idnits 2.17.1 draft-ietf-ippm-model-based-metrics-11.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 : ---------------------------------------------------------------------------- ** The document seems to lack separate sections for Informative/Normative References. All references will be assumed normative when checking for downward references. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year == Line 1545 has weird spacing: '... and n = h1...' -- The document date (June 30, 2017) is 2492 days in the past. Is this intentional? Checking references for intended status: Experimental ---------------------------------------------------------------------------- ** Obsolete normative reference: RFC 2861 (Obsoleted by RFC 7661) == Outdated reference: A later version (-15) exists of draft-ietf-tcpm-rack-02 Summary: 2 errors (**), 0 flaws (~~), 3 warnings (==), 1 comment (--). 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: January 1, 2018 AT&T Labs 6 June 30, 2017 8 Model Based Metrics for Bulk Transport Capacity 9 draft-ietf-ippm-model-based-metrics-11.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 IP diagnostics are built using test 28 streams and statistical criteria for evaluating the packet transfer 29 that mimic TCP over the complete path. The temporal structure of the 30 test stream (bursts, etc) mimic TCP or other transport protocol 31 carrying bulk data over a long path. However they are constructed to 32 be independent of the details of the subpath under test, end systems 33 or applications. Likewise the success criteria evaluates the packet 34 transfer statistics of the subpath against criteria determined by 35 protocol performance models applied to the Target Transport 36 Performance of the complete path. The success criteria also does not 37 depend on the details of the subpath, end systems or application. 39 Model Based Metrics exhibit several important new properties not 40 present in other Bulk Transport Capacity Metrics, including the 41 ability to reason about concatenated or overlapping subpaths. The 42 results are vantage independent which is critical for supporting 43 independent validation of tests by comparing results from multiple 44 measurement points. 46 This document provides a framework for designing suites of IP 47 diagnostic tests that are tailored to confirming that infrastructure 48 can meet the predetermined Target Transport Performance. It does not 49 fully specify the IP diagnostics tests needed to assure any specific 50 target 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 January 1, 2018. 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 . . . . . . . . . . . . . . . . . . . . . . . . . . 8 89 3. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 10 90 4. Background . . . . . . . . . . . . . . . . . . . . . . . . . 17 91 4.1. TCP properties . . . . . . . . . . . . . . . . . . . . . 18 92 4.2. Diagnostic Approach . . . . . . . . . . . . . . . . . . . 20 93 4.3. New requirements relative to RFC 2330 . . . . . . . . . . 21 94 5. Common Models and Parameters . . . . . . . . . . . . . . . . 22 95 5.1. Target End-to-end parameters . . . . . . . . . . . . . . 22 96 5.2. Common Model Calculations . . . . . . . . . . . . . . . . 23 97 5.3. Parameter Derating . . . . . . . . . . . . . . . . . . . 24 98 5.4. Test Preconditions . . . . . . . . . . . . . . . . . . . 24 99 6. Generating test streams . . . . . . . . . . . . . . . . . . . 25 100 6.1. Mimicking slowstart . . . . . . . . . . . . . . . . . . . 26 101 6.2. Constant window pseudo CBR . . . . . . . . . . . . . . . 27 102 6.3. Scanned window pseudo CBR . . . . . . . . . . . . . . . . 28 103 6.4. Concurrent or channelized testing . . . . . . . . . . . . 29 104 7. Interpreting the Results . . . . . . . . . . . . . . . . . . 30 105 7.1. Test outcomes . . . . . . . . . . . . . . . . . . . . . . 30 106 7.2. Statistical criteria for estimating run_length . . . . . 31 107 7.3. Reordering Tolerance . . . . . . . . . . . . . . . . . . 34 108 8. IP Diagnostic Tests . . . . . . . . . . . . . . . . . . . . . 34 109 8.1. Basic Data Rate and Packet Transfer Tests . . . . . . . . 35 110 8.1.1. Delivery Statistics at Paced Full Data Rate . . . . . 35 111 8.1.2. Delivery Statistics at Full Data Windowed Rate . . . 36 112 8.1.3. Background Packet Transfer Statistics Tests . . . . . 36 113 8.2. Standing Queue Tests . . . . . . . . . . . . . . . . . . 36 114 8.2.1. Congestion Avoidance . . . . . . . . . . . . . . . . 38 115 8.2.2. Bufferbloat . . . . . . . . . . . . . . . . . . . . . 38 116 8.2.3. Non excessive loss . . . . . . . . . . . . . . . . . 38 117 8.2.4. Duplex Self Interference . . . . . . . . . . . . . . 39 118 8.3. Slowstart tests . . . . . . . . . . . . . . . . . . . . . 39 119 8.3.1. Full Window slowstart test . . . . . . . . . . . . . 39 120 8.3.2. Slowstart AQM test . . . . . . . . . . . . . . . . . 40 121 8.4. Sender Rate Burst tests . . . . . . . . . . . . . . . . . 40 122 8.5. Combined and Implicit Tests . . . . . . . . . . . . . . . 41 123 8.5.1. Sustained Bursts Test . . . . . . . . . . . . . . . . 41 124 8.5.2. Passive Measurements . . . . . . . . . . . . . . . . 42 125 9. An Example . . . . . . . . . . . . . . . . . . . . . . . . . 43 126 9.1. Observations about applicability . . . . . . . . . . . . 44 127 10. Validation . . . . . . . . . . . . . . . . . . . . . . . . . 45 128 11. Security Considerations . . . . . . . . . . . . . . . . . . . 46 129 12. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . 47 130 13. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 47 131 14. References . . . . . . . . . . . . . . . . . . . . . . . . . 47 132 Appendix A. Model Derivations . . . . . . . . . . . . . . . . . 51 133 A.1. Queueless Reno . . . . . . . . . . . . . . . . . . . . . 52 134 Appendix B. The effects of ACK scheduling . . . . . . . . . . . 53 135 Appendix C. Version Control . . . . . . . . . . . . . . . . . . 54 136 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 54 138 1. Introduction 140 Model Based Metrics (MBM) rely on peer-reviewed mathematical models 141 to specify a Targeted Suite of IP Diagnostic tests, designed to 142 assess whether common transport protocols can be expected to meet a 143 predetermined Target Transport Performance over an Internet path. 144 This note describes the modeling framework to derive the test 145 parameters for assessing an Internet path's ability to support a 146 predetermined Bulk Transport Capacity. 148 Each test in the Targeted IP Diagnostic Suite (TIDS) measures some 149 aspect of IP packet transfer needed to meet the Target Transport 150 Performance. For Bulk Transport Capacity the TIDS includes IP 151 diagnostic tests to verify that there is: sufficient IP capacity 152 (data rate); sufficient queue space at bottlenecks to absorb and 153 deliver typical transport bursts; and that the background packet loss 154 ratio is low enough not to interfere with congestion control; and 155 other properties described below. Unlike typical IPPM metrics which 156 yield measures of network properties, Model Based Metrics nominally 157 yield pass/fail evaluations of the ability of standard transport 158 protocols to meet the specific performance objective over some 159 network path. 161 In most cases, the IP diagnostic tests can be implemented by 162 combining existing IPPM metrics with additional controls for 163 generating test streams having a specified temporal structure (bursts 164 or standing queues caused by constant bit rate streams, etc.) and 165 statistical criteria for evaluating packet transfer. The temporal 166 structure of the test streams mimic transport protocol behavior over 167 the complete path; the statistical criteria models the transport 168 protocol's response to less than ideal IP packet transfer. 170 This note addresses Bulk Transport Capacity. It describes an 171 alternative to the approach presented in "A Framework for Defining 172 Empirical Bulk Transfer Capacity Metrics" [RFC3148]. Other Model 173 Based Metrics may cover other applications and transports, such as 174 VoIP over UDP and RTP, and new transport protocols. 176 The MBM approach, mapping Target Transport Performance to a Targeted 177 IP Diagnostic Suite (TIDS) of IP tests, solves some intrinsic 178 problems with using TCP or other throughput maximizing protocols for 179 measurement. In particular all throughput maximizing protocols (and 180 TCP congestion control in particular) cause some level of congestion 181 in order to detect when they have reached the available capacity 182 limitation of the network. This self inflicted congestion obscures 183 the network properties of interest and introduces non-linear dynamic 184 equilibrium behaviors that make any resulting measurements useless as 185 metrics because they have no predictive value for conditions or paths 186 different than that of the measurement itself. In order to prevent 187 these effects it is necessary to avoid the effects of TCP congestion 188 control in the measurement method. These issues are discussed at 189 length in Section 4. Readers whom are unfamiliar with basic 190 properties of TCP and TCP-like congestion control may find it easier 191 to start at Section 4 or Section 4.1. 193 A Targeted IP Diagnostic Suite does not have such difficulties. IP 194 diagnostics can be constructed such that they make strong statistical 195 statements about path properties that are independent of the 196 measurement details, such as vantage and choice of measurement 197 points. 199 1.1. Version Control 201 RFC Editor: Please remove this entire subsection prior to 202 publication. 204 REF Editor: The reference to draft-ietf-tcpm-rack is to attribute an 205 idea. This document should not block waiting for the completion of 206 that one. 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: Thu Jun 29 19:08:08 PDT 2017 214 Changes since -10 draft: 216 o A few more nits from various sources. 217 o (From IETF LC review comments.) 218 o David Mandelberg: design metrics to prevent DDOS. 219 o From Robert Sparks: 221 * Remove all legacy 2119 language. 222 * Fixed Xr notation inconsistency. 223 * Adjusted abstract: tests are only partially specified. 224 * Avoid rather than suppress the effects of congestion control 225 * Removed the unnecessary, excessively abstract and unclear 226 thought about IP vs TCP measurements. 227 * Changed "thwarted" to "not fulfilled". 228 * Qualified language about burst models. 229 * Replaced "infinitesimal" with other language. 230 * Added citations for the reordering strawman. 231 * Pointed out that psuedo CBR tests depend on self clock. 232 * Fixed some run on sentences. 233 o Update language to reflect RFC7567, AQM recommendations. 234 o Suggestion from Merry Mou (MIT) 236 Changes since -09 draft: 238 o Five last minute editing nits. 240 Changes since -08 draft: 242 o Language, spelling and usage nits. 243 o Expanded the abstract describe the models. 244 o Remove superfluous standards like language 245 o Remove superfluous "future technology" language. 246 o Interconnects -> network interconnections. 247 o Added more labels to Figure 1. 248 o Defined Bulk Transport. 249 o Clarified "implied bottleneck IP capacity" 250 o Clarified the history of the BTC metrics. 251 o Clarified stochastic vs non-stochastic test traffic generation. 252 o Reworked Fig 2 and 6.1 "Mimicking slowstart" 253 o Described the unsynchronized parallel stream failure case. 254 o Discussed how to measure devices that use virtual queues. 255 o Changed section 8.5.2 (Streaming Media) to be Passive 256 Measurements. 258 Changes since -07 draft: 260 o Sharpened the use of "statistical criteria" 261 o Sharpened the definition of test_window, and removed related 262 redundant text in several places 263 o Clarified "equilibrium" as "dynamic equilibrium, similar to 264 processes observed in chemistry" 265 o Properly explained "Heisenberg" as "observer effect" 266 o Added the observation from RFC 6576 that HW and SW congestion 267 control implementations do not generally give the same results. 268 o Noted that IP and application metrics differ as to how overhead is 269 handled. MBM is explicit about how it handles overhead. 270 o Clarified the language and added a new reference about the 271 problems caused by token bucket policers. 272 o Added an subsection in the example that comments on some of issues 273 that need to be mentioned in a future usage or applicability doc. 274 o Updated ippm-2680-bis to RFC7680 275 o Many terminology, punctuation and spelling nits. 277 Changes since -06 draft: 279 o More language nits: 281 * "Targeted IP Diagnostic Suite (TIDS)" replaces "Targeted 282 Diagnostic Suite (TDS)". 283 * "implied bottleneck IP capacity" replaces "implied bottleneck 284 IP rate". 285 * Updated to ECN CE Marks. 286 * Added "specified temporal structure" 287 * "test stream" replaces "test traffic" 288 * "packet transfer" replaces "packet delivery" 289 * Reworked discussion of slowstart, bursts and pacing. 291 * RFC 7567 replaces RFC 2309. 293 Changes since -05 draft: 295 o Wordsmithing on sections overhauled in -05 draft. 296 o Reorganized the document: 298 * Relocated subsection "Preconditions". 299 * Relocated subsection "New Requirements relative to RFC 2330". 300 o Addressed nits and not so nits by Ruediger Geib. (Thanks!) 301 o Substantially tightened the entire definitions section. 302 o Many terminology changes, to better conform to other docs : 304 * IP rate and IP capacity (following RFC 5136) replaces various 305 forms of link data rate. 306 * subpath replaces link. 307 * target_window_size replaces target_pipe_size. 308 * implied bottleneck IP rate replaces effective bottleneck link 309 rate. 310 * Packet delivery statistics replaces delivery statistics. 312 Changes since -04 draft: 314 o The introduction was heavily overhauled: split into a separate 315 introduction and overview. 316 o The new shorter introduction: 318 * Is a problem statement; 319 * This document provides a framework; 320 * That it replaces TCP measurement by IP tests; 321 * That the results are pass/fail. 322 o Added a diagram of the framework to the overview 323 o and introduces all of the elements of the framework. 324 o Renumbered sections, reducing the depth of some section numbers. 325 o Updated definitions to better agree with other documents: 327 * Reordered section 2 328 * Bulk [data] performance -> Bulk Transport Capacity, everywhere 329 including the title. 330 * loss rate and loss probability -> packet loss ratio 331 * end-to-end path -> complete path 332 * [end-to-end][target] performance -> Target Transport 333 Performance 334 * load test -> capacity test 336 2. Overview 338 This document describes a modeling framework for deriving a Targeted 339 IP Diagnostic Suite from a predetermined Target Transport 340 Performance. It is not a complete specification, and relies on other 341 standards documents to define important details such as packet Type-P 342 selection, sampling techniques, vantage selection, etc. We imagine 343 Fully Specified - Targeted IP Diagnostic Suites (FS-TIDS), that 344 define all of these details. We use Targeted IP Diagnostic Suite 345 (TIDS) to refer to the subset of such a specification that is in 346 scope for this document. This terminology is defined in Section 3. 348 Section 4 describes some key aspects of TCP behavior and what they 349 imply about the requirements for IP packet transfer. Most of the IP 350 diagnostic tests needed to confirm that the path meets these 351 properties can be built on existing IPPM metrics, with the addition 352 of statistical criteria for evaluating packet transfer and in a few 353 cases, new mechanisms to implement the required temporal structure. 354 (One group of tests, the standing queue tests described in 355 Section 8.2, don't correspond to existing IPPM metrics, but suitable 356 new IPPM metrics can be patterned after the existing definitions.) 358 Figure 1 shows the MBM modeling and measurement framework. The 359 Target Transport Performance, at the top of the figure, is determined 360 by the needs of the user or application, outside the scope of this 361 document. For Bulk Transport Capacity, the main performance 362 parameter of interest is the Target Data Rate. However, since TCP's 363 ability to compensate for less than ideal network conditions is 364 fundamentally affected by the Round Trip Time (RTT) and the Maximum 365 Transmission Unit (MTU) of the complete path, these parameters must 366 also be specified in advance based on knowledge about the intended 367 application setting. They may reflect a specific application over a 368 real path through the Internet or an idealized application and 369 hypothetical path representing a typical user community. Section 5 370 describes the common parameters and models derived from the Target 371 Transport Performance. 373 Target Transport Performance 374 (Target Data Rate, Target RTT and Target MTU) 375 | 376 ________V_________ 377 | mathematical | 378 | models | 379 | | 380 ------------------ 381 Traffic parameters | | Statistical criteria 382 | | 383 _______V____________V____Targeted_______ 384 | | * * * | Diagnostic Suite | 385 _____|_______V____________V________________ | 386 __|____________V____________V______________ | | 387 | IP diagnostic tests | | | 388 | | | | | | 389 | _____________V__ __V____________ | | | 390 | | traffic | | Delivery | | | | 391 | | pattern | | Evaluation | | | | 392 | | generation | | | | | | 393 | -------v-------- ------^-------- | | | 394 | | v test stream via ^ | | |-- 395 | | -->======================>-- | | | 396 | | subpath under test | |- 397 ----V----------------------------------V--- | 398 | | | | | | 399 V V V V V V 400 fail/inconclusive pass/fail/inconclusive 401 (traffic generation status) (test result) 403 Overall Modeling Framework 405 Figure 1 407 Mathematical TCP models are used to determine Traffic parameters and 408 subsequently to design traffic patterns that mimic TCP or other 409 transport protocol delivering bulk data and operating at the Target 410 Data Rate, MTU and RTT over a full range of conditions, including 411 flows that are bursty at multiple time scales. The traffic patterns 412 are generated based on the three Target parameters of complete path 413 and independent of the properties of individual subpaths using the 414 techniques described in Section 6. As much as possible the test 415 streams are generated deterministically (precomputed) to minimize the 416 extent to which test methodology, measurement points, measurement 417 vantage or path partitioning affect the details of the measurement 418 traffic. 420 Section 7 describes packet transfer statistics and methods to test 421 them against the statistical criteria provided by the mathematical 422 models. Since the statistical criteria typically apply to the 423 complete path (a composition of subpaths) [RFC6049], in situ testing 424 requires that the end-to-end statistical criteria be apportioned as 425 separate criteria for each subpath. Subpaths that are expected to be 426 bottlenecks would then be permitted to contribute a larger fraction 427 of the end-to-end packet loss budget. In compensation, subpaths that 428 are not expected to exhibit bottlenecks must be constrained to 429 contribute less packet loss. Thus the statistical criteria for each 430 subpath in each test of a TIDS is an apportioned share of the end-to- 431 end statistical criteria for the complete path which was determined 432 by the mathematical model. 434 Section 8 describes the suite of individual tests needed to verify 435 all of required IP delivery properties. A subpath passes if and only 436 if all of the individual IP diagnostic tests pass. Any subpath that 437 fails any test indicates that some users are likely to fail to attain 438 their Target Transport Performance under some conditions. In 439 addition to passing or failing, a test can be deemed to be 440 inconclusive for a number of reasons including: the precomputed 441 traffic pattern was not accurately generated; the measurement results 442 were not statistically significant; and others such as failing to 443 meet some required test preconditions. If all tests pass but some 444 are inconclusive, then the entire suite is deemed to be inconclusive. 446 In Section 9 we present an example TIDS that might be representative 447 of High Definition (HD) video, and illustrate how Model Based Metrics 448 can be used to address difficult measurement situations, such as 449 confirming that inter-carrier exchanges have sufficient performance 450 and capacity to deliver HD video between ISPs. 452 Since there is some uncertainty in the modeling process, Section 10 453 describes a validation procedure to diagnose and minimize false 454 positive and false negative results. 456 3. Terminology 458 Terms containing underscores (rather than spaces) appear in equations 459 and typically have algorithmic definitions. 461 General Terminology: 463 Target: A general term for any parameter specified by or derived 464 from the user's application or transport performance requirements. 465 Target Transport Performance: Application or transport performance 466 target values for the complete path. For Bulk Transport Capacity 467 defined in this note the Target Transport Performance includes the 468 Target Data Rate, Target RTT and Target MTU as described below. 469 Target Data Rate: The specified application data rate required for 470 an application's proper operation. Conventional Bulk Transport 471 Capacity (BTC) metrics are focused on the Target Data Rate, 472 however these metrics had little or no predictive value because 473 they do not consider the effects of the other two parameters of 474 the Target Transport Performance, the RTT and MTU of the complete 475 paths. 476 Target RTT (Round Trip Time): The specified baseline (minimum) RTT 477 of the longest complete path over which the user expects to be 478 able to meet the target performance. TCP and other transport 479 protocol's ability to compensate for path problems is generally 480 proportional to the number of round trips per second. The Target 481 RTT determines both key parameters of the traffic patterns (e.g. 482 burst sizes) and the thresholds on acceptable IP packet transfer 483 statistics. The Target RTT must be specified considering 484 appropriate packets sizes: MTU sized packets on the forward path, 485 ACK sized packets (typically header_overhead) on the return path. 486 Note that Target RTT is specified and not measured, MBM 487 measurements derived for a given target_RTT will be applicable to 488 any path with a smaller RTTs. 489 Target MTU (Maximum Transmission Unit): The specified maximum MTU 490 supported by the complete path the over which the application 491 expects to meet the target performance. In this document assume a 492 1500 Byte MTU unless otherwise specified. If some subpath has a 493 smaller MTU, then it becomes the Target MTU for the complete path, 494 and all model calculations and subpath tests must use the same 495 smaller MTU. 496 Targeted IP Diagnostic Suite (TIDS): A set of IP diagnostic tests 497 designed to determine if an otherwise ideal complete path 498 containing the subpath under test can sustain flows at a specific 499 target_data_rate using target_MTU sized packets when the RTT of 500 the complete path is target_RTT. 501 Fully Specified Targeted IP Diagnostic Suite (FS-TIDS): A TIDS 502 together with additional specification such as measurement packet 503 type ("type-p" [RFC2330]), etc. which are out of scope for this 504 document, but need to be drawn from other standards documents. 505 Bulk Transport Capacity: Bulk Transport Capacity Metrics evaluate an 506 Internet path's ability to carry bulk data, such as large files, 507 streaming (non-real time) video, and under some conditions, web 508 images and other content. Prior efforts to define BTC metrics 509 have been based on [RFC3148], which predates our understanding of 510 TCP and the requirements described in Section 4. In general "Bulk 511 Transport" indicates that performance is determined by the 512 interplay between the network, cross traffic and congestion 513 control in the transport protocol. It excludes situations where 514 performance is dominated by the RTT alone (e.g. transactions) or 515 bottlenecks elsewhere, such as in the application itself. 516 IP diagnostic tests: Measurements or diagnostics to determine if 517 packet transfer statistics meet some precomputed target. 518 traffic patterns: The temporal patterns or burstiness of traffic 519 generated by applications over transport protocols such as TCP. 520 There are several mechanisms that cause bursts at various time 521 scales as described in Section 4.1. Our goal here is to mimic the 522 range of common patterns (burst sizes and rates, etc), without 523 tying our applicability to specific applications, implementations 524 or technologies, which are sure to become stale. 525 packet transfer statistics: Raw, detailed or summary statistics 526 about packet transfer properties of the IP layer including packet 527 losses, ECN Congestion Experienced (CE) marks, reordering, or any 528 other properties that may be germane to transport performance. 529 packet loss ratio: As defined in [RFC7680]. 530 apportioned: To divide and allocate, for example budgeting packet 531 loss across multiple subpaths such that the losses will accumulate 532 to less than a specified end-to-end loss ratio. Apportioning 533 metrics is essentially the inverse of the process described in 534 [RFC5835]. 535 open loop: A control theory term used to describe a class of 536 techniques where systems that naturally exhibit circular 537 dependencies can be analyzed by suppressing some of the 538 dependencies, such that the resulting dependency graph is acyclic. 540 Terminology about paths, etc. See [RFC2330] and [RFC7398] for 541 existing terms and definitions. 543 data sender: Host sending data and receiving ACKs. 544 data receiver: Host receiving data and sending ACKs. 545 complete path: The end-to-end path from the data sender to the data 546 receiver. 547 subpath: A portion of the complete path. Note that there is no 548 requirement that subpaths be non-overlapping. A subpath can be a 549 small as a single device, link or interface. 550 measurement point: Measurement points as described in [RFC7398]. 551 test path: A path between two measurement points that includes a 552 subpath of the complete path under test. If the measurement 553 points are off path, the test path may include "test leads" 554 between the measurement points and the subpath. 555 dominant bottleneck: The bottleneck that generally determines most 556 of packet transfer statistics for the entire path. It typically 557 determines a flow's self clock timing, packet loss and ECN 558 Congestion Experienced (CE) marking rate, with other potential 559 bottlenecks having less effect on the packet transfer statistics. 560 See Section 4.1 on TCP properties. 562 front path: The subpath from the data sender to the dominant 563 bottleneck. 564 back path: The subpath from the dominant bottleneck to the receiver. 565 return path: The path taken by the ACKs from the data receiver to 566 the data sender. 567 cross traffic: Other, potentially interfering, traffic competing for 568 network resources (bandwidth and/or queue capacity). 570 Properties determined by the complete path and application. These 571 are described in more detail in Section 5.1. 573 Application Data Rate: General term for the data rate as seen by the 574 application above the transport layer in bytes per second. This 575 is the payload data rate, and explicitly excludes transport and 576 lower level headers (TCP/IP or other protocols), retransmissions 577 and other overhead that is not part to the total quantity of data 578 delivered to the application. 579 IP rate: The actual number of IP-layer bytes delivered through a 580 subpath, per unit time, including TCP and IP headers, retransmits 581 and other TCP/IP overhead. Follows from IP-type-P Link Usage 582 [RFC5136]. 583 IP capacity: The maximum number of IP-layer bytes that can be 584 transmitted through a subpath, per unit time, including TCP and IP 585 headers, retransmits and other TCP/IP overhead. Follows from IP- 586 type-P Link Capacity [RFC5136]. 587 bottleneck IP capacity: The IP capacity of the dominant bottleneck 588 in the forward path. All throughput maximizing protocols estimate 589 this capacity by observing the IP rate delivered through the 590 bottleneck. Most protocols derive their self clocks from the 591 timing of this data. See Section 4.1 and Appendix B for more 592 details. 593 implied bottleneck IP capacity: This is the bottleneck IP capacity 594 implied by the ACKs returning from the receiver. It is determined 595 by looking at how much application data the ACK stream at the 596 sender reports delivered to the data receiver per unit time at 597 various time scales. If the return path is thinning, batching or 598 otherwise altering the ACK timing the implied bottleneck IP 599 capacity over short time scales might be substantially larger than 600 the bottleneck IP capacity averaged over a full RTT. Since TCP 601 derives its clock from the data delivered through the bottleneck, 602 the front path must have sufficient buffering to absorb any data 603 bursts at the dimensions (size and IP rate) implied by the ACK 604 stream, which are potentially doubled during slowstart. If the 605 return path is not altering the ACK stream, then the implied 606 bottleneck IP capacity will be the same as the bottleneck IP 607 capacity. See Section 4.1 and Appendix B for more details. 608 sender interface rate: The IP rate which corresponds to the IP 609 capacity of the data sender's interface. Due to sender efficiency 610 algorithms including technologies such as TCP segmentation offload 611 (TSO), nearly all modern servers deliver data in bursts at full 612 interface link rate. Today 1 or 10 Gb/s are typical. 613 Header_overhead: The IP and TCP header sizes, which are the portion 614 of each MTU not available for carrying application payload. 615 Without loss of generality this is assumed to be the size for 616 returning acknowledgments (ACKs). For TCP, the Maximum Segment 617 Size (MSS) is the Target MTU minus the header_overhead. 619 Basic parameters common to models and subpath tests are defined here 620 are described in more detail in Section 5.2. Note that these are 621 mixed between application transport performance (excludes headers) 622 and IP performance (which include TCP headers and retransmissions as 623 part of the IP payload). 625 Window [size]: The total quantity of data carried by packets in- 626 flight plus the data represented by ACKs circulating in the 627 network is referred to as the window. See Section 4.1. Sometimes 628 used with other qualifiers (congestion window, cwnd or receiver 629 window) to indicate which mechanism is controlling the window. 630 pipe size: A general term for number of packets needed in flight 631 (the window size) to exactly fill some network path or subpath. 632 It corresponds to the window size which maximizes network power, 633 the observed data rate divided by the observed RTT. Often used 634 with additional qualifiers to specify which path, or under what 635 conditions, etc. 636 target_window_size: The average number of packets in flight (the 637 window size) needed to meet the Target Data Rate, for the 638 specified Target RTT, and MTU. It implies the scale of the bursts 639 that the network might experience. 640 run length: A general term for the observed, measured, or specified 641 number of packets that are (expected to be) delivered between 642 losses or ECN Congestion Experienced (CE) marks. Nominally one 643 over the sum of the loss and ECN CE marking probabilities, if 644 there are independently and identically distributed. 645 target_run_length: The target_run_length is an estimate of the 646 minimum number of non-congestion marked packets needed between 647 losses or ECN Congestion Experienced (CE) marks necessary to 648 attain the target_data_rate over a path with the specified 649 target_RTT and target_MTU, as computed by a mathematical model of 650 TCP congestion control. A reference calculation is shown in 651 Section 5.2 and alternatives in Appendix A 652 reference target_run_length: target_run_length computed precisely by 653 the method in Section 5.2. This is likely to be slightly more 654 conservative than required by modern TCP implementations. 656 Ancillary parameters used for some tests: 658 derating: Under some conditions the standard models are too 659 conservative. The modeling framework permits some latitude in 660 relaxing or "derating" some test parameters as described in 661 Section 5.3 in exchange for a more stringent TIDS validation 662 procedures, described in Section 10. Models can be derated by 663 including a multiplicative derating factor to make tests less 664 stringent. 665 subpath_IP_capacity: The IP capacity of a specific subpath. 666 test path: A subpath of a complete path under test. 667 test_path_RTT: The RTT observed between two measurement points using 668 packet sizes that are consistent with the transport protocol. 669 This is generally MTU sized packets of the forward path, 670 header_overhead sized packets on the return path. 671 test_path_pipe: The pipe size of a test path. Nominally the 672 test_path_RTT times the test path IP_capacity. 673 test_window: The smallest window sufficient to meet or exceed the 674 target_rate when operating with a pure self clock over a test 675 path. The test_window is typically given by 676 ceiling(target_data_rate*test_path_RTT/(target_MTU- 677 header_overhead)) but see the discussion in Appendix B about the 678 effects of channel scheduling on RTT. On some test paths the 679 test_window may need to be adjusted slightly to compensate for the 680 RTT being inflated by the devices that schedule packets. 682 The terminology below is used to define temporal patterns for test 683 stream. These patterns are designed to mimic TCP behavior, as 684 described in Section 4.1. 686 packet headway: Time interval between packets, specified from the 687 start of one to the start of the next. e.g. If packets are sent 688 with a 1 mS headway, there will be exactly 1000 packets per 689 second. 690 burst headway: Time interval between bursts, specified from the 691 start of the first packet one burst to the start of the first 692 packet of the next burst. e.g. If 4 packet bursts are sent with a 693 1 mS burst headway, there will be exactly 4000 packets per second. 694 paced single packets: Send individual packets at the specified rate 695 or packet headway. 696 paced bursts: Send bursts on a timer. Specify any 3 of: average 697 data rate, packet size, burst size (number of packets) and burst 698 headway (burst start to start). By default the bursts are assumed 699 to occur at full sender interface rate, such that the packet 700 headway within each burst is the minimum supported by the sender's 701 interface. Under some conditions it is useful to explicitly 702 specify the packet headway within each burst. 703 slowstart rate: Mimic TCP slowstart by sending 4 packet paced bursts 704 at an average data rate equal to twice the implied bottleneck IP 705 capacity (but not more than the sender interface rate). This is a 706 two level burst pattern described in more detail in Section 6.1. 707 If the implied bottleneck IP capacity is more than half of the 708 sender interface rate, slowstart rate becomes sender interface 709 rate. 710 slowstart burst: Mimic one round of TCP slowstart by sending a 711 specified number of packets packets in a two level burst pattern 712 that resembles slowstart. 713 repeated slowstart bursts: Repeat Slowstart bursts once per 714 target_RTT. For TCP each burst would be twice as large as the 715 prior burst, and the sequence would end at the first ECN CE mark 716 or lost packet. For measurement, all slowstart bursts would be 717 the same size (nominally target_window_size but other sizes might 718 be specified), and the ECN CE marks and lost packets are counted. 720 The tests described in this note can be grouped according to their 721 applicability. 723 Capacity tests: Capacity tests determine if a network subpath has 724 sufficient capacity to deliver the Target Transport Performance. 725 As long as the test stream is within the proper envelope for the 726 Target Transport Performance, the average packet losses or ECN 727 Congestion Experienced (CE) marks must be below the statistical 728 criteria computed by the model. As such, capacity tests reflect 729 parameters that can transition from passing to failing as a 730 consequence of cross traffic, additional presented load or the 731 actions of other network users. By definition, capacity tests 732 also consume significant network resources (data capacity and/or 733 queue buffer space), and the test schedules must be balanced by 734 their cost. 735 Monitoring tests: Monitoring tests are designed to capture the most 736 important aspects of a capacity test, but without presenting 737 excessive ongoing load themselves. As such they may miss some 738 details of the network's performance, but can serve as a useful 739 reduced-cost proxy for a capacity test, for example to support 740 continuous production network monitoring. 741 Engineering tests: Engineering tests evaluate how network algorithms 742 (such as AQM and channel allocation) interact with TCP-style self 743 clocked protocols and adaptive congestion control based on packet 744 loss and ECN Congestion Experienced (CE) marks. These tests are 745 likely to have complicated interactions with cross traffic and 746 under some conditions can be inversely sensitive to load. For 747 example a test to verify that an AQM algorithm causes ECN CE marks 748 or packet drops early enough to limit queue occupancy may 749 experience a false pass result in the presence of cross traffic. 750 It is important that engineering tests be performed under a wide 751 range of conditions, including both in situ and bench testing, and 752 over a wide variety of load conditions. Ongoing monitoring is 753 less likely to be useful for engineering tests, although sparse in 754 situ testing might be appropriate. 756 4. Background 758 At the time the "Framework for IP Performance Metrics" [RFC2330] was 759 published (1998), sound Bulk Transport Capacity (BTC) measurement was 760 known to be well beyond our capabilities. Even when Framework for 761 Empirical BTC Metrics [RFC3148] was published, we knew that we didn't 762 really understand the problem. Now, by hindsight we understand why 763 assessing BTC is such a hard problem: 765 o TCP is a control system with circular dependencies - everything 766 affects performance, including components that are explicitly not 767 part of the test (for example, the host processing power is not 768 in-scope of path performance tests). 769 o Congestion control is a dynamic equilibrium process, similar to 770 processes observed in chemistry and other fields. The network and 771 transport protocols find an operating point which balances between 772 opposing forces: the transport protocol pushing harder (raising 773 the data rate and/or window) while the network pushes back 774 (raising packet loss ratio, RTT and/or ECN CE marks). By design 775 TCP congestion control keeps raising the data rate until the 776 network gives some indication that its capacity has been exceeded 777 by dropping packets or adding ECN CE marks. If a TCP sender 778 accurately fills a path to its IP capacity, (e.g. the bottleneck 779 is 100% utilized), then packet losses and ECN CE marks are mostly 780 determined by the TCP sender and how aggressively it seeks 781 additional capacity, and not the network itself, since the network 782 must send exactly the signals that TCP needs to set its rate. 783 o TCP's ability to compensate for network impairments (such as loss, 784 delay and delay variation, outside of those caused by TCP itself) 785 is directly proportional to the number of send-ACK round trip 786 exchanges per second (i.e. inversely proportional to the RTT). As 787 a consequence an impaired subpath may pass a short RTT local test 788 even though it fails when the subpath is extended by an 789 effectively perfect network to some larger RTT. 790 o TCP has an extreme form of the Observer Effect (colloquially know 791 as the Heisenberg effect). Measurement and cross traffic interact 792 in unknown and ill defined ways. The situation is actually worse 793 than the traditional physics problem where you can at least 794 estimate bounds on the relative momentum of the measurement and 795 measured particles. For network measurement you can not in 796 general determine even the order of magnitude of the effect. It 797 is possible to construct measurement scenarios where the 798 measurement traffic starves real user traffic, yielding an overly 799 inflated measurement. The inverse is also possible: the user 800 traffic can fill the network, such that the measurement traffic 801 detects only minimal available capacity. You can not in general 802 determine which scenario might be in effect, so you can not gauge 803 the relative magnitude of the uncertainty introduced by 804 interactions with other network traffic. 805 o As a consequence of the properties listed above it is difficult, 806 if not impossible, for two independent implementations (HW or SW) 807 of TCP congestion control to produce equivalent performance 808 results [RFC6576] under the same network conditions, 810 These properties are a consequence of the dynamic equilibrium 811 behavior intrinsic to how all throughput maximizing protocols 812 interact with the Internet. These protocols rely on control systems 813 based on estimated network metrics to regulate the quantity of data 814 to send into the network. The packet sending characteristics in turn 815 alter the network properties estimated by the control system metrics, 816 such that there are circular dependencies between every transmission 817 characteristic and every estimated metric. Since some of these 818 dependencies are nonlinear, the entire system is nonlinear, and any 819 change anywhere causes a difficult to predict response in network 820 metrics. As a consequence Bulk Transport Capacity metrics have not 821 fulfilled the analytic framework envisioned in [RFC2330] 823 Model Based Metrics overcome these problems by making the measurement 824 system open loop: the packet transfer statistics (akin to the network 825 estimators) do not affect the traffic or traffic patterns (bursts), 826 which are computed on the basis of the Target Transport Performance. 827 A path or subpath meeting the Target Transfer Performance 828 requirements would exhibit packet transfer statistics and estimated 829 metrics that would not cause the control system to slow the traffic 830 below the Target Data Rate. 832 4.1. TCP properties 834 TCP and other self clocked protocols (e.g. SCTP) carry the vast 835 majority of all Internet data. Their dominant bulk data transport 836 behavior is to have an approximately fixed quantity of data and 837 acknowledgments (ACKs) circulating in the network. The data receiver 838 reports arriving data by returning ACKs to the data sender, the data 839 sender typically responds by sending approximately the same quantity 840 of data back into the network. The total quantity of data plus the 841 data represented by ACKs circulating in the network is referred to as 842 the window. The mandatory congestion control algorithms 843 incrementally adjust the window by sending slightly more or less data 844 in response to each ACK. The fundamentally important property of 845 this system is that it is self clocked: The data transmissions are a 846 reflection of the ACKs that were delivered by the network, the ACKs 847 are a reflection of the data arriving from the network. 849 A number of protocol features cause bursts of data, even in idealized 850 networks that can be modeled as simple queuing systems. 852 During slowstart the IP rate is doubled on each RTT by sending twice 853 as much data as was delivered to the receiver during the prior RTT. 854 Each returning ACK causes the sender to transmit twice the data the 855 ACK reported arriving at the receiver. For slowstart to be able to 856 fill the pipe, the network must be able to tolerate slowstart bursts 857 up to the full pipe size inflated by the anticipated window reduction 858 on the first loss or ECN CE mark. For example, with classic Reno 859 congestion control, an optimal slowstart has to end with a burst that 860 is twice the bottleneck rate for one RTT in duration. This burst 861 causes a queue which is equal to the pipe size (i.e. the window is 862 twice the pipe size) so when the window is halved in response to the 863 first packet loss, the new window will be the pipe size. 865 Note that if the bottleneck IP rate is less that half of the capacity 866 of the front path (which is almost always the case), the slowstart 867 bursts will not by themselves cause significant queues anywhere else 868 along the front path; they primarily exercise the queue at the 869 dominant bottleneck. 871 Several common efficiency algorithms also cause bursts. The self 872 clock is typically applied to groups of packets: the receiver's 873 delayed ACK algorithm generally sends only one ACK per two data 874 segments. Furthermore the modern senders use TCP segmentation 875 offload (TSO) to reduce CPU overhead. The sender's software stack 876 builds super sized TCP segments that the TSO hardware splits into MTU 877 sized segments on the wire. The net effect of TSO, delayed ACK and 878 other efficiency algorithms is to send bursts of segments at full 879 sender interface rate. 881 Note that these efficiency algorithms are almost always in effect, 882 including during slowstart, such that slowstart typically has a two 883 level burst structure. Section 6.1 describes slowstart in more 884 detail. 886 Additional sources of bursts include TCP's initial window [RFC6928], 887 application pauses, channel allocation mechanisms and network devices 888 that schedule ACKs. Appendix B describes these last two items. If 889 the application pauses (stops reading or writing data) for some 890 fraction of an RTT, many TCP implementations catch up to their 891 earlier window size by sending a burst of data at the full sender 892 interface rate. To fill a network with a realistic application, the 893 network has to be able to tolerate sender interface rate bursts large 894 enough to restore the prior window following application pauses. 896 Although the sender interface rate bursts are typically smaller than 897 the last burst of a slowstart, they are at a higher IP rate so they 898 potentially exercise queues at arbitrary points along the front path 899 from the data sender up to and including the queue at the dominant 900 bottleneck. It is known that these bursts can hurt network 901 performance, especially in conjunction with other queue pressure, 902 however we are not aware of any models for how frequent sender rate 903 bursts the network should be able to tolerate at various burst sizes. 905 In conclusion, to verify that a path can meet a Target Transport 906 Performance, it is necessary to independently confirm that the path 907 can tolerate bursts at the scales that can be caused by the above 908 mechanisms. Three cases are believed to be sufficient: 910 o Two level slowstart bursts sufficient to get connections started 911 properly. 912 o Ubiquitous sender interface rate bursts caused by efficiency 913 algorithms. We assume 4 packet bursts to be the most common case, 914 since it matches the effects of delayed ACK during slowstart. 915 These bursts should be assumed not to significantly affect packet 916 transfer statistics. 917 o Infrequent sender interface rate bursts that are full 918 target_window_size. Target_run_length may be derated for these 919 large fast bursts. 921 If a subpath can meet the required packet loss ratio for bursts at 922 all of these scales then it has sufficient buffering at all potential 923 bottlenecks to tolerate any of the bursts that are likely introduced 924 by TCP or other transport protocols. 926 4.2. Diagnostic Approach 928 A complete path of a given RTT and MTU, which are equal to or smaller 929 than the Target RTT and equal to or larger than the Target MTU 930 respectively, is expected to be able to attain a specified Bulk 931 Transport Capacity when all of the following conditions are met: 933 1. The IP capacity is above the Target Data Rate by sufficient 934 margin to cover all TCP/IP overheads. This can be confirmed by 935 the tests described in Section 8.1 or any number of IP capacity 936 tests adapted to implement MBM. 937 2. The observed packet transfer statistics are better than required 938 by a suitable TCP performance model (e.g. fewer packet losses or 939 ECN CE marks). See Section 8.1 or any number of low or fixed 940 rate packet loss tests outside of MBM. 941 3. There is sufficient buffering at the dominant bottleneck to 942 absorb a slowstart bursts large enough to get the flow out of 943 slowstart at a suitable window size. See Section 8.3. 945 4. There is sufficient buffering in the front path to absorb and 946 smooth sender interface rate bursts at all scales that are likely 947 to be generated by the application, any channel arbitration in 948 the ACK path or any other mechanisms. See Section 8.4. 949 5. When there is a slowly rising standing queue at the bottleneck 950 the onset of packet loss has to be at an appropriate point (time 951 or queue depth) and progressive [RFC7567]. See Section 8.2. 952 6. When there is a standing queue at a bottleneck for a shared media 953 subpath (e.g. half duplex), there must be a suitable bounds on 954 the interaction between ACKs and data, for example due to the 955 channel arbitration mechanism. See Section 8.2.4. 957 Note that conditions 1 through 4 require capacity tests for 958 validation, and thus may need to be monitored on an ongoing basis. 959 Conditions 5 and 6 require engineering tests, which are best 960 performed in controlled environments such as a bench test. They 961 won't generally fail due to load, but may fail in the field due to 962 configuration errors, etc. and should be spot checked. 964 A tool that can perform many of the tests is available from 965 [MBMSource]. 967 4.3. New requirements relative to RFC 2330 969 Model Based Metrics are designed to fulfill some additional 970 requirements that were not recognized at the time RFC 2330 was 971 written [RFC2330]. These missing requirements may have significantly 972 contributed to policy difficulties in the IP measurement space. Some 973 additional requirements are: 975 o IP metrics must be actionable by the ISP - they have to be 976 interpreted in terms of behaviors or properties at the IP or lower 977 layers, that an ISP can test, repair and verify. 978 o Metrics should be spatially composable, such that measures of 979 concatenated paths should be predictable from subpaths. 980 o Metrics must be vantage point invariant over a significant range 981 of measurement point choices, including off path measurement 982 points. The only requirements on MP selection should be that the 983 RTT between the MPs is below some reasonable bound, and that the 984 effects of the "test leads" connecting MPs to the subpath under 985 test can be can be calibrated out of the measurements. The latter 986 might be be accomplished if the test leads are effectively ideal 987 or their properties can be deducted from the measurements between 988 the MPs. While many of tests require that the test leads have at 989 least as much IP capacity as the subpath under test, some do not, 990 for example Background Packet Transfer Tests described in 991 Section 8.1.3. 993 o Metric measurements should be repeatable by multiple parties with 994 no specialized access to MPs or diagnostic infrastructure. It 995 should be possible for different parties to make the same 996 measurement and observe the same results. In particular it is 997 specifically important that both a consumer (or their delegate) 998 and ISP be able to perform the same measurement and get the same 999 result. Note that vantage independence is key to meeting this 1000 requirement. 1002 5. Common Models and Parameters 1004 5.1. Target End-to-end parameters 1006 The target end-to-end parameters are the Target Data Rate, Target RTT 1007 and Target MTU as defined in Section 3. These parameters are 1008 determined by the needs of the application or the ultimate end user 1009 and the complete Internet path over which the application is expected 1010 to operate. The target parameters are in units that make sense to 1011 upper layers: payload bytes delivered to the application, above TCP. 1012 They exclude overheads associated with TCP and IP headers, 1013 retransmits and other protocols (e.g. DNS). Note that IP-based 1014 network services include TCP headers and retransmissions as part of 1015 delivered payload, and this difference is recognized in calculations 1016 below (header_overhead). 1018 Other end-to-end parameters defined in Section 3 include the 1019 effective bottleneck data rate, the sender interface data rate and 1020 the TCP and IP header sizes. 1022 The target_data_rate must be smaller than all subpath IP capacities 1023 by enough headroom to carry the transport protocol overhead, 1024 explicitly including retransmissions and an allowance for 1025 fluctuations in TCP's actual data rate. Specifying a 1026 target_data_rate with insufficient headroom is likely to result in 1027 brittle measurements having little predictive value. 1029 Note that the target parameters can be specified for a hypothetical 1030 path, for example to construct TIDS designed for bench testing in the 1031 absence of a real application; or for a live in situ test of 1032 production infrastructure. 1034 The number of concurrent connections is explicitly not a parameter to 1035 this model. If a subpath requires multiple connections in order to 1036 meet the specified performance, that must be stated explicitly and 1037 the procedure described in Section 6.4 applies. 1039 5.2. Common Model Calculations 1041 The Target Transport Performance is used to derive the 1042 target_window_size and the reference target_run_length. 1044 The target_window_size, is the average window size in packets needed 1045 to meet the target_rate, for the specified target_RTT and target_MTU. 1046 It is given by: 1048 target_window_size = ceiling( target_rate * target_RTT / ( target_MTU 1049 - header_overhead ) ) 1051 Target_run_length is an estimate of the minimum required number of 1052 unmarked packets that must be delivered between losses or ECN 1053 Congestion Experienced (CE) marks, as computed by a mathematical 1054 model of TCP congestion control. The derivation here follows 1055 [MSMO97], and by design is quite conservative. 1057 Reference target_run_length is derived as follows: assume the 1058 subpath_IP_capacity is infinitesimally larger than the 1059 target_data_rate plus the required header_overhead. Then 1060 target_window_size also predicts the onset of queuing. A larger 1061 window will cause a standing queue at the bottleneck. 1063 Assume the transport protocol is using standard Reno style Additive 1064 Increase, Multiplicative Decrease (AIMD) congestion control [RFC5681] 1065 (but not Appropriate Byte Counting [RFC3465]) and the receiver is 1066 using standard delayed ACKs. Reno increases the window by one packet 1067 every pipe_size worth of ACKs. With delayed ACKs this takes 2 Round 1068 Trip Times per increase. To exactly fill the pipe, the spacing of 1069 losses must be no closer than when the peak of the AIMD sawtooth 1070 reached exactly twice the target_window_size. Otherwise, the 1071 multiplicative window reduction triggered by the loss would cause the 1072 network to be under-filled. Following [MSMO97] the number of packets 1073 between losses must be the area under the AIMD sawtooth. They must 1074 be no more frequent than every 1 in 1075 ((3/2)*target_window_size)*(2*target_window_size) packets, which 1076 simplifies to: 1078 target_run_length = 3*(target_window_size^2) 1080 Note that this calculation is very conservative and is based on a 1081 number of assumptions that may not apply. Appendix A discusses these 1082 assumptions and provides some alternative models. If a different 1083 model is used, a FS-TIDS must document the actual method for 1084 computing target_run_length and ratio between alternate 1085 target_run_length and the reference target_run_length calculated 1086 above, along with a discussion of the rationale for the underlying 1087 assumptions. 1089 These two parameters, target_window_size and target_run_length, 1090 directly imply most of the individual parameters for the tests in 1091 Section 8. 1093 5.3. Parameter Derating 1095 Since some aspects of the models are very conservative, the MBM 1096 framework permits some latitude in derating test parameters. Rather 1097 than trying to formalize more complicated models we permit some test 1098 parameters to be relaxed as long as they meet some additional 1099 procedural constraints: 1101 o The FS-TIDS must document and justify the actual method used to 1102 compute the derated metric parameters. 1103 o The validation procedures described in Section 10 must be used to 1104 demonstrate the feasibility of meeting the Target Transport 1105 Performance with infrastructure that just barely passes the 1106 derated tests. 1107 o The validation process for a FS-TIDS itself must be documented is 1108 such a way that other researchers can duplicate the validation 1109 experiments. 1111 Except as noted, all tests below assume no derating. Tests where 1112 there is not currently a well established model for the required 1113 parameters explicitly include derating as a way to indicate 1114 flexibility in the parameters. 1116 5.4. Test Preconditions 1118 Many tests have preconditions which are required to assure their 1119 validity. Examples include: the presence or non-presence of cross 1120 traffic on specific subpaths; negotiating ECN; and appropriate 1121 preamble packet stream to testing to put reactive network elements 1122 into the proper states [RFC7312]. If preconditions are not properly 1123 satisfied for some reason, the tests should be considered to be 1124 inconclusive. In general it is useful to preserve diagnostic 1125 information as to why the preconditions were not met, and any test 1126 data that was collected even if it is not useful for the intended 1127 test. Such diagnostic information and partial test data may be 1128 useful for improving the test or test procedures themselves. 1130 It is important to preserve the record that a test was scheduled, 1131 because otherwise precondition enforcement mechanisms can introduce 1132 sampling bias. For example, canceling tests due to cross traffic on 1133 subscriber access links might introduce sampling bias in tests of the 1134 rest of the network by reducing the number of tests during peak 1135 network load. 1137 Test preconditions and failure actions must be specified in a FS- 1138 TIDS. 1140 6. Generating test streams 1142 Many important properties of Model Based Metrics, such as vantage 1143 independence, are a consequence of using test streams that have 1144 temporal structures that mimic TCP or other transport protocols 1145 running over a complete path. As described in Section 4.1, self 1146 clocked protocols naturally have burst structures related to the RTT 1147 and pipe size of the complete path. These bursts naturally get 1148 larger (contain more packets) as either the Target RTT or Target Data 1149 Rate get larger, or the Target MTU gets smaller. An implication of 1150 these relationships is that test streams generated by running self 1151 clocked protocols over short subpaths may not adequately exercise the 1152 queuing at any bottleneck to determine if the subpath can support the 1153 full Target Transport Performance over the complete path. 1155 Failing to authentically mimic TCP's temporal structure is part of 1156 the reason why simple performance tools such as iPerf, netperf, nc, 1157 etc have the reputation of yielding false pass results over short 1158 test paths, even when some subpath has a flaw. 1160 The definitions in Section 3 are sufficient for most test streams. 1161 We describe the slowstart and standing queue test streams in more 1162 detail. 1164 In conventional measurement practice stochastic processes are used to 1165 eliminate many unintended correlations and sample biases. However 1166 MBM tests are designed to explicitly mimic temporal correlations 1167 caused by network or protocol elements themselves. Some portions of 1168 these systems, such as traffic arrival (test scheduling) are 1169 naturally stochastic. Other behaviors, such as back-to-back packet 1170 transmissions, are dominated by implementation specific deterministic 1171 effects. Although these behaviors always contain non-deterministic 1172 elements and might be modeled stochastically, these details typically 1173 do not contribute significantly to the overall system behavior. 1174 Furthermore, it is known that real protocols are subject to failures 1175 caused by network property estimators suffering from bias due to 1176 correlation in their own traffic. For example TCP's RTT estimator 1177 used to determine the Retransmit Time Out (RTO), can be fooled by 1178 periodic cross traffic or start-stop applications. For these reasons 1179 many details of the test streams are specified deterministically. 1181 It may prove useful to introduce fine grained noise sources into the 1182 models used for generating test streams in an update of Model Based 1183 Metrics, but the complexity is not warranted at the time this 1184 document was written. 1186 6.1. Mimicking slowstart 1188 TCP slowstart has a two level burst structure as shown in Figure 2. 1189 The fine time structure is caused by efficiency algorithms that 1190 deliberately batch work (CPU, channel allocation, etc) to better 1191 amortize certain network and host overheads. ACKs passing through 1192 the return path typically cause the sender to transmit small bursts 1193 of data at full sender interface rate. For example TCP Segmentation 1194 Offload (TSO) and Delayed Acknowledgment both contribute to this 1195 effect. During slowstart these bursts are at the same headway as the 1196 returning ACKs, but are typically twice as large (e.g. having twice 1197 as much data) as the ACK reported was delivered to the receiver. Due 1198 to variations in delayed ACK and algorithms such as Appropriate Byte 1199 Counting [RFC3465], different pairs of senders and receivers produce 1200 slightly different burst patterns. Without loss of generality, we 1201 assume each ACK causes 4 packet sender interface rate bursts at an 1202 average headway equal to the ACK headway, and corresponding to 1203 sending at an average rate equal to twice the effective bottleneck IP 1204 rate. Each slowstart burst consists of a series of 4 packet sender 1205 interface rate bursts such that the total number of packets is the 1206 current window size (as of the last packet in the burst). 1208 The coarse time structure is due to each RTT being a reflection of 1209 the prior RTT. For real transport protocols, each slowstart burst is 1210 twice as large (twice the window) as the previous burst but is spread 1211 out in time by the network bottleneck, such that each successive RTT 1212 exhibits the same effective bottleneck IP rate. The slowstart phase 1213 ends on the first lost packet or ECN mark, which is intended to 1214 happen after successive slowstart bursts merge in time: the next 1215 burst starts before the bottleneck queue is fully drained and the 1216 prior burst is complete. 1218 For diagnostic tests described below we preserve the fine time 1219 structure but manipulate the coarse structure of the slowstart bursts 1220 (burst size and headway) to measure the ability of the dominant 1221 bottleneck to absorb and smooth slowstart bursts. 1223 Note that a stream of repeated slowstart bursts has three different 1224 average rates, depending on the averaging time interval. At the 1225 finest time scale (a few packet times at the sender interface) the 1226 peak of the average IP rate is the same as the sender interface rate; 1227 at a medium timescale (a few ACK times at the dominant bottleneck) 1228 the peak of the average IP rate is twice the implied bottleneck IP 1229 capacity; and at time scales longer than the target_RTT and when the 1230 burst size is equal to the target_window_size, the average rate is 1231 equal to the target_data_rate. This pattern corresponds to repeating 1232 the last RTT of TCP slowstart when delayed ACK and sender side byte 1233 counting are present but without the limits specified in Appropriate 1234 Byte Counting [RFC3465]. 1236 time ==> ( - equals one packet) 1238 Fine time structure of the packet stream: 1240 ---- ---- ---- ---- ---- 1242 |<>| sender interface rate bursts (typically 3 or 4 packets) 1243 |<===>| burst headway (from the ACK headway) 1245 \____repeating sender______/ 1246 rate bursts 1248 Coarse (RTT level) time structure of the packet stream: 1250 ---- ---- ---- ---- ---- ---- ---- ... 1252 |<========================>| slowstart burst size (from the window) 1253 |<==============================================>| slowstart headway 1254 (from the RTT) 1255 \__________________________/ \_________ ... 1256 one slowstart burst Repeated slowstart bursts 1258 Multiple levels of Slowstart Bursts 1260 Figure 2 1262 6.2. Constant window pseudo CBR 1264 Implement pseudo constant bit rate by running a standard self clocked 1265 protocol such as TCP with a fixed window size. If that window size 1266 is test_window, the data rate will be slightly above the target_rate. 1268 Since the test_window is constrained to be an integer number of 1269 packets, for small RTTs or low data rates there may not be 1270 sufficiently precise control over the data rate. Rounding the 1271 test_window up (as defined above) is likely to result in data rates 1272 that are higher than the target rate, but reducing the window by one 1273 packet may result in data rates that are too small. Also cross 1274 traffic potentially raises the RTT, implicitly reducing the rate. 1276 Cross traffic that raises the RTT nearly always makes the test more 1277 strenuous (more demanding for the network path). 1279 Note that Constant window pseudo CBR (and Scanned window pseudo CBR 1280 in the next section) both rely on a self clock which is at least 1281 partially derived from the properties of the subnet under test. This 1282 introduces the possibility that the subnet under test exhibits 1283 behaviors such as extreme RTT fluctuations that prevent these 1284 algorithms from accurately controlling data rates. 1286 A FS-TIDS specifying a constant window CBR test must explicitly 1287 indicate under what conditions errors in the data rate cause tests to 1288 be inconclusive. Conventional paced measurement traffic may be more 1289 appropriate for these environments. 1291 6.3. Scanned window pseudo CBR 1293 Scanned window pseudo CBR is similar to the constant window CBR 1294 described above, except the window is scanned across a range of sizes 1295 designed to include two key events, the onset of queuing and the 1296 onset of packet loss or ECN CE marks. The window is scanned by 1297 incrementing it by one packet every 2*target_window_size delivered 1298 packets. This mimics the additive increase phase of standard Reno 1299 TCP congestion avoidance when delayed ACKs are in effect. Normally 1300 the window increases separated by intervals slightly longer than 1301 twice the target_RTT. 1303 There are two ways to implement this test: one built by applying a 1304 window clamp to standard congestion control in a standard protocol 1305 such as TCP and the other built by stiffening a non-standard 1306 transport protocol. When standard congestion control is in effect, 1307 any losses or ECN CE marks cause the transport to revert to a window 1308 smaller than the clamp such that the scanning clamp loses control the 1309 window size. The NPAD pathdiag tool is an example of this class of 1310 algorithms [Pathdiag]. 1312 Alternatively a non-standard congestion control algorithm can respond 1313 to losses by transmitting extra data, such that it maintains the 1314 specified window size independent of losses or ECN CE marks. Such a 1315 stiffened transport explicitly violates mandatory Internet congestion 1316 control [RFC5681] and is not suitable for in situ testing. It is 1317 only appropriate for engineering testing under laboratory conditions. 1318 The Windowed Ping tool implements such a test [WPING]. The tool 1319 described in the paper has been updated.[mpingSource] 1321 The test procedures in Section 8.2 describe how to the partition the 1322 scans into regions and how to interpret the results. 1324 6.4. Concurrent or channelized testing 1326 The procedures described in this document are only directly 1327 applicable to single stream measurement, e.g. one TCP connection or 1328 measurement stream. In an ideal world, we would disallow all 1329 performance claims based multiple concurrent streams, but this is not 1330 practical due to at least two issues. First, many very high rate 1331 link technologies are channelized and at last partially pin the flow 1332 to channel mapping to minimize packet reordering within flows. 1333 Second, TCP itself has scaling limits. Although the former problem 1334 might be overcome through different design decisions, the later 1335 problem is more deeply rooted. 1337 All congestion control algorithms that are philosophically aligned 1338 with the standard [RFC5681] (e.g. claim some level of TCP 1339 compatibility, friendliness or fairness) have scaling limits, in the 1340 sense that as a long fast network (LFN) with a fixed RTT and MTU gets 1341 faster, these congestion control algorithms get less accurate and as 1342 a consequence have difficulty filling the network [CCscaling]. These 1343 properties are a consequence of the original Reno AIMD congestion 1344 control design and the requirement in [RFC5681] that all transport 1345 protocols have similar responses to congestion. 1347 There are a number of reasons to want to specify performance in terms 1348 of multiple concurrent flows, however this approach is not 1349 recommended for data rates below several megabits per second, which 1350 can be attained with run lengths under 10000 packets on many paths. 1351 Since the required run length goes as the square of the data rate, at 1352 higher rates the run lengths can be unreasonably large, and multiple 1353 flows might be the only feasible approach. 1355 If multiple flows are deemed necessary to meet aggregate performance 1356 targets then this must be stated in both the design of the TIDS and 1357 in any claims about network performance. The IP diagnostic tests 1358 must be performed concurrently with the specified number of 1359 connections. For the tests that use bursty test streams, the bursts 1360 should be synchronized across streams unless there is a priori 1361 knowledge that the applications have some explicit mechanism to 1362 stagger their own bursts. In the absences of an explicit mechanism 1363 to stagger bursts many network and application artifacts will 1364 sometimes implicitly synchronize bursts. A test that does not 1365 control burst synchronization may be prone to false pass results for 1366 some applications. 1368 7. Interpreting the Results 1370 7.1. Test outcomes 1372 To perform an exhaustive test of a complete network path, each test 1373 of the TIDS is applied to each subpath of the complete path. If any 1374 subpath fails any test then a standard transport protocol running 1375 over the complete path can also be expected to fail to attain the 1376 Target Transport Performance under some conditions. 1378 In addition to passing or failing, a test can be deemed to be 1379 inconclusive for a number of reasons. Proper instrumentation and 1380 treatment of inconclusive outcomes is critical to the accuracy and 1381 robustness of Model Based Metrics. Tests can be inconclusive if the 1382 precomputed traffic pattern or data rates were not accurately 1383 generated; the measurement results were not statistically 1384 significant; and others causes such as failing to meet some required 1385 preconditions for the test. See Section 5.4 1387 For example consider a test that implements Constant Window Pseudo 1388 CBR (Section 6.2) by adding rate controls and detailed IP packet 1389 transfer instrumentation to TCP (e.g. [RFC4898]). TCP includes 1390 built in control systems which might interfere with the sending data 1391 rate. If such a test meets the required packet transfer statistics 1392 (e.g. run length) while failing to attain the specified data rate it 1393 must be treated as an inconclusive result, because we can not a 1394 priori determine if the reduced data rate was caused by a TCP problem 1395 or a network problem, or if the reduced data rate had a material 1396 effect on the observed packet transfer statistics. 1398 Note that for capacity tests, if the observed packet transfer 1399 statistics meet the statistical criteria for failing (accepting 1400 hypnosis H1 in Section 7.2), the test can can be considered to have 1401 failed because it doesn't really matter that the test didn't attain 1402 the required data rate. 1404 The really important new properties of MBM, such as vantage 1405 independence, are a direct consequence of opening the control loops 1406 in the protocols, such that the test stream does not depend on 1407 network conditions or IP packets received. Any mechanism that 1408 introduces feedback between the path's measurements and the test 1409 stream generation is at risk of introducing nonlinearities that spoil 1410 these properties. Any exceptional event that indicates that such 1411 feedback has happened should cause the test to be considered 1412 inconclusive. 1414 One way to view inconclusive tests is that they reflect situations 1415 where a test outcome is ambiguous between limitations of the network 1416 and some unknown limitation of the IP diagnostic test itself, which 1417 may have been caused by some uncontrolled feedback from the network. 1419 Note that procedures that attempt to search the target parameter 1420 space to find the limits on some parameter such as target_data_rate 1421 are at risk of breaking the location independent properties of Model 1422 Based Metrics, if any part of the boundary between passing and 1423 inconclusive or failing results is sensitive to RTT (which is 1424 normally the case). For example the maximum data rate for a marginal 1425 link (e.g. exhibiting excess errors) is likely to be sensitive to 1426 the test_path_RTT. The maximum observed data rate over the test path 1427 has very little value for predicting the maximum rate over a 1428 different path. 1430 One of the goals for evolving TIDS designs will be to keep sharpening 1431 distinction between inconclusive, passing and failing tests. The 1432 criteria for for passing, failing and inconclusive tests must be 1433 explicitly stated for every test in the TIDS or FS-TIDS. 1435 One of the goals of evolving the testing process, procedures, tools 1436 and measurement point selection should be to minimize the number of 1437 inconclusive tests. 1439 It may be useful to keep raw packet transfer statistics and ancillary 1440 metrics [RFC3148] for deeper study of the behavior of the network 1441 path and to measure the tools themselves. Raw packet transfer 1442 statistics can help to drive tool evolution. Under some conditions 1443 it might be possible to re-evaluate the raw data for satisfying 1444 alternate Target Transport Performance. However it is important to 1445 guard against sampling bias and other implicit feedback which can 1446 cause false results and exhibit measurement point vantage 1447 sensitivity. Simply applying different delivery criteria based on a 1448 different Target Transport Performance is insufficient if the test 1449 traffic patterns (bursts, etc.) does not match the alternate Target 1450 Transport Performance. 1452 7.2. Statistical criteria for estimating run_length 1454 When evaluating the observed run_length, we need to determine 1455 appropriate packet stream sizes and acceptable error levels for 1456 efficient measurement. In practice, can we compare the empirically 1457 estimated packet loss and ECN Congestion Experienced (CE) marking 1458 ratios with the targets as the sample size grows? How large a sample 1459 is needed to say that the measurements of packet transfer indicate a 1460 particular run length is present? 1462 The generalized measurement can be described as recursive testing: 1463 send packets (individually or in patterns) and observe the packet 1464 transfer performance (packet loss ratio or other metric, any marking 1465 we define). 1467 As each packet is sent and measured, we have an ongoing estimate of 1468 the performance in terms of the ratio of packet loss or ECN CE mark 1469 to total packets (i.e. an empirical probability). We continue to 1470 send until conditions support a conclusion or a maximum sending limit 1471 has been reached. 1473 We have a target_mark_probability, 1 mark per target_run_length, 1474 where a "mark" is defined as a lost packet, a packet with ECN CE 1475 mark, or other signal. This constitutes the null Hypothesis: 1477 H0: no more than one mark in target_run_length = 1478 3*(target_window_size)^2 packets 1480 and we can stop sending packets if on-going measurements support 1481 accepting H0 with the specified Type I error = alpha (= 0.05 for 1482 example). 1484 We also have an alternative Hypothesis to evaluate: if performance is 1485 significantly lower than the target_mark_probability. Based on 1486 analysis of typical values and practical limits on measurement 1487 duration, we choose four times the H0 probability: 1489 H1: one or more marks in (target_run_length/4) packets 1491 and we can stop sending packets if measurements support rejecting H0 1492 with the specified Type II error = beta (= 0.05 for example), thus 1493 preferring the alternate hypothesis H1. 1495 H0 and H1 constitute the Success and Failure outcomes described 1496 elsewhere in the memo, and while the ongoing measurements do not 1497 support either hypothesis the current status of measurements is 1498 inconclusive. 1500 The problem above is formulated to match the Sequential Probability 1501 Ratio Test (SPRT) [Wald45] and [Montgomery90]. Note that as 1502 originally framed the events under consideration were all 1503 manufacturing defects. In networking, ECN CE marks and lost packets 1504 are not defects but signals, indicating that the transport protocol 1505 should slow down. 1507 The Sequential Probability Ratio Test also starts with a pair of 1508 hypothesis specified as above: 1510 H0: p0 = one defect in target_run_length 1511 H1: p1 = one defect in target_run_length/4 1512 As packets are sent and measurements collected, the tester evaluates 1513 the cumulative defect count against two boundaries representing H0 1514 Acceptance or Rejection (and acceptance of H1): 1516 Acceptance line: Xa = -h1 + s*n 1517 Rejection line: Xr = h2 + s*n 1519 where n increases linearly for each packet sent and 1521 h1 = { log((1-alpha)/beta) }/k 1522 h2 = { log((1-beta)/alpha) }/k 1523 k = log{ (p1(1-p0)) / (p0(1-p1)) } 1524 s = [ log{ (1-p0)/(1-p1) } ]/k 1526 for p0 and p1 as defined in the null and alternative Hypotheses 1527 statements above, and alpha and beta as the Type I and Type II 1528 errors. 1530 The SPRT specifies simple stopping rules: 1532 o Xa < defect_count(n) < Xr: continue testing 1533 o defect_count(n) <= Xa: Accept H0 1534 o defect_count(n) >= Xr: Accept H1 1536 The calculations above are implemented in the R-tool for Statistical 1537 Analysis [Rtool] , in the add-on package for Cross-Validation via 1538 Sequential Testing (CVST) [CVST]. 1540 Using the equations above, we can calculate the minimum number of 1541 packets (n) needed to accept H0 when x defects are observed. For 1542 example, when x = 0: 1544 Xa = 0 = -h1 + s*n 1545 and n = h1 / s 1547 Note that the derivations in [Wald45] and [Montgomery90] differ. 1548 Montgomery's simplified derivation of SPRT may assume a Bernoulli 1549 processes, where the packet loss probabilities are independent and 1550 identically distributed, making the SPRT more accessible. Wald's 1551 seminal paper showed that this assumption is not necessary. It helps 1552 to remember that the goal of SPRT is not to estimate the value of the 1553 packet loss rate, but only whether or not the packet loss ratio is 1554 likely low enough (when we accept the H0 null hypothesis) yielding 1555 success; or too high (when we accept the H1 alternate hypothesis) 1556 yielding failure. 1558 7.3. Reordering Tolerance 1560 All tests must be instrumented for packet level reordering [RFC4737]. 1561 However, there is no consensus for how much reordering should be 1562 acceptable. Over the last two decades the general trend has been to 1563 make protocols and applications more tolerant to reordering (see for 1564 example [RFC4015]), in response to the gradual increase in reordering 1565 in the network. This increase has been due to the deployment of 1566 technologies such as multithreaded routing lookups and Equal Cost 1567 MultiPath (ECMP) routing. These techniques increase parallelism in 1568 network and are critical to enabling overall Internet growth to 1569 exceed Moore's Law. 1571 Note that transport retransmission strategies can trade off 1572 reordering tolerance vs how quickly they can repair losses vs 1573 overhead from spurious retransmissions. In advance of new 1574 retransmission strategies we propose the following strawman: 1575 Transport protocols should be able to adapt to reordering as long as 1576 the reordering extent is not more than the maximum of one quarter 1577 window or 1 mS, whichever is larger. (These values come from 1578 experience prototyping Early Retransmit [RFC5827] and related 1579 algorithms. They agree with the values being proposed for "RACK: a 1580 time-based fast loss detection algorithm" [I-D.ietf-tcpm-rack].) 1581 Within this limit on reorder extent, there should be no bound on 1582 reordering density. 1584 By implication, recording which is less than these bounds should not 1585 be treated as a network impairment. However [RFC4737] still applies: 1586 reordering should be instrumented and the maximum reordering that can 1587 be properly characterized by the test (because of the bound on 1588 history buffers) should be recorded with the measurement results. 1590 Reordering tolerance and diagnostic limitations, such as the size of 1591 the history buffer used to diagnose packets that are way out-of- 1592 order, must be specified in a FSTIDS. 1594 8. IP Diagnostic Tests 1596 The IP diagnostic tests below are organized by traffic pattern: basic 1597 data rate and packet transfer statistics, standing queues, slowstart 1598 bursts, and sender rate bursts. We also introduce some combined 1599 tests which are more efficient when networks are expected to pass, 1600 but conflate diagnostic signatures when they fail. 1602 There are a number of test details which are not fully defined here. 1603 They must be fully specified in a FS-TIDS. From a standardization 1604 perspective, this lack of specificity will weaken this version of 1605 Model Based Metrics, however it is anticipated that this weakness is 1606 than offset by the extent to which MBM suppresses the problems caused 1607 by using transport protocols for measurement. e.g. non-specific MBM 1608 metrics are likely to have better repeatability than many existing 1609 BTC like metrics. Once we have good field experience, the missing 1610 details can be fully specified. 1612 8.1. Basic Data Rate and Packet Transfer Tests 1614 We propose several versions of the basic data rate and packet 1615 transfer statistics test. All measure the number of packets 1616 delivered between losses or ECN Congestion Experienced (CE) marks, 1617 using a data stream that is rate controlled at approximately the 1618 target_data_rate. 1620 The tests below differ in how the data rate is controlled. The data 1621 can be paced on a timer, or window controlled (and self clocked). 1622 The first two tests implicitly confirm that sub_path has sufficient 1623 raw capacity to carry the target_data_rate. They are recommended for 1624 relatively infrequent testing, such as an installation or periodic 1625 auditing process. The third, background packet transfer statistics, 1626 is a low rate test designed for ongoing monitoring for changes in 1627 subpath quality. 1629 All rely on the data receiver accumulating packet transfer statistics 1630 as described in Section 7.2 to score the outcome: 1632 Pass: it is statistically significant that the observed interval 1633 between losses or ECN CE marks is larger than the target_run_length. 1635 Fail: it is statistically significant that the observed interval 1636 between losses or ECN CE marks is smaller than the target_run_length. 1638 A test is considered to be inconclusive if it failed to generate the 1639 data rate as specified below, meet the qualifications defined in 1640 Section 5.4 or neither run length statistical hypothesis was 1641 confirmed in the allotted test duration. 1643 8.1.1. Delivery Statistics at Paced Full Data Rate 1645 Confirm that the observed run length is at least the 1646 target_run_length while relying on timer to send data at the 1647 target_rate using the procedure described in in Section 6.1 with a 1648 burst size of 1 (single packets) or 2 (packet pairs). 1650 The test is considered to be inconclusive if the packet transmission 1651 can not be accurately controlled for any reason. 1653 RFC 6673 [RFC6673] is appropriate for measuring packet transfer 1654 statistics at full data rate. 1656 8.1.2. Delivery Statistics at Full Data Windowed Rate 1658 Confirm that the observed run length is at least the 1659 target_run_length while sending at an average rate approximately 1660 equal to the target_data_rate, by controlling (or clamping) the 1661 window size of a conventional transport protocol to test_window. 1663 Since losses and ECN CE marks cause transport protocols to reduce 1664 their data rates, this test is expected to be less precise about 1665 controlling its data rate. It should not be considered inconclusive 1666 as long as at least some of the round trips reached the full 1667 target_data_rate without incurring losses or ECN CE marks. To pass 1668 this test the network must deliver target_window_size packets in 1669 target_RTT time without any losses or ECN CE marks at least once per 1670 two target_window_size round trips, in addition to meeting the run 1671 length statistical test. 1673 8.1.3. Background Packet Transfer Statistics Tests 1675 The background run length is a low rate version of the target target 1676 rate test above, designed for ongoing lightweight monitoring for 1677 changes in the observed subpath run length without disrupting users. 1678 It should be used in conjunction with one of the above full rate 1679 tests because it does not confirm that the subpath can support raw 1680 data rate. 1682 RFC 6673 [RFC6673] is appropriate for measuring background packet 1683 transfer statistics. 1685 8.2. Standing Queue Tests 1687 These engineering tests confirm that the bottleneck is well behaved 1688 across the onset of packet loss, which typically follows after the 1689 onset of queuing. Well behaved generally means lossless for 1690 transient queues, but once the queue has been sustained for a 1691 sufficient period of time (or reaches a sufficient queue depth) there 1692 should be a small number of losses or ECN CE marks to signal to the 1693 transport protocol that it should reduce its window or data rate. 1694 Losses that are too early can prevent the transport from averaging at 1695 the target_data_rate. Losses that are too late indicate that the 1696 queue might not have an appropriate AQM [RFC7567] and as a 1697 consequence subject to bufferbloat [wikiBloat]. Queues without AQM 1698 have the potential to inflict excess delays on all flows sharing the 1699 bottleneck. Excess losses (more than half of the window) at the 1700 onset of loss make loss recovery problematic for the transport 1701 protocol. Non-linear, erratic or excessive RTT increases suggest 1702 poor interactions between the channel acquisition algorithms and the 1703 transport self clock. All of the tests in this section use the same 1704 basic scanning algorithm, described here, but score the link or 1705 subpath on the basis of how well it avoids each of these problems. 1707 Some network technologies rely on virtual queues or other techniques 1708 to meter traffic without adding any queuing delay, in which case the 1709 data rate will vary with the window size all the way up to the onset 1710 of load induced packet loss or ECN CE marks. For these technologies, 1711 the discussion of queuing in Section 6.3 does not apply, but it is 1712 still necessary to confirm that the onset of losses or ECN CE marks 1713 be at an appropriate point and progressive. If the network 1714 bottleneck does not introduce significant queuing delay, modify the 1715 procedure described in Section 6.3 to start the scan at a window 1716 equal to or slightly smaller than the test_window. 1718 Use the procedure in Section 6.3 to sweep the window across the onset 1719 of queuing and the onset of loss. The tests below all assume that 1720 the scan emulates standard additive increase and delayed ACK by 1721 incrementing the window by one packet for every 2*target_window_size 1722 packets delivered. A scan can typically be divided into three 1723 regions: below the onset of queuing, a standing queue, and at or 1724 beyond the onset of loss. 1726 Below the onset of queuing the RTT is typically fairly constant, and 1727 the data rate varies in proportion to the window size. Once the data 1728 rate reaches the subpath IP rate, the data rate becomes fairly 1729 constant, and the RTT increases in proportion to the increase in 1730 window size. The precise transition across the start of queuing can 1731 be identified by the maximum network power, defined to be the ratio 1732 data rate over the RTT. The network power can be computed at each 1733 window size, and the window with the maximum is taken as the start of 1734 the queuing region. 1736 If there is random background loss (e.g. bit errors, etc), precise 1737 determination of the onset of queue induced packet loss may require 1738 multiple scans. Above the onset of queuing loss, all transport 1739 protocols are expected to experience periodic losses determined by 1740 the interaction between the congestion control and AQM algorithms. 1741 For standard congestion control algorithms the periodic losses are 1742 likely to be relatively widely spaced and the details are typically 1743 dominated by the behavior of the transport protocol itself. For the 1744 stiffened transport protocols case (with non-standard, aggressive 1745 congestion control algorithms) the details of periodic losses will be 1746 dominated by how the window increase function responds to loss. 1748 8.2.1. Congestion Avoidance 1750 A subpath passes the congestion avoidance standing queue test if more 1751 than target_run_length packets are delivered between the onset of 1752 queuing (as determined by the window with the maximum network power 1753 as described above) and the first loss or ECN CE mark. If this test 1754 is implemented using a standard congestion control algorithm with a 1755 clamp, it can be performed in situ in the production internet as a 1756 capacity test. For an example of such a test see [Pathdiag]. 1758 For technologies that do not have conventional queues, use the 1759 test_window in place of the onset of queuing. i.e. A subpath passes 1760 the congestion avoidance standing queue test if more than 1761 target_run_length packets are delivered between start of the scan at 1762 test_window and the first loss or ECN CE mark. 1764 8.2.2. Bufferbloat 1766 This test confirms that there is some mechanism to limit buffer 1767 occupancy (e.g. that prevents bufferbloat). Note that this is not 1768 strictly a requirement for single stream bulk transport capacity, 1769 however if there is no mechanism to limit buffer queue occupancy then 1770 a single stream with sufficient data to deliver is likely to cause 1771 the problems described in [RFC7567], and [wikiBloat]. This may cause 1772 only minor symptoms for the dominant flow, but has the potential to 1773 make the subpath unusable for other flows and applications. 1775 Pass if the onset of loss occurs before a standing queue has 1776 introduced more delay than than twice target_RTT, or other well 1777 defined and specified limit. Note that there is not yet a model for 1778 how much standing queue is acceptable. The factor of two chosen here 1779 reflects a rule of thumb. In conjunction with the previous test, 1780 this test implies that the first loss should occur at a queuing delay 1781 which is between one and two times the target_RTT. 1783 Specified RTT limits that are larger than twice the target_RTT must 1784 be fully justified in the FS-TIDS. 1786 8.2.3. Non excessive loss 1788 This test confirms that the onset of loss is not excessive. Pass if 1789 losses are equal or less than the increase in the cross traffic plus 1790 the test stream window increase since the previous RTT. This could 1791 be restated as non-decreasing total throughput of the subpath at the 1792 onset of loss. (Note that when there is a transient drop in subpath 1793 throughput and there is not already a standing queue, a subpath that 1794 passes other queue tests in this document will have sufficient queue 1795 space to hold one full RTT worth of data). 1797 Note that token bucket policers will not pass this test, which is as 1798 intended. TCP often stumbles badly if more than a small fraction of 1799 the packets are dropped in one RTT. Many TCP implementations will 1800 require a timeout and slowstart to recover their self clock. Even if 1801 they can recover from the massive losses the sudden change in 1802 available capacity at the bottleneck wastes serving and front path 1803 capacity until TCP can adapt to the new rate [Policing]. 1805 8.2.4. Duplex Self Interference 1807 This engineering test confirms a bound on the interactions between 1808 the forward data path and the ACK return path when they share a half 1809 duplex link. 1811 Some historical half duplex technologies had the property that each 1812 direction held the channel until it completely drained its queue. 1813 When a self clocked transport protocol, such as TCP, has data and 1814 ACKs passing in opposite directions through such a link, the behavior 1815 often reverts to stop-and-wait. Each additional packet added to the 1816 window raises the observed RTT by two packet times, once as the 1817 additional packet passes through the data path, and once for the 1818 additional delay incurred by the ACK waiting on the return path. 1820 The duplex self interference test fails if the RTT rises by more than 1821 a fixed bound above the expected queuing time computed from the 1822 excess window divided by the subpath IP Capacity. This bound must be 1823 smaller than target_RTT/2 to avoid reverting to stop and wait 1824 behavior. (e.g. Data packets and ACKs both have to be released at 1825 least twice per RTT.) 1827 8.3. Slowstart tests 1829 These tests mimic slowstart: data is sent at twice the effective 1830 bottleneck rate to exercise the queue at the dominant bottleneck. 1832 8.3.1. Full Window slowstart test 1834 This is a capacity test to confirm that slowstart is not likely to 1835 exit prematurely. Send slowstart bursts that are target_window_size 1836 total packets. 1838 Accumulate packet transfer statistics as described in Section 7.2 to 1839 score the outcome. Pass if it is statistically significant that the 1840 observed number of good packets delivered between losses or ECN CE 1841 marks is larger than the target_run_length. Fail if it is 1842 statistically significant that the observed interval between losses 1843 or ECN CE marks is smaller than the target_run_length. 1845 It is deemed inconclusive if the elapsed time to send the data burst 1846 is not less than half of the time to receive the ACKs. (i.e. It is 1847 acceptable to send data too fast, but sending it slower than twice 1848 the actual bottleneck rate as indicated by the ACKs is deemed 1849 inconclusive). The headway for the slowstart bursts should be the 1850 target_RTT. 1852 Note that these are the same parameters as the Sender Full Window 1853 burst test, except the burst rate is at slowstart rate, rather than 1854 sender interface rate. 1856 8.3.2. Slowstart AQM test 1858 Do a continuous slowstart (send data continuously at twice the 1859 implied IP bottleneck capacity), until the first loss, stop, allow 1860 the network to drain and repeat, gathering statistics on how many 1861 packets were delivered before the loss, the pattern of losses, 1862 maximum observed RTT and window size. Justify the results. There is 1863 not currently sufficient theory justifying requiring any particular 1864 result, however design decisions that affect the outcome of this 1865 tests also affect how the network balances between long and short 1866 flows (the "mice vs elephants" problem). The queue sojourn time for 1867 the first packet delivered after the first loss should be at least 1868 one half of the target_RTT. 1870 This is an engineering test: It should be performed on a quiescent 1871 network or testbed, since cross traffic has the potential to change 1872 the results in ill defined ways. 1874 8.4. Sender Rate Burst tests 1876 These tests determine how well the network can deliver bursts sent at 1877 sender's interface rate. Note that this test most heavily exercises 1878 the front path, and is likely to include infrastructure may be out of 1879 scope for an access ISP, even though the bursts might be caused by 1880 ACK compression, thinning or channel arbitration in the access ISP. 1881 See Appendix B. 1883 Also, there are a several details about sender interface rate bursts 1884 that are not fully defined here. These details, such as the assumed 1885 sender interface rate, should be explicitly stated is a FS-TIDS. 1887 Current standards permit TCP to send full window bursts following an 1888 application pause. (Congestion Window Validation [RFC2861] and 1889 updates to support Rate-Limited Traffic [RFC7661], are not required). 1890 Since full window bursts are consistent with standard behavior, it is 1891 desirable that the network be able to deliver such bursts, otherwise 1892 application pauses will cause unwarranted losses. Note that the AIMD 1893 sawtooth requires a peak window that is twice target_window_size, so 1894 the worst case burst may be 2*target_window_size. 1896 It is also understood in the application and serving community that 1897 interface rate bursts have a cost to the network that has to be 1898 balanced against other costs in the servers themselves. For example 1899 TCP Segmentation Offload (TSO) reduces server CPU in exchange for 1900 larger network bursts, which increase the stress on network buffer 1901 memory. Some newer TCP implementations can pace traffic at scale 1902 [TSO_pacing][TSO_fq_pacing]. It remains to be determined if and how 1903 quickly these changes will be deployed. 1905 There is not yet theory to unify these costs or to provide a 1906 framework for trying to optimize global efficiency. We do not yet 1907 have a model for how much server rate bursts should be tolerated by 1908 the network. Some bursts must be tolerated by the network, but it is 1909 probably unreasonable to expect the network to be able to efficiently 1910 deliver all data as a series of bursts. 1912 For this reason, this is the only test for which we encourage 1913 derating. A TIDS could include a table of pairs of derating 1914 parameters: burst sizes and how much each burst size is permitted to 1915 reduce the run length, relative to to the target_run_length. 1917 8.5. Combined and Implicit Tests 1919 Combined tests efficiently confirm multiple network properties in a 1920 single test, possibly as a side effect of normal content delivery. 1921 They require less measurement traffic than other testing strategies 1922 at the cost of conflating diagnostic signatures when they fail. 1923 These are by far the most efficient for monitoring networks that are 1924 nominally expected to pass all tests. 1926 8.5.1. Sustained Bursts Test 1928 The sustained burst test implements a combined worst case version of 1929 all of the capacity tests above. It is simply: 1931 Send target_window_size bursts of packets at server interface rate 1932 with target_RTT burst headway (burst start to next burst start). 1933 Verify that the observed packet transfer statistics meets the 1934 target_run_length. 1936 Key observations: 1938 o The subpath under test is expected to go idle for some fraction of 1939 the time, determined by the difference between the time to drain 1940 the queue at the subpath_IP_capacity, and the target_RTT. If the 1941 queue does not drain completely it may be an indication that the 1942 the subpath has insufficient IP capacity or that there is some 1943 other problem with the test (e.g. inconclusive). 1944 o The burst sensitivity can be derated by sending smaller bursts 1945 more frequently. E.g. send target_window_size*derate packet 1946 bursts every target_RTT*derate, where "derate" is less than one. 1947 o When not derated, this test is the most strenuous capacity test. 1948 o A subpath that passes this test is likely to be able to sustain 1949 higher rates (close to subpath_IP_capacity) for paths with RTTs 1950 significantly smaller than the target_RTT. 1951 o This test can be implemented with instrumented TCP [RFC4898], 1952 using a specialized measurement application at one end [MBMSource] 1953 and a minimal service at the other end [RFC0863] [RFC0864]. 1954 o This test is efficient to implement, since it does not require 1955 per-packet timers, and can make use of TSO in modern NIC hardware. 1956 o If a subpath is known to pass the Standing Queue engineering tests 1957 (particularly that it has a progressive onset of loss at an 1958 appropriate queue depth), then the Sustained Burst Test is 1959 sufficient to assure that the subpath under test will not impair 1960 Bulk Transport Capacity at the target performance under all 1961 conditions. See Section 8.2 for a discussion of the standing 1962 queue tests. 1964 Note that this test is clearly independent of the subpath RTT, or 1965 other details of the measurement infrastructure, as long as the 1966 measurement infrastructure can accurately and reliably deliver the 1967 required bursts to the subpath under test. 1969 8.5.2. Passive Measurements 1971 Any non-throughput maximizing application, such as fixed rate 1972 streaming media, can be used to implement passive or hybrid (defined 1973 in [RFC7799]) versions of Model Based Metrics with some additional 1974 instrumentation and possibly a traffic shaper or other controls in 1975 the servers. The essential requirement is that the data transmission 1976 be constrained such that even with arbitrary application pauses and 1977 bursts, the data rate and burst sizes stay within the envelope 1978 defined by the individual tests described above. 1980 If the application's serving data rate can be constrained to be less 1981 than or equal to the target_data_rate and the serving_RTT (the RTT 1982 between the sender and client) is less than the target_RTT, this 1983 constraint is most easily implemented by clamping the transport 1984 window size to serving_window_clamp, set to the test_window, computed 1985 for the actual serving path. 1987 Under the above constraints the serving_window_clamp will limit the 1988 both the serving data rate and burst sizes to be no larger than the 1989 procedures in Section 8.1.2 and Section 8.4 or Section 8.5.1. Since 1990 the serving RTT is smaller than the target_RTT, the worst case bursts 1991 that might be generated under these conditions will be smaller than 1992 called for by Section 8.4 and the sender rate burst sizes are 1993 implicitly derated by the serving_window_clamp divided by the 1994 target_window_size at the very least. (Depending on the application 1995 behavior, the data might be significantly smoother than specified by 1996 any of the burst tests.) 1998 In an alternative implementation the data rate and bursts might be 1999 explicitly controlled by a programmable traffic shaper or pacing at 2000 the sender. This would provide better control over transmissions but 2001 is more complicated to implement, although the required technology is 2002 available [TSO_pacing][TSO_fq_pacing]. 2004 Note that these techniques can be applied to any content delivery 2005 that can operated at a constrained data rate to inhibit TCP 2006 equilibrium behavior. 2008 Furthermore note that Dynamic Adaptive Streaming over HTTP (DASH) is 2009 generally in conflict with passive Model Based Metrics measurement, 2010 because it is a rate maximizing protocol. It can still meet the 2011 requirement here if the rate can be capped, for example by knowing a 2012 priori the maximum rate needed to deliver a particular piece of 2013 content. 2015 9. An Example 2017 In this section we illustrate a TIDS designed to confirm that an 2018 access ISP can reliably deliver HD video from multiple content 2019 providers to all of their customers. With modern codecs, minimal HD 2020 video (720p) generally fits in 2.5 Mb/s. Due to their geographical 2021 size, network topology and modem characteristics the ISP determines 2022 that most content is within a 50 mS RTT of their users (This example 2023 RTT is a sufficient to cover the propagation delay to continental 2024 Europe or either US coast with low delay modems or somewhat smaller 2025 geographical regions if the modems require additional delay to 2026 implement advanced compression and error recovery). 2028 2.5 Mb/s over a 50 ms path 2030 +----------------------+-------+---------+ 2031 | End-to-End Parameter | value | units | 2032 +----------------------+-------+---------+ 2033 | target_rate | 2.5 | Mb/s | 2034 | target_RTT | 50 | ms | 2035 | target_MTU | 1500 | bytes | 2036 | header_overhead | 64 | bytes | 2037 | | | | 2038 | target_window_size | 11 | packets | 2039 | target_run_length | 363 | packets | 2040 +----------------------+-------+---------+ 2042 Table 1 2044 Table 1 shows the default TCP model with no derating, and as such is 2045 quite conservative. The simplest TIDS would be to use the sustained 2046 burst test, described in Section 8.5.1. Such a test would send 11 2047 packet bursts every 50mS, and confirming that there was no more than 2048 1 packet loss per 33 bursts (363 total packets in 1.650 seconds). 2050 Since this number represents is the entire end-to-end loss budget, 2051 independent subpath tests could be implemented by apportioning the 2052 packet loss ratio across subpaths. For example 50% of the losses 2053 might be allocated to the access or last mile link to the user, 40% 2054 to the network interconnections with other ISPs and 1% to each 2055 internal hop (assuming no more than 10 internal hops). Then all of 2056 the subpaths can be tested independently, and the spatial composition 2057 of passing subpaths would be expected to be within the end-to-end 2058 loss budget. 2060 9.1. Observations about applicability 2062 Guidance on deploying and using MBM belong in a future document. 2063 However this example illustrates some the issues that may need to be 2064 considered. 2066 Note that another ISP, with different geographical coverage, topology 2067 or modem technology may need to assume a different target_RTT, and as 2068 a consequence different target_window_size and target_run_length, 2069 even for the same target_data rate. One of the implications of this 2070 is that infrastructure shared by multiple ISPs, such as inter- 2071 exchange points (IXPs) and other interconnects may need to be 2072 evaluated on the basis of the most stringent target_window_size and 2073 target_run_length of any participating ISP. One way to do this might 2074 be to choose target parameters for evaluating such shared 2075 infrastructure on the basis of a hypothetical reference path that 2076 does not necessarily match any actual paths. 2078 Testing interconnects has generally been problematic: conventional 2079 performance tests run between measurement points adjacent to either 2080 side of the interconnect are not generally useful. Unconstrained TCP 2081 tests, such as iPerf [iPerf] are usually overly aggressive due to the 2082 small RTT (often less than 1 mS). With a short RTT these tools are 2083 likely to report inflated data rates because on a short RTT these 2084 tools can tolerate very high packet loss ratios and can push other 2085 cross traffic off of the network. As a consequence these 2086 measurements are useless for predicting actual user performance over 2087 longer paths, and may themselves be quite disruptive. Model Based 2088 Metrics solves this problem. The interconnect can be evaluated with 2089 the same TIDS as other subpaths. Continuing our example, if the 2090 interconnect is apportioned 40% of the losses, 11 packet bursts sent 2091 every 50mS should have fewer than one loss per 82 bursts (902 2092 packets). 2094 10. Validation 2096 Since some aspects of the models are likely to be too conservative, 2097 Section 5.2 permits alternate protocol models and Section 5.3 permits 2098 test parameter derating. If either of these techniques are used, we 2099 require demonstrations that such a TIDS can robustly detect subpaths 2100 that will prevent authentic applications using state-of-the-art 2101 protocol implementations from meeting the specified Target Transport 2102 Performance. This correctness criteria is potentially difficult to 2103 prove, because it implicitly requires validating a TIDS against all 2104 possible paths and subpaths. The procedures described here are still 2105 experimental. 2107 We suggest two approaches, both of which should be applied: first, 2108 publish a fully open description of the TIDS, including what 2109 assumptions were used and and how it was derived, such that the 2110 research community can evaluate the design decisions, test them and 2111 comment on their applicability; and second, demonstrate that 2112 applications do meet the Target Transport Performance when running 2113 over a network testbed which has the tightest possible constraints 2114 that still allow the tests in the TIDS to pass. 2116 This procedure resembles an epsilon-delta proof in calculus. 2117 Construct a test network such that all of the individual tests of the 2118 TIDS pass by only small (infinitesimal) margins, and demonstrate that 2119 a variety of authentic applications running over real TCP 2120 implementations (or other protocols as appropriate) meets the Target 2121 Transport Performance over such a network. The workloads should 2122 include multiple types of streaming media and transaction oriented 2123 short flows (e.g. synthetic web traffic). 2125 For example, for the HD streaming video TIDS described in Section 9, 2126 the IP capacity should be exactly the header_overhead above 2.5 Mb/s, 2127 the per packet random background loss ratio should be 1/363, for a 2128 run length of 363 packets, the bottleneck queue should be 11 packets 2129 and the front path should have just enough buffering to withstand 11 2130 packet interface rate bursts. We want every one of the TIDS tests to 2131 fail if we slightly increase the relevant test parameter, so for 2132 example sending a 12 packet burst should cause excess (possibly 2133 deterministic) packet drops at the dominant queue at the bottleneck. 2134 This network has the tightest possible constraints that can be 2135 expected to pass the TIDS, yet it should be possible for a real 2136 application using a stock TCP implementation in the vendor's default 2137 configuration to attain 2.5 Mb/s over an 50 mS path. 2139 The most difficult part of setting up such a testbed is arranging for 2140 it to have the tightest possible constraints that still allow it to 2141 pass the individual tests. Two approaches are suggested: 2142 constraining (configuring) the network devices not to use all 2143 available resources (e.g. by limiting available buffer space or data 2144 rate); and pre-loading subpaths with cross traffic. Note that is it 2145 important that a single tightly constrained environment just barely 2146 passes all tests, otherwise there is a chance that TCP can exploit 2147 extra latitude in some parameters (such as data rate) to partially 2148 compensate for constraints in other parameters (queue space, or vice- 2149 versa). 2151 To the extent that a TIDS is used to inform public dialog it should 2152 be fully publicly documented, including the details of the tests, 2153 what assumptions were used and how it was derived. All of the 2154 details of the validation experiment should also be published with 2155 sufficient detail for the experiments to be replicated by other 2156 researchers. All components should either be open source of fully 2157 described proprietary implementations that are available to the 2158 research community. 2160 11. Security Considerations 2162 Measurement is often used to inform business and policy decisions, 2163 and as a consequence is potentially subject to manipulation. Model 2164 Based Metrics are expected to be a huge step forward because 2165 equivalent measurements can be performed from multiple vantage 2166 points, such that performance claims can be independently validated 2167 by multiple parties. 2169 Much of the acrimony in the Net Neutrality debate is due to the 2170 historical lack of any effective vantage independent tools to 2171 characterize network performance. Traditional methods for measuring 2172 Bulk Transport Capacity are sensitive to RTT and as a consequence 2173 often yield very different results when run local to an ISP or 2174 interconnect and when run over a customer's complete path. Neither 2175 the ISP nor customer can repeat the others measurements, leading to 2176 high levels of distrust and acrimony. Model Based Metrics are 2177 expected to greatly improve this situation. 2179 Note that in situ measurements sometimes requires sending synthetic 2180 measurement traffic between arbitrary locations in the network, and 2181 as such are potentially attractive platforms for launching DDOS 2182 attacks. All active measurement tools and protocols must be deigned 2183 to minimize the opportunities for these misuses. See the discussion 2184 in section 7 of [RFC7594]. 2186 This document only describes a framework for designing Fully 2187 Specified Targeted IP Diagnostic Suite. Each FS-TIDS must include 2188 its own security section. 2190 12. Acknowledgments 2192 Ganga Maguluri suggested the statistical test for measuring loss 2193 probability in the target run length. Alex Gilgur and Merry Mou for 2194 helping with the statistics. 2196 Meredith Whittaker for improving the clarity of the communications. 2198 Ruediger Geib provided feedback which greatly improved the document. 2200 This work was inspired by Measurement Lab: open tools running on an 2201 open platform, using open tools to collect open data. See 2202 http://www.measurementlab.net/ 2204 13. IANA Considerations 2206 This document has no actions for IANA. 2208 14. References 2210 [RFC0863] Postel, J., "Discard Protocol", STD 21, RFC 863, May 1983. 2212 [RFC0864] Postel, J., "Character Generator Protocol", STD 22, 2213 RFC 864, May 1983. 2215 [RFC2330] Paxson, V., Almes, G., Mahdavi, J., and M. Mathis, 2216 "Framework for IP Performance Metrics", RFC 2330, May 2217 1998. 2219 [RFC2861] Handley, M., Padhye, J., and S. Floyd, "TCP Congestion 2220 Window Validation", RFC 2861, June 2000. 2222 [RFC3148] Mathis, M. and M. Allman, "A Framework for Defining 2223 Empirical Bulk Transfer Capacity Metrics", RFC 3148, July 2224 2001. 2226 [RFC3465] Allman, M., "TCP Congestion Control with Appropriate Byte 2227 Counting (ABC)", RFC 3465, February 2003. 2229 [RFC4015] Ludwig, R. and A. Gurtov, "The Eifel Response Algorithm 2230 for TCP", RFC 4015, February 2005. 2232 [RFC4737] Morton, A., Ciavattone, L., Ramachandran, G., Shalunov, 2233 S., and J. Perser, "Packet Reordering Metrics", RFC 4737, 2234 November 2006. 2236 [RFC4898] Mathis, M., Heffner, J., and R. Raghunarayan, "TCP 2237 Extended Statistics MIB", RFC 4898, May 2007. 2239 [RFC5136] Chimento, P. and J. Ishac, "Defining Network Capacity", 2240 RFC 5136, February 2008. 2242 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion 2243 Control", RFC 5681, September 2009. 2245 [RFC5827] Allman, M., Avrachenkov, K., Ayesta, U., Blanton, J., and 2246 P. Hurtig, "Early Retransmit for TCP and Stream Control 2247 Transmission Protocol (SCTP)", RFC 5827, 2248 DOI 10.17487/RFC5827, May 2010, 2249 . 2251 [RFC5835] Morton, A. and S. Van den Berghe, "Framework for Metric 2252 Composition", RFC 5835, April 2010. 2254 [RFC6049] Morton, A. and E. Stephan, "Spatial Composition of 2255 Metrics", RFC 6049, January 2011. 2257 [RFC6576] Geib, R., Ed., Morton, A., Fardid, R., and A. Steinmitz, 2258 "IP Performance Metrics (IPPM) Standard Advancement 2259 Testing", BCP 176, RFC 6576, DOI 10.17487/RFC6576, March 2260 2012, . 2262 [RFC6673] Morton, A., "Round-Trip Packet Loss Metrics", RFC 6673, 2263 August 2012. 2265 [RFC6928] Chu, J., Dukkipati, N., Cheng, Y., and M. Mathis, 2266 "Increasing TCP's Initial Window", RFC 6928, 2267 DOI 10.17487/RFC6928, April 2013, 2268 . 2270 [RFC7312] Fabini, J. and A. Morton, "Advanced Stream and Sampling 2271 Framework for IP Performance Metrics (IPPM)", RFC 7312, 2272 August 2014. 2274 [RFC7398] Bagnulo, M., Burbridge, T., Crawford, S., Eardley, P., and 2275 A. Morton, "A Reference Path and Measurement Points for 2276 Large-Scale Measurement of Broadband Performance", 2277 RFC 7398, February 2015. 2279 [RFC7567] Baker, F., Ed. and G. Fairhurst, Ed., "IETF 2280 Recommendations Regarding Active Queue Management", 2281 BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015, 2282 . 2284 [RFC7594] Eardley, P., Morton, A., Bagnulo, M., Burbridge, T., 2285 Aitken, P., and A. Akhter, "A Framework for Large-Scale 2286 Measurement of Broadband Performance (LMAP)", RFC 7594, 2287 DOI 10.17487/RFC7594, September 2015, 2288 . 2290 [RFC7661] Fairhurst, G., Sathiaseelan, A., and R. Secchi, "Updating 2291 TCP to Support Rate-Limited Traffic", RFC 7661, 2292 DOI 10.17487/RFC7661, October 2015, 2293 . 2295 [RFC7680] Almes, G., Kalidindi, S., Zekauskas, M., and A. Morton, 2296 Ed., "A One-Way Loss Metric for IP Performance Metrics 2297 (IPPM)", STD 82, RFC 7680, DOI 10.17487/RFC7680, January 2298 2016, . 2300 [RFC7799] Morton, A., "Active and Passive Metrics and Methods (with 2301 Hybrid Types In-Between)", RFC 7799, DOI 10.17487/RFC7799, 2302 May 2016, . 2304 [I-D.ietf-tcpm-rack] 2305 Cheng, Y., Cardwell, N., and N. Dukkipati, "RACK: a time- 2306 based fast loss detection algorithm for TCP", draft-ietf- 2307 tcpm-rack-02 (work in progress), March 2017. 2309 [MSMO97] Mathis, M., Semke, J., Mahdavi, J., and T. Ott, "The 2310 Macroscopic Behavior of the TCP Congestion Avoidance 2311 Algorithm", Computer Communications Review volume 27, 2312 number3, July 1997. 2314 [WPING] Mathis, M., "Windowed Ping: An IP Level Performance 2315 Diagnostic", INET 94, June 1994. 2317 [mpingSource] 2318 Fan, X., Mathis, M., and D. Hamon, "Git Repository for 2319 mping: An IP Level Performance Diagnostic", Sept 2013, 2320 . 2322 [MBMSource] 2323 Hamon, D., Stuart, S., and H. Chen, "Git Repository for 2324 Model Based Metrics", Sept 2013, . 2327 [Pathdiag] 2328 Mathis, M., Heffner, J., O'Neil, P., and P. Siemsen, 2329 "Pathdiag: Automated TCP Diagnosis", Passive and Active 2330 Measurement , June 2008. 2332 [iPerf] Wikipedia Contributors, , "iPerf", Wikipedia, The Free 2333 Encyclopedia , cited March 2015, 2334 . 2337 [Wald45] Wald, A., "Sequential Tests of Statistical Hypotheses", 2338 The Annals of Mathematical Statistics, Vol. 16, No. 2, pp. 2339 117-186, Published by: Institute of Mathematical 2340 Statistics, Stable URL: 2341 http://www.jstor.org/stable/2235829, June 1945. 2343 [Montgomery90] 2344 Montgomery, D., "Introduction to Statistical Quality 2345 Control - 2nd ed.", ISBN 0-471-51988-X, 1990. 2347 [Rtool] R Development Core Team, , "R: A language and environment 2348 for statistical computing. R Foundation for Statistical 2349 Computing, Vienna, Austria. ISBN 3-900051-07-0, URL 2350 http://www.R-project.org/", , 2011. 2352 [CVST] Krueger, T. and M. Braun, "R package: Fast Cross- 2353 Validation via Sequential Testing", version 0.1, 11 2012. 2355 [AFD] Pan, R., Breslau, L., Prabhakar, B., and S. Shenker, 2356 "Approximate fairness through differential dropping", 2357 SIGCOMM Comput. Commun. Rev. 33, 2, April 2003. 2359 [wikiBloat] 2360 Wikipedia, , "Bufferbloat", http://en.wikipedia.org/ 2361 w/ index.php?title=Bufferbloat&oldid=608805474, March 2362 2015. 2364 [CCscaling] 2365 Fernando, F., Doyle, J., and S. Steven, "Scalable laws for 2366 stable network congestion control", Proceedings of 2367 Conference on Decision and 2368 Control, http://www.ee.ucla.edu/~paganini, December 2001. 2370 [TSO_pacing] 2371 Corbet, J., "TSO sizing and the FQ scheduler", 2372 LWN.net https://lwn.net/Articles/564978/, Aug 2013. 2374 [TSO_fq_pacing] 2375 Dumazet, E. and Y. Chen, "TSO, fair queuing, pacing: 2376 three's a charm", Proceedings of IETF 88, TCPM WG 2377 https://www.ietf.org/proceedings/88/slides/slides-88-tcpm- 2378 9.pdf, Nov 2013. 2380 [Policing] 2381 Flach, T., Papageorge, P., Terzis, A., Pedrosa, L., Cheng, 2382 Y., Karim, T., Katz-Bassett, E., and R. Govindan, "An 2383 Internet-Wide Analysis of Traffic Policing", ACM SIGCOMM , 2384 August 2016. 2386 Appendix A. Model Derivations 2388 The reference target_run_length described in Section 5.2 is based on 2389 very conservative assumptions: that all excess data in flight 2390 (window) above the target_window_size contributes to a standing queue 2391 that raises the RTT, and that classic Reno congestion control with 2392 delayed ACKs are in effect. In this section we provide two 2393 alternative calculations using different assumptions. 2395 It may seem out of place to allow such latitude in a measurement 2396 method, but this section provides offsetting requirements. 2398 The estimates provided by these models make the most sense if network 2399 performance is viewed logarithmically. In the operational Internet, 2400 data rates span more than 8 orders of magnitude, RTT spans more than 2401 3 orders of magnitude, and packet loss ratio spans at least 8 orders 2402 of magnitude if not more. When viewed logarithmically (as in 2403 decibels), these correspond to 80 dB of dynamic range. On an 80 dB 2404 scale, a 3 dB error is less than 4% of the scale, even though it 2405 represents a factor of 2 in untransformed parameter. 2407 This document gives a lot of latitude for calculating 2408 target_run_length, however people designing a TIDS should consider 2409 the effect of their choices on the ongoing tussle about the relevance 2410 of "TCP friendliness" as an appropriate model for Internet capacity 2411 allocation. Choosing a target_run_length that is substantially 2412 smaller than the reference target_run_length specified in Section 5.2 2413 strengthens the argument that it may be appropriate to abandon "TCP 2414 friendliness" as the Internet fairness model. This gives developers 2415 incentive and permission to develop even more aggressive applications 2416 and protocols, for example by increasing the number of connections 2417 that they open concurrently. 2419 A.1. Queueless Reno 2421 In Section 5.2 models were derived based on the assumption that the 2422 subpath IP rate matches the target rate plus overhead, such that the 2423 excess window needed for the AIMD sawtooth causes a fluctuating queue 2424 at the bottleneck. 2426 An alternate situation would be a bottleneck where there is no 2427 significant queue and losses are caused by some mechanism that does 2428 not involve extra delay, for example by the use of a virtual queue as 2429 done in Approximate Fair Dropping [AFD]. A flow controlled by such a 2430 bottleneck would have a constant RTT and a data rate that fluctuates 2431 in a sawtooth due to AIMD congestion control. Assume the losses are 2432 being controlled to make the average data rate meet some goal which 2433 is equal or greater than the target_rate. The necessary run length 2434 to meet the target_rate can be computed as follows: 2436 For some value of Wmin, the window will sweep from Wmin packets to 2437 2*Wmin packets in 2*Wmin RTT (due to delayed ACK). Unlike the 2438 queuing case where Wmin = target_window_size, we want the average of 2439 Wmin and 2*Wmin to be the target_window_size, so the average data 2440 rate is the target rate. Thus we want Wmin = 2441 (2/3)*target_window_size. 2443 Between losses each sawtooth delivers (1/2)(Wmin+2*Wmin)(2Wmin) 2444 packets in 2*Wmin round trip times. 2446 Substituting these together we get: 2448 target_run_length = (4/3)(target_window_size^2) 2449 Note that this is 44% of the reference_run_length computed earlier. 2450 This makes sense because under the assumptions in Section 5.2 the 2451 AMID sawtooth caused a queue at the bottleneck, which raised the 2452 effective RTT by 50%. 2454 Appendix B. The effects of ACK scheduling 2456 For many network technologies simple queuing models don't apply: the 2457 network schedules, thins or otherwise alters the timing of ACKs and 2458 data, generally to raise the efficiency of the channel allocation 2459 algorithms when confronted with relatively widely spaced small ACKs. 2460 These efficiency strategies are ubiquitous for half duplex, wireless 2461 and broadcast media. 2463 Altering the ACK stream by holding or thinning ACKs typically has two 2464 consequences: it raises the implied bottleneck IP capacity, making 2465 the fine grained slowstart bursts either faster or larger and it 2466 raises the effective RTT by the average time that the ACKs and data 2467 are delayed. The first effect can be partially mitigated by re- 2468 clocking ACKs once they are beyond the bottleneck on the return path 2469 to the sender, however this further raises the effective RTT. 2471 The most extreme example of this sort of behavior would be a half 2472 duplex channel that is not released as long as the endpoint currently 2473 holding the channel has more traffic (data or ACKs) to send. Such 2474 environments cause self clocked protocols under full load to revert 2475 to extremely inefficient stop and wait behavior. The channel 2476 constrains the protocol to send an entire window of data as a single 2477 contiguous burst on the forward path, followed by the entire window 2478 of ACKs on the return path. 2480 If a particular return path contains a subpath or device that alters 2481 the timing of the ACK stream, then the entire front path from the 2482 sender up to the bottleneck must be tested at the burst parameters 2483 implied by the ACK scheduling algorithm. The most important 2484 parameter is the Implied Bottleneck IP Capacity, which is the average 2485 rate at which the ACKs advance snd.una. Note that thinning the ACK 2486 stream (relying on the cumulative nature of seg.ack to permit 2487 discarding some ACKs) causes most TCP implementations to send 2488 interface rate bursts to offset the longer times between ACKs in 2489 order to maintain the average data rate. 2491 Note that due to ubiquitous self clocking in Internet protocols, ill 2492 conceived channel allocation mechanisms are likely to increases the 2493 queuing stress on the front path because they cause larger full 2494 sender rate data bursts. 2496 Holding data or ACKs for channel allocation or other reasons (such as 2497 forward error correction) always raises the effective RTT relative to 2498 the minimum delay for the path. Therefore it may be necessary to 2499 replace target_RTT in the calculation in Section 5.2 by an 2500 effective_RTT, which includes the target_RTT plus a term to account 2501 for the extra delays introduced by these mechanisms. 2503 Appendix C. Version Control 2505 This section to be removed prior to publication. 2507 Formatted: Thu Apr 7 18:12:37 PDT 2016 2509 Authors' Addresses 2511 Matt Mathis 2512 Google, Inc 2513 1600 Amphitheater Parkway 2514 Mountain View, California 94043 2515 USA 2517 Email: mattmathis@google.com 2519 Al Morton 2520 AT&T Labs 2521 200 Laurel Avenue South 2522 Middletown, NJ 07748 2523 USA 2525 Phone: +1 732 420 1571 2526 Email: acmorton@att.com