standardization 1558 perspective, this lack of specificity will weaken this version of 1559 Model Based Metrics, however it is anticipated that this weakness is 1560 than offset by the extent to which MBM suppresses the problems caused 1561 by using transport protocols for measurement. e.g. non-specific MBM 1562 metrics are likely to have better repeatability than many existing 1563 BTC like metrics. Once we have good field experience, the missing 1564 details can be fully specified. 1566 8.1. Basic Data Rate and Packet Transfer Tests 1568 We propose several versions of the basic data rate and packet 1569 transfer statistics test. All measure the number of packets 1570 delivered between losses or ECN Congestion Experienced (CE) marks, 1571 using a data stream that is rate controlled at approximately the 1572 target_data_rate. 1574 The tests below differ in how the data rate is controlled. The data 1575 can be paced on a timer, or window controlled (and self clocked). 1576 The first two tests implicitly confirm that sub_path has sufficient 1577 raw capacity to carry the target_data_rate. They are recommended for 1578 relatively infrequent testing, such as an installation or periodic 1579 auditing process. The third, background packet transfer statistics, 1580 is a low rate test designed for ongoing monitoring for changes in 1581 subpath quality. 1583 All rely on the data receiver accumulating packet transfer statistics 1584 as described in Section 7.2 to score the outcome: 1586 Pass: it is statistically significant that the observed interval 1587 between losses or ECN CE marks is larger than the target_run_length. 1589 Fail: it is statistically significant that the observed interval 1590 between losses or ECN CE marks is smaller than the target_run_length. 1592 A test is considered to be inconclusive if it failed to generate the 1593 data rate as specified below, meet the qualifications defined in 1594 Section 5.4 or neither run length statistical hypothesis was 1595 confirmed in the allotted test duration. 1597 8.1.1. Delivery Statistics at Paced Full Data Rate 1599 Confirm that the observed run length is at least the 1600 target_run_length while relying on timer to send data at the 1601 target_rate using the procedure described in in Section 6.1 with a 1602 burst size of 1 (single packets) or 2 (packet pairs). 1604 The test is considered to be inconclusive if the packet transmission 1605 can not be accurately controlled for any reason. 1607 RFC 6673 [RFC6673] is appropriate for measuring packet transfer 1608 statistics at full data rate. 1610 8.1.2. Delivery Statistics at Full Data Windowed Rate 1612 Confirm that the observed run length is at least the 1613 target_run_length while sending at an average rate approximately 1614 equal to the target_data_rate, by controlling (or clamping) the 1615 window size of a conventional transport protocol to test_window. 1617 Since losses and ECN CE marks cause transport protocols to reduce 1618 their data rates, this test is expected to be less precise about 1619 controlling its data rate. It should not be considered inconclusive 1620 as long as at least some of the round trips reached the full 1621 target_data_rate without incurring losses or ECN CE marks. To pass 1622 this test the network must deliver target_window_size packets in 1623 target_RTT time without any losses or ECN CE marks at least once per 1624 two target_window_size round trips, in addition to meeting the run 1625 length statistical test. 1627 8.1.3. Background Packet Transfer Statistics Tests 1629 The background run length is a low rate version of the target target 1630 rate test above, designed for ongoing lightweight monitoring for 1631 changes in the observed subpath run length without disrupting users. 1632 It should be used in conjunction with one of the above full rate 1633 tests because it does not confirm that the subpath can support raw 1634 data rate. 1636 RFC 6673 [RFC6673] is appropriate for measuring background packet 1637 transfer statistics. 1639 8.2. Standing Queue Tests 1641 These engineering tests confirm that the bottleneck is well behaved 1642 across the onset of packet loss, which typically follows after the 1643 onset of queuing. Well behaved generally means lossless for 1644 transient queues, but once the queue has been sustained for a 1645 sufficient period of time (or reaches a sufficient queue depth) there 1646 should be a small number of losses or ECN CE marks to signal to the 1647 transport protocol that it should reduce its window. Losses that are 1648 too early can prevent the transport from averaging at the 1649 target_data_rate. Losses that are too late indicate that the queue 1650 might be subject to bufferbloat [wikiBloat] and inflict excess 1651 queuing delays on all flows sharing the bottleneck queue. Excess 1652 losses (more than half of the window) at the onset of congestion make 1653 loss recovery problematic for the transport protocol. Non-linear, 1654 erratic or excessive RTT increases suggest poor interactions between 1655 the channel acquisition algorithms and the transport self clock. All 1656 of the tests in this section use the same basic scanning algorithm, 1657 described here, but score the link or subpath on the basis of how 1658 well it avoids each of these problems. 1660 Some network technologies rely on virtual queues or other techniques 1661 to meter traffic without adding any queuing delay, in which case the 1662 data rate will vary with the window size all the way up to the onset 1663 of load induced packet loss or ECN CE marks. For theses 1664 technologies, the discussion of queuing in Section 6.3 does not 1665 apply, but it is still necessary to confirm that the onset of losses 1666 or ECN CE marks be at an appropriate point and progressive. If the 1667 network bottleneck does not introduce significant queuing delay, 1668 modify the procedure described in Section 6.3 to start scan at a 1669 window equal to or slightly smaller than the test_window. 1671 Use the procedure in Section 6.3 to sweep the window across the onset 1672 of queuing and the onset of loss. The tests below all assume that 1673 the scan emulates standard additive increase and delayed ACK by 1674 incrementing the window by one packet for every 2*target_window_size 1675 packets delivered. A scan can typically be divided into three 1676 regions: below the onset of queuing, a standing queue, and at or 1677 beyond the onset of loss. 1679 Below the onset of queuing the RTT is typically fairly constant, and 1680 the data rate varies in proportion to the window size. Once the data 1681 rate reaches the subpath IP rate, the data rate becomes fairly 1682 constant, and the RTT increases in proportion to the increase in 1683 window size. The precise transition across the start of queuing can 1684 be identified by the maximum network power, defined to be the ratio 1685 data rate over the RTT. The network power can be computed at each 1686 window size, and the window with the maximum is taken as the start of 1687 the queuing region. 1689 If there is random background loss (e.g. bit errors, etc), precise 1690 determination of the onset of queue induced packet loss may require 1691 multiple scans. Above the onset of queuing loss, all transport 1692 protocols are expected to experience periodic losses determined by 1693 the interaction between the congestion control and AQM algorithms. 1694 For standard congestion control algorithms the periodic losses are 1695 likely to be relatively widely spaced and the details are typically 1696 dominated by the behavior of the transport protocol itself. For the 1697 stiffened transport protocols case (with non-standard, aggressive 1698 congestion control algorithms) the details of periodic losses will be 1699 dominated by how the window increase function responds to loss. 1701 8.2.1. Congestion Avoidance 1703 A subpath passes the congestion avoidance standing queue test if more 1704 than target_run_length packets are delivered between the onset of 1705 queuing (as determined by the window with the maximum network power 1706 as described above) and the first loss or ECN CE mark. If this test 1707 is implemented using a standard congestion control algorithm with a 1708 clamp, it can be performed in situ in the production internet as a 1709 capacity test. For an example of such a test see [Pathdiag]. 1711 For technologies that do not have conventional queues, use the 1712 test_window in place of the onset of queuing. i.e. A subpath passes 1713 the congestion avoidance standing queue test if more than 1714 target_run_length packets are delivered between start of the scan at 1715 test_window and the first loss or ECN CE mark. 1717 8.2.2. Bufferbloat 1719 This test confirms that there is some mechanism to limit buffer 1720 occupancy (e.g. that prevents bufferbloat). Note that this is not 1721 strictly a requirement for single stream bulk transport capacity, 1722 however if there is no mechanism to limit buffer queue occupancy then 1723 a single stream with sufficient data to deliver is likely to cause 1724 the problems described in [RFC7567], and [wikiBloat]. This may cause 1725 only minor symptoms for the dominant flow, but has the potential to 1726 make the subpath unusable for other flows and applications. 1728 Pass if the onset of loss occurs before a standing queue has 1729 introduced more delay than than twice target_RTT, or other well 1730 defined and specified limit. Note that there is not yet a model for 1731 how much standing queue is acceptable. The factor of two chosen here 1732 reflects a rule of thumb. In conjunction with the previous test, 1733 this test implies that the first loss should occur at a queuing delay 1734 which is between one and two times the target_RTT. 1736 Specified RTT limits that are larger than twice the target_RTT must 1737 be fully justified in the FS-TIDS. 1739 8.2.3. Non excessive loss 1741 This test confirms that the onset of loss is not excessive. Pass if 1742 losses are equal or less than the increase in the cross traffic plus 1743 the test stream window increase since the previous RTT. This could 1744 be restated as non-decreasing total throughput of the subpath at the 1745 onset of loss. (Note that when there is a transient drop in subpath 1746 throughput and there is not already a standing queue, a subpath that 1747 passes other queue tests in this document will have sufficient queue 1748 space to hold one full RTT worth of data). 1750 Note that token bucket policers will not pass this test, which is as 1751 intended. TCP often stumbles badly if more than a small fraction of 1752 the packets are dropped in one RTT. Many TCP implementations will 1753 require a timeout and slowstart to recover their self clock. Even if 1754 they can recover from the massive losses the sudden change in 1755 available capacity at the bottleneck wastes serving and front path 1756 capacity until TCP can adapt to the new rate [Policing]. 1758 8.2.4. Duplex Self Interference 1760 This engineering test confirms a bound on the interactions between 1761 the forward data path and the ACK return path. 1763 Some historical half duplex technologies had the property that each 1764 direction held the channel until it completely drained its queue. 1765 When a self clocked transport protocol, such as TCP, has data and 1766 ACKs passing in opposite directions through such a link, the behavior 1767 often reverts to stop-and-wait. Each additional packet added to the 1768 window raises the observed RTT by two packet times, once as it passes 1769 through the data path, and once for the additional delay incurred by 1770 the ACK waiting on the return path. 1772 The duplex self interference test fails if the RTT rises by more than 1773 a fixed bound above the expected queugit staing time computed from 1774 trom the excess window divided by the subpath IP Capacity. This 1775 bound must be smaller than target_RTT/2 to avoid reverting to stop 1776 and wait behavior. (e.g. Data packets and ACKs both have to be 1777 released at least twice per RTT.) 1779 8.3. Slowstart tests 1781 These tests mimic slowstart: data is sent at twice the effective 1782 bottleneck rate to exercise the queue at the dominant bottleneck. 1784 8.3.1. Full Window slowstart test 1786 This is a capacity test to confirm that slowstart is not likely to 1787 exit prematurely. Send slowstart bursts that are target_window_size 1788 total packets. 1790 Accumulate packet transfer statistics as described in Section 7.2 to 1791 score the outcome. Pass if it is statistically significant that the 1792 observed number of good packets delivered between losses or ECN CE 1793 marks is larger than the target_run_length. Fail if it is 1794 statistically significant that the observed interval between losses 1795 or ECN CE marks is smaller than the target_run_length. 1797 It is deemed inconclusive if the elapsed time to send the data burst 1798 is not less than half of the time to receive the ACKs. (i.e. It is 1799 acceptable to send data too fast, but sending it slower than twice 1800 the actual bottleneck rate as indicated by the ACKs is deemed 1801 inconclusive). The headway for the slowstart bursts should be the 1802 target_RTT. 1804 Note that these are the same parameters as the Sender Full Window 1805 burst test, except the burst rate is at slowstart rate, rather than 1806 sender interface rate. 1808 8.3.2. Slowstart AQM test 1810 Do a continuous slowstart (send data continuously at twice the 1811 implied IP bottleneck capacity), until the first loss, stop, allow 1812 the network to drain and repeat, gathering statistics on how many 1813 packets were delivered before the loss, the pattern of losses, 1814 maximum observed RTT and window size. Justify the results. There is 1815 not currently sufficient theory justifying requiring any particular 1816 result, however design decisions that affect the outcome of this 1817 tests also affect how the network balances between long and short 1818 flows (the "mice vs elephants" problem). The queue sojourn time for 1819 the first packet delivered after the first loss should be at least 1820 one half of the target_RTT. 1822 This is an engineering test: It should be performed on a quiescent 1823 network or testbed, since cross traffic has the potential to change 1824 the results in ill defined ways. 1826 8.4. Sender Rate Burst tests 1828 These tests determine how well the network can deliver bursts sent at 1829 sender's interface rate. Note that this test most heavily exercises 1830 the front path, and is likely to include infrastructure may be out of 1831 scope for an access ISP, even though the bursts might be caused by 1832 ACK compression, thinning or channel arbitration in the access ISP. 1833 See Appendix B. 1835 Also, there are a several details about sender interface rate bursts 1836 that are not fully defined here. These details, such as the assumed 1837 sender interface rate, should be explicitly stated is a FS-TIDS. 1839 Current standards permit TCP to send full window bursts following an 1840 application pause. (Congestion Window Validation [RFC2861] and 1841 updates to support Rate-Limited Traffic [RFC7661], are not required). 1842 Since full window bursts are consistent with standard behavior, it is 1843 desirable that the network be able to deliver such bursts, otherwise 1844 application pauses will cause unwarranted losses. Note that the AIMD 1845 sawtooth requires a peak window that is twice target_window_size, so 1846 the worst case burst may be 2*target_window_size. 1848 It is also understood in the application and serving community that 1849 interface rate bursts have a cost to the network that has to be 1850 balanced against other costs in the servers themselves. For example 1851 TCP Segmentation Offload (TSO) reduces server CPU in exchange for 1852 larger network bursts, which increase the stress on network buffer 1853 memory. Some newer TCP implementations can pace traffic at scale 1854 [TSO_pacing][TSO_fq_pacing]. It remains to be determined if and how 1855 quickly these changes will be deployed. 1857 There is not yet theory to unify these costs or to provide a 1858 framework for trying to optimize global efficiency. We do not yet 1859 have a model for how much the network should tolerate server rate 1860 bursts. Some bursts must be tolerated by the network, but it is 1861 probably unreasonable to expect the network to be able to efficiently 1862 deliver all data as a series of bursts. 1864 For this reason, this is the only test for which we encourage 1865 derating. A TIDS could include a table of pairs of derating 1866 parameters: burst sizes and how much each burst size is permitted to 1867 reduce the run length, relative to to the target_run_length. 1869 8.5. Combined and Implicit Tests 1871 Combined tests efficiently confirm multiple network properties in a 1872 single test, possibly as a side effect of normal content delivery. 1873 They require less measurement traffic than other testing strategies 1874 at the cost of conflating diagnostic signatures when they fail. 1875 These are by far the most efficient for monitoring networks that are 1876 nominally expected to pass all tests. 1878 8.5.1. Sustained Bursts Test 1880 The sustained burst test implements a combined worst case version of 1881 all of the capacity tests above. It is simply: 1883 Send target_window_size bursts of packets at server interface rate 1884 with target_RTT burst headway (burst start to next burst start). 1885 Verify that the observed packet transfer statistics meets the 1886 target_run_length. 1888 Key observations: 1890 o The subpath under test is expected to go idle for some fraction of 1891 the time, determined by the difference between the time to drain 1892 the queue at the subpath_IP_capacity, and the target_RTT. If the 1893 queue does not drain completely it may be an indication that the 1894 the subpath has insufficient IP capacity or that there is some 1895 other problem with the test (e.g. inconclusive). 1896 o The burst sensitivity can be derated by sending smaller bursts 1897 more frequently. E.g. send target_window_size*derate packet 1898 bursts every target_RTT*derate, where "derate" is less than one. 1899 o When not derated, this test is the most strenuous capacity test. 1901 o A subpath that passes this test is likely to be able to sustain 1902 higher rates (close to subpath_IP_capacity) for paths with RTTs 1903 significantly smaller than the target_RTT. 1904 o This test can be implemented with instrumented TCP [RFC4898], 1905 using a specialized measurement application at one end [MBMSource] 1906 and a minimal service at the other end [RFC0863] [RFC0864]. 1907 o This test is efficient to implement, since it does not require 1908 per-packet timers, and can make use of TSO in modern NIC hardware. 1909 o If a subpath is known to pass the Standing Queue engineering tests 1910 (particularly that it has a progressive onset of loss at an 1911 appropriate queue depth), then the Sustained Burst Test is 1912 sufficient to assure that the subpath under test will not impair 1913 Bulk Transport Capacity at the target performance under all 1914 conditions. See Section 8.2 for a discussion of the standing 1915 queue tests. 1917 Note that this test is clearly independent of the subpath RTT, or 1918 other details of the measurement infrastructure, as long as the 1919 measurement infrastructure can accurately and reliably deliver the 1920 required bursts to the subpath under test. 1922 8.5.2. Passive Measurements 1924 Any non-throughput maximizing application, such as fixed rate 1925 streaming media, can be used to implement passive or hybrid (defined 1926 in [RFC7799]) versions of Model Based Metrics with some additional 1927 instrumentation and possibly a traffic shaper or other controls in 1928 the servers. The essential requirement is that the data transmission 1929 be constrained such that even with arbitrary application pauses and 1930 bursts, the data rate and burst sizes stay within the envelope 1931 defined by the individual tests described above. 1933 If the application's serving data rate can be constrained to be less 1934 than or equal to the target_data_rate and the serving_RTT (the RTT 1935 between the sender and client) is less than the target_RTT, this 1936 constraint is most easily implemented by clamping the transport 1937 window size to serving_window_clamp, set to the test_window, computed 1938 for the actual serving path. 1940 Under the above constraints the serving_window_clamp will limit the 1941 both the serving data rate and burst sizes to be no larger than the 1942 procedures in Section 8.1.2 and Section 8.4 or Section 8.5.1. Since 1943 the serving RTT is smaller than the target_RTT, the worst case bursts 1944 that might be generated under these conditions will be smaller than 1945 called for by Section 8.4 and the sender rate burst sizes are 1946 implicitly derated by the serving_window_clamp divided by the 1947 target_window_size at the very least. (Depending on the application 1948 behavior, the data might be significantly smoother than specified by 1949 any of the burst tests.) 1951 In an alternative implementation the data rate and bursts might be 1952 explicitly controlled by a programmable traffic shaper or pacing at 1953 the sender. This would provide better control over transmissions but 1954 is more complicated to implement, although the required technology is 1955 available [TSO_pacing][TSO_fq_pacing]. 1957 Note that these techniques can be applied to any content delivery 1958 that can operated at a constrained data rate to inhibit TCP 1959 equilibrium behavior. 1961 Furthermore note that Dynamic Adaptive Streaming over HTTP (DASH) is 1962 generally in conflict with passive Model Based Metrics measurement, 1963 because it is a rate maximizing protocol. It can still meet the 1964 requirement here if the rate can be capped, for example by knowing a 1965 priori the maximum rate needed to deliver a particular piece of 1966 content. 1968 9. An Example 1970 In this section we illustrate a TIDS designed to confirm that an 1971 access ISP can reliably deliver HD video from multiple content 1972 providers to all of their customers. With modern codecs, minimal HD 1973 video (720p) generally fits in 2.5 Mb/s. Due to their geographical 1974 size, network topology and modem characteristics the ISP determines 1975 that most content is within a 50 mS RTT of their users (This example 1976 RTT is a sufficient to cover the propagation delay to continental 1977 Europe or either US coast with low delay modems or somewhat smaller 1978 geographical regions if the modems require additional delay to 1979 implement advanced compression and error recovery). 1981 2.5 Mb/s over a 50 ms path 1983 +----------------------+-------+---------+ 1984 | End-to-End Parameter | value | units | 1985 +----------------------+-------+---------+ 1986 | target_rate | 2.5 | Mb/s | 1987 | target_RTT | 50 | ms | 1988 | target_MTU | 1500 | bytes | 1989 | header_overhead | 64 | bytes | 1990 | | | | 1991 | target_window_size | 11 | packets | 1992 | target_run_length | 363 | packets | 1993 +----------------------+-------+---------+ 1995 Table 1 1997 Table 1 shows the default TCP model with no derating, and as such is 1998 quite conservative. The simplest TIDS would be to use the sustained 1999 burst test, described in Section 8.5.1. Such a test would send 11 2000 packet bursts every 50mS, and confirming that there was no more than 2001 1 packet loss per 33 bursts (363 total packets in 1.650 seconds). 2003 Since this number represents is the entire end-to-end loss budget, 2004 independent subpath tests could be implemented by apportioning the 2005 packet loss ratio across subpaths. For example 50% of the losses 2006 might be allocated to the access or last mile link to the user, 40% 2007 to the network interconnections with other ISPs and 1% to each 2008 internal hop (assuming no more than 10 internal hops). Then all of 2009 the subpaths can be tested independently, and the spatial composition 2010 of passing subpaths would be expected to be within the end-to-end 2011 loss budget. 2013 9.1. Observations about applicability 2015 Guidance on deploying and using MBM belong in a future document. 2016 However this example illustrates some the issues that may need to be 2017 considered. 2019 Note that another ISP, with different geographical coverage, topology 2020 or modem technology may need to assume a different target_RTT, and as 2021 a consequence different target_window_size and target_run_length, 2022 even for the same target_data rate. One of the implications of this 2023 is that infrastructure shared by multiple ISPs, such as inter- 2024 exchange points (IXPs) and other interconnects may need to be 2025 evaluated on the basis of the most stringent target_window_size and 2026 target_run_length of any participating ISP. One way to do this might 2027 be to choose target parameters for evaluating such shared 2028 infrastructure on the basis of a hypothetical reference path that 2029 does not necessarily match any actual paths. 2031 Testing interconnects has generally been problematic: conventional 2032 performance tests run between measurement points adjacent to either 2033 side of the interconnect are not generally useful. Unconstrained TCP 2034 tests, such as iPerf [iPerf] are usually overly aggressive due to the 2035 small RTT (often less than 1 mS). With a short RTT these tools are 2036 likely to report inflated data rates because on a short RTT these 2037 tools can tolerate very high packet loss ratios and can push other 2038 cross traffic off of the network. As a consequence these 2039 measurements are useless for predicting actual user performance over 2040 longer paths, and may themselves be quite disruptive. Model Based 2041 Metrics solves this problem. The interconnect can be evaluated with 2042 the same TIDS as other subpaths. Continuing our example, if the 2043 interconnect is apportioned 40% of the losses, 11 packet bursts sent 2044 every 50mS should have fewer than one loss per 82 bursts (902 2045 packets). 2047 10. Validation 2049 Since some aspects of the models are likely to be too conservative, 2050 Section 5.2 permits alternate protocol models and Section 5.3 permits 2051 test parameter derating. If either of these techniques are used, we 2052 require demonstrations that such a TIDS can robustly detect subpaths 2053 that will prevent authentic applications using state-of-the-art 2054 protocol implementations from meeting the specified Target Transport 2055 Performance. This correctness criteria is potentially difficult to 2056 prove, because it implicitly requires validating a TIDS against all 2057 possible paths and subpaths. The procedures described here are still 2058 experimental. 2060 We suggest two approaches, both of which should be applied: first, 2061 publish a fully open description of the TIDS, including what 2062 assumptions were used and and how it was derived, such that the 2063 research community can evaluate the design decisions, test them and 2064 comment on their applicability; and second, demonstrate that an 2065 applications running over an infinitesimally passing testbed do meet 2066 the performance targets. 2068 An infinitesimally passing testbed resembles a epsilon-delta proof in 2069 calculus. Construct a test network such that all of the individual 2070 tests of the TIDS pass by only small (infinitesimal) margins, and 2071 demonstrate that a variety of authentic applications running over 2072 real TCP implementations (or other protocol as appropriate) meets the 2073 Target Transport Performance over such a network. The workloads 2074 should include multiple types of streaming media and transaction 2075 oriented short flows (e.g. synthetic web traffic). 2077 For example, for the HD streaming video TIDS described in Section 9, 2078 the IP capacity should be exactly the header_overhead above 2.5 Mb/s, 2079 the per packet random background loss ratio should be 1/363, for a 2080 run length of 363 packets, the bottleneck queue should be 11 packets 2081 and the front path should have just enough buffering to withstand 11 2082 packet interface rate bursts. We want every one of the TIDS tests to 2083 fail if we slightly increase the relevant test parameter, so for 2084 example sending a 12 packet bursts should cause excess (possibly 2085 deterministic) packet drops at the dominant queue at the bottleneck. 2086 On this infinitesimally passing network it should be possible for a 2087 real application using a stock TCP implementation in the vendor's 2088 default configuration to attain 2.5 Mb/s over an 50 mS path. 2090 The most difficult part of setting up such a testbed is arranging for 2091 it to infinitesimally pass the individual tests. Two approaches are 2092 suggested: constraining the network devices not to use all available 2093 resources (e.g. by limiting available buffer space or data rate); and 2094 pre-loading subpaths with cross traffic. Note that is it important 2095 that a single environment be constructed which infinitesimally passes 2096 all tests at the same time, otherwise there is a chance that TCP can 2097 exploit extra latitude in some parameters (such as data rate) to 2098 partially compensate for constraints in other parameters (queue 2099 space, or vice-versa). 2101 To the extent that a TIDS is used to inform public dialog it should 2102 be fully publicly documented, including the details of the tests, 2103 what assumptions were used and how it was derived. All of the 2104 details of the validation experiment should also be published with 2105 sufficient detail for the experiments to be replicated by other 2106 researchers. All components should either be open source of fully 2107 described proprietary implementations that are available to the 2108 research community. 2110 11. Security Considerations 2112 Measurement is often used to inform business and policy decisions, 2113 and as a consequence is potentially subject to manipulation. Model 2114 Based Metrics are expected to be a huge step forward because 2115 equivalent measurements can be performed from multiple vantage 2116 points, such that performance claims can be independently validated 2117 by multiple parties. 2119 Much of the acrimony in the Net Neutrality debate is due to the 2120 historical lack of any effective vantage independent tools to 2121 characterize network performance. Traditional methods for measuring 2122 Bulk Transport Capacity are sensitive to RTT and as a consequence 2123 often yield very different results when run local to an ISP or 2124 interconnect and when run over a customer's complete path. Neither 2125 the ISP nor customer can repeat the others measurements, leading to 2126 high levels of distrust and acrimony. Model Based Metrics are 2127 expected to greatly improve this situation. 2129 This document only describes a framework for designing Fully 2130 Specified Targeted IP Diagnostic Suite. Each FS-TIDS MUST include 2131 its own security section. 2133 12. Acknowledgments 2135 Ganga Maguluri suggested the statistical test for measuring loss 2136 probability in the target run length. Alex Gilgur for helping with 2137 the statistics. 2139 Meredith Whittaker for improving the clarity of the communications. 2141 Ruediger Geib provided feedback which greatly improved the document. 2143 This work was inspired by Measurement Lab: open tools running on an 2144 open platform, using open tools to collect open data. See 2145 http://www.measurementlab.net/ 2147 13. IANA Considerations 2149 This document has no actions for IANA. 2151 14. References 2153 14.1. Normative References 2155 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 2156 Requirement Levels", BCP 14, RFC 2119, March 1997. 2158 14.2. Informative References 2160 [RFC0863] Postel, J., "Discard Protocol", STD 21, RFC 863, May 1983. 2162 [RFC0864] Postel, J., "Character Generator Protocol", STD 22, 2163 RFC 864, May 1983. 2165 [RFC2330] Paxson, V., Almes, G., Mahdavi, J., and M. Mathis, 2166 "Framework for IP Performance Metrics", RFC 2330, May 2167 1998. 2169 [RFC2861] Handley, M., Padhye, J., and S. Floyd, "TCP Congestion 2170 Window Validation", RFC 2861, June 2000. 2172 [RFC3148] Mathis, M. and M. Allman, "A Framework for Defining 2173 Empirical Bulk Transfer Capacity Metrics", RFC 3148, July 2174 2001. 2176 [RFC3465] Allman, M., "TCP Congestion Control with Appropriate Byte 2177 Counting (ABC)", RFC 3465, February 2003. 2179 [RFC4015] Ludwig, R. and A. Gurtov, "The Eifel Response Algorithm 2180 for TCP", RFC 4015, February 2005. 2182 [RFC4737] Morton, A., Ciavattone, L., Ramachandran, G., Shalunov, 2183 S., and J. Perser, "Packet Reordering Metrics", RFC 4737, 2184 November 2006. 2186 [RFC4898] Mathis, M., Heffner, J., and R. Raghunarayan, "TCP 2187 Extended Statistics MIB", RFC 4898, May 2007. 2189 [RFC5136] Chimento, P. and J. Ishac, "Defining Network Capacity", 2190 RFC 5136, February 2008. 2192 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion 2193 Control", RFC 5681, September 2009. 2195 [RFC5835] Morton, A. and S. Van den Berghe, "Framework for Metric 2196 Composition", RFC 5835, April 2010. 2198 [RFC6049] Morton, A. and E. Stephan, "Spatial Composition of 2199 Metrics", RFC 6049, January 2011. 2201 [RFC6576] Geib, R., Ed., Morton, A., Fardid, R., and A. Steinmitz, 2202 "IP Performance Metrics (IPPM) Standard Advancement 2203 Testing", BCP 176, RFC 6576, DOI 10.17487/RFC6576, March 2204 2012, . 2206 [RFC6673] Morton, A., "Round-Trip Packet Loss Metrics", RFC 6673, 2207 August 2012. 2209 [RFC6928] Chu, J., Dukkipati, N., Cheng, Y., and M. Mathis, 2210 "Increasing TCP's Initial Window", RFC 6928, 2211 DOI 10.17487/RFC6928, April 2013, 2212 . 2214 [RFC7312] Fabini, J. and A. Morton, "Advanced Stream and Sampling 2215 Framework for IP Performance Metrics (IPPM)", RFC 7312, 2216 August 2014. 2218 [RFC7398] Bagnulo, M., Burbridge, T., Crawford, S., Eardley, P., and 2219 A. Morton, "A Reference Path and Measurement Points for 2220 Large-Scale Measurement of Broadband Performance", 2221 RFC 7398, February 2015. 2223 [RFC7567] Baker, F., Ed. and G. Fairhurst, Ed., "IETF 2224 Recommendations Regarding Active Queue Management", 2225 BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015, 2226 . 2228 [RFC7661] Fairhurst, G., Sathiaseelan, A., and R. Secchi, "Updating 2229 TCP to Support Rate-Limited Traffic", RFC 7661, 2230 DOI 10.17487/RFC7661, October 2015, 2231 . 2233 [RFC7680] Almes, G., Kalidindi, S., Zekauskas, M., and A. Morton, 2234 Ed., "A One-Way Loss Metric for IP Performance Metrics 2235 (IPPM)", STD 82, RFC 7680, DOI 10.17487/RFC7680, January 2236 2016, . 2238 [RFC7799] Morton, A., "Active and Passive Metrics and Methods (with 2239 Hybrid Types In-Between)", RFC 7799, DOI 10.17487/RFC7799, 2240 May 2016, . 2242 [MSMO97] Mathis, M., Semke, J., Mahdavi, J., and T. Ott, "The 2243 Macroscopic Behavior of the TCP Congestion Avoidance 2244 Algorithm", Computer Communications Review volume 27, 2245 number3, July 1997. 2247 [WPING] Mathis, M., "Windowed Ping: An IP Level Performance 2248 Diagnostic", INET 94, June 1994. 2250 [mpingSource] 2251 Fan, X., Mathis, M., and D. Hamon, "Git Repository for 2252 mping: An IP Level Performance Diagnostic", Sept 2013, 2253 . 2255 [MBMSource] 2256 Hamon, D., Stuart, S., and H. Chen, "Git Repository for 2257 Model Based Metrics", Sept 2013, . 2260 [Pathdiag] 2261 Mathis, M., Heffner, J., O'Neil, P., and P. Siemsen, 2262 "Pathdiag: Automated TCP Diagnosis", Passive and Active 2263 Measurement , June 2008. 2265 [iPerf] Wikipedia Contributors, , "iPerf", Wikipedia, The Free 2266 Encyclopedia , cited March 2015, 2267 . 2270 [StatQC] Montgomery, D., "Introduction to Statistical Quality 2271 Control - 2nd ed.", ISBN 0-471-51988-X, 1990. 2273 [Rtool] R Development Core Team, , "R: A language and environment 2274 for statistical computing. R Foundation for Statistical 2275 Computing, Vienna, Austria. ISBN 3-900051-07-0, URL 2276 http://www.R-project.org/", , 2011. 2278 [CVST] Krueger, T. and M. Braun, "R package: Fast Cross- 2279 Validation via Sequential Testing", version 0.1, 11 2012. 2281 [AFD] Pan, R., Breslau, L., Prabhakar, B., and S. Shenker, 2282 "Approximate fairness through differential dropping", 2283 SIGCOMM Comput. Commun. Rev. 33, 2, April 2003. 2285 [wikiBloat] 2286 Wikipedia, , "Bufferbloat", http://en.wikipedia.org/ 2287 w/ index.php?title=Bufferbloat&oldid=608805474, March 2288 2015. 2290 [CCscaling] 2291 Fernando, F., Doyle, J., and S. Steven, "Scalable laws for 2292 stable network congestion control", Proceedings of 2293 Conference on Decision and 2294 Control, http://www.ee.ucla.edu/~paganini, December 2001. 2296 [TSO_pacing] 2297 Corbet, J., "TSO sizing and the FQ scheduler", 2298 LWN.net https://lwn.net/Articles/564978/, Aug 2013. 2300 [TSO_fq_pacing] 2301 Dumazet, E. and Y. Chen, "TSO, fair queuing, pacing: 2302 three's a charm", Proceedings of IETF 88, TCPM WG 2303 https://www.ietf.org/proceedings/88/slides/slides-88-tcpm- 2304 9.pdf, Nov 2013. 2306 [Policing] 2307 Flach, T., Papageorge, P., Terzis, A., Pedrosa, L., Cheng, 2308 Y., Karim, T., Katz-Bassett, E., and R. Govindan, "An 2309 Internet-Wide Analysis of Traffic Policing", ACM SIGCOMM , 2310 August 2016. 2312 Appendix A. Model Derivations 2314 The reference target_run_length described in Section 5.2 is based on 2315 very conservative assumptions: that all excess data in flight 2316 (window) above the target_window_size contributes to a standing queue 2317 that raises the RTT, and that classic Reno congestion control with 2318 delayed ACKs are in effect. In this section we provide two 2319 alternative calculations using different assumptions. 2321 It may seem out of place to allow such latitude in a measurement 2322 method, but this section provides offsetting requirements. 2324 The estimates provided by these models make the most sense if network 2325 performance is viewed logarithmically. In the operational Internet, 2326 data rates span more than 8 orders of magnitude, RTT spans more than 2327 3 orders of magnitude, and packet loss ratio spans at least 8 orders 2328 of magnitude if not more. When viewed logarithmically (as in 2329 decibels), these correspond to 80 dB of dynamic range. On an 80 dB 2330 scale, a 3 dB error is less than 4% of the scale, even though it 2331 represents a factor of 2 in untransformed parameter. 2333 This document gives a lot of latitude for calculating 2334 target_run_length, however people designing a TIDS should consider 2335 the effect of their choices on the ongoing tussle about the relevance 2336 of "TCP friendliness" as an appropriate model for Internet capacity 2337 allocation. Choosing a target_run_length that is substantially 2338 smaller than the reference target_run_length specified in Section 5.2 2339 strengthens the argument that it may be appropriate to abandon "TCP 2340 friendliness" as the Internet fairness model. This gives developers 2341 incentive and permission to develop even more aggressive applications 2342 and protocols, for example by increasing the number of connections 2343 that they open concurrently. 2345 A.1. Queueless Reno 2347 In Section 5.2 models were derived based on the assumption that the 2348 subpath IP rate matches the target rate plus overhead, such that the 2349 excess window needed for the AIMD sawtooth causes a fluctuating queue 2350 at the bottleneck. 2352 An alternate situation would be a bottleneck where there is no 2353 significant queue and losses are caused by some mechanism that does 2354 not involve extra delay, for example by the use of a virtual queue as 2355 done in Approximate Fair Dropping [AFD]. A flow controlled by such a 2356 bottleneck would have a constant RTT and a data rate that fluctuates 2357 in a sawtooth due to AIMD congestion control. Assume the losses are 2358 being controlled to make the average data rate meet some goal which 2359 is equal or greater than the target_rate. The necessary run length 2360 to meet the target_rate can be computed as follows: 2362 For some value of Wmin, the window will sweep from Wmin packets to 2363 2*Wmin packets in 2*Wmin RTT (due to delayed ACK). Unlike the 2364 queuing case where Wmin = target_window_size, we want the average of 2365 Wmin and 2*Wmin to be the target_window_size, so the average data 2366 rate is the target rate. Thus we want Wmin = 2367 (2/3)*target_window_size. 2369 Between losses each sawtooth delivers (1/2)(Wmin+2*Wmin)(2Wmin) 2370 packets in 2*Wmin round trip times. 2372 Substituting these together we get: 2374 target_run_length = (4/3)(target_window_size^2) 2376 Note that this is 44% of the reference_run_length computed earlier. 2377 This makes sense because under the assumptions in Section 5.2 the 2378 AMID sawtooth caused a queue at the bottleneck, which raised the 2379 effective RTT by 50%. 2381 Appendix B. The effects of ACK scheduling 2383 For many network technologies simple queuing models don't apply: the 2384 network schedules, thins or otherwise alters the timing of ACKs and 2385 data, generally to raise the efficiency of the channel allocation 2386 algorithms when confronted with relatively widely spaced small ACKs. 2387 These efficiency strategies are ubiquitous for half duplex, wireless 2388 and broadcast media. 2390 Altering the ACK stream by holding or thinning ACKs typically has two 2391 consequences: it raises the implied bottleneck IP capacity, making 2392 the fine grained slowstart bursts either faster or larger and it 2393 raises the effective RTT by the average time that the ACKs and data 2394 are delayed. The first effect can be partially mitigated by re- 2395 clocking ACKs once they are beyond the bottleneck on the return path 2396 to the sender, however this further raises the effective RTT. 2398 The most extreme example of this sort of behavior would be a half 2399 duplex channel that is not released as long as the endpoint currently 2400 holding the channel has more traffic (data or ACKs) to send. Such 2401 environments cause self clocked protocols under full load to revert 2402 to extremely inefficient stop and wait behavior. The channel 2403 constrains the protocol to send an entire window of data as a single 2404 contiguous burst on the forward path, followed by the entire window 2405 of ACKs on the return path. 2407 If a particular return path contains a subpath or device that alters 2408 the timing of the ACK stream, then the entire front path from the 2409 sender up to the bottleneck must be tested at the burst parameters 2410 implied by the ACK scheduling algorithm. The most important 2411 parameter is the Implied Bottleneck IP Capacity, which is the average 2412 rate at which the ACKs advance snd.una. Note that thinning the ACK 2413 stream (relying on the cumulative nature of seg.ack to permit 2414 discarding some ACKs) causes most TCP implementations to send 2415 interface rate bursts to offset the longer times between ACKs in 2416 order to maintain the average data rate. 2418 Note that due to ubiquitous self clocking in Internet protocols, ill 2419 conceived channel allocation mechanisms are likely to increases the 2420 queuing stress on the front path because they cause larger full 2421 sender rate data bursts. 2423 Holding data or ACKs for channel allocation or other reasons (such as 2424 forward error correction) always raises the effective RTT relative to 2425 the minimum delay for the path. Therefore it may be necessary to 2426 replace target_RTT in the calculation in Section 5.2 by an 2427 effective_RTT, which includes the target_RTT plus a term to account 2428 for the extra delays introduced by these mechanisms. 2430 Appendix C. Version Control 2432 This section to be removed prior to publication. 2434 Formatted: Thu Apr 7 18:12:37 PDT 2016 2436 Authors' Addresses 2438 Matt Mathis 2439 Google, Inc 2440 1600 Amphitheater Parkway 2441 Mountain View, California 94043 2442 USA 2444 Email: mattmathis@google.com 2446 Al Morton 2447 AT&T Labs 2448 200 Laurel Avenue South 2449 Middletown, NJ 07748 2450 USA 2452 Phone: +1 732 420 1571 2453 Email: acmorton@att.com