idnits 2.17.1 draft-ietf-ippm-model-based-metrics-13.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- No issues found here. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year == Line 1567 has weird spacing: '... and n = h1...' -- The document date (September 15, 2017) is 2412 days in the past. Is this intentional? Checking references for intended status: Experimental ---------------------------------------------------------------------------- -- Obsolete informational reference (is this intentional?): RFC 2861 (Obsoleted by RFC 7661) == Outdated reference: A later version (-15) exists of draft-ietf-tcpm-rack-02 Summary: 0 errors (**), 0 flaws (~~), 3 warnings (==), 2 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 IP Performance Working Group M. Mathis 3 Internet-Draft Google, Inc 4 Intended status: Experimental A. Morton 5 Expires: March 19, 2018 AT&T Labs 6 September 15, 2017 8 Model Based Metrics for Bulk Transport Capacity 9 draft-ietf-ippm-model-based-metrics-13.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 mathematical models to specify a Targeted 23 Suite of IP Diagnostic tests, designed to assess whether common 24 transport protocols can be expected to meet a predetermined Target 25 Transport Performance over an Internet path. 27 For Bulk Transport Capacity the 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 Status of This Memo 41 This Internet-Draft is submitted in full conformance with the 42 provisions of BCP 78 and BCP 79. 44 Internet-Drafts are working documents of the Internet Engineering 45 Task Force (IETF). Note that other groups may also distribute 46 working documents as Internet-Drafts. The list of current Internet- 47 Drafts is at http://datatracker.ietf.org/drafts/current/. 49 Internet-Drafts are draft documents valid for a maximum of six months 50 and may be updated, replaced, or obsoleted by other documents at any 51 time. It is inappropriate to use Internet-Drafts as reference 52 material or to cite them other than as "work in progress." 54 This Internet-Draft will expire on March 19, 2018. 56 Copyright Notice 58 Copyright (c) 2017 IETF Trust and the persons identified as the 59 document authors. All rights reserved. 61 This document is subject to BCP 78 and the IETF Trust's Legal 62 Provisions Relating to IETF Documents 63 (http://trustee.ietf.org/license-info) in effect on the date of 64 publication of this document. Please review these documents 65 carefully, as they describe your rights and restrictions with respect 66 to this document. Code Components extracted from this document must 67 include Simplified BSD License text as described in Section 4.e of 68 the Trust Legal Provisions and are provided without warranty as 69 described in the Simplified BSD License. 71 Table of Contents 73 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3 74 1.1. Version Control . . . . . . . . . . . . . . . . . . . . . 5 75 2. Overview . . . . . . . . . . . . . . . . . . . . . . . . . . 8 76 3. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 10 77 4. Background . . . . . . . . . . . . . . . . . . . . . . . . . 17 78 4.1. TCP properties . . . . . . . . . . . . . . . . . . . . . 18 79 4.2. Diagnostic Approach . . . . . . . . . . . . . . . . . . . 20 80 4.3. New requirements relative to RFC 2330 . . . . . . . . . . 21 81 5. Common Models and Parameters . . . . . . . . . . . . . . . . 22 82 5.1. Target End-to-end parameters . . . . . . . . . . . . . . 22 83 5.2. Common Model Calculations . . . . . . . . . . . . . . . . 23 84 5.3. Parameter Derating . . . . . . . . . . . . . . . . . . . 24 85 5.4. Test Preconditions . . . . . . . . . . . . . . . . . . . 24 86 6. Generating test streams . . . . . . . . . . . . . . . . . . . 25 87 6.1. Mimicking slowstart . . . . . . . . . . . . . . . . . . . 26 88 6.2. Constant window pseudo CBR . . . . . . . . . . . . . . . 27 89 6.3. Scanned window pseudo CBR . . . . . . . . . . . . . . . . 28 90 6.4. Concurrent or channelized testing . . . . . . . . . . . . 29 91 7. Interpreting the Results . . . . . . . . . . . . . . . . . . 30 92 7.1. Test outcomes . . . . . . . . . . . . . . . . . . . . . . 30 93 7.2. Statistical criteria for estimating run_length . . . . . 31 94 7.3. Reordering Tolerance . . . . . . . . . . . . . . . . . . 34 95 8. IP Diagnostic Tests . . . . . . . . . . . . . . . . . . . . . 34 96 8.1. Basic Data Rate and Packet Transfer Tests . . . . . . . . 35 97 8.1.1. Delivery Statistics at Paced Full Data Rate . . . . . 35 98 8.1.2. Delivery Statistics at Full Data Windowed Rate . . . 35 99 8.1.3. Background Packet Transfer Statistics Tests . . . . . 35 100 8.2. Standing Queue Tests . . . . . . . . . . . . . . . . . . 36 101 8.2.1. Congestion Avoidance . . . . . . . . . . . . . . . . 37 102 8.2.2. Bufferbloat . . . . . . . . . . . . . . . . . . . . . 37 103 8.2.3. Non excessive loss . . . . . . . . . . . . . . . . . 38 104 8.2.4. Duplex Self Interference . . . . . . . . . . . . . . 38 105 8.3. Slowstart tests . . . . . . . . . . . . . . . . . . . . . 39 106 8.3.1. Full Window slowstart test . . . . . . . . . . . . . 39 107 8.3.2. Slowstart AQM test . . . . . . . . . . . . . . . . . 39 108 8.4. Sender Rate Burst tests . . . . . . . . . . . . . . . . . 40 109 8.5. Combined and Implicit Tests . . . . . . . . . . . . . . . 41 110 8.5.1. Sustained Bursts Test . . . . . . . . . . . . . . . . 41 111 8.5.2. Passive Measurements . . . . . . . . . . . . . . . . 42 112 9. An Example . . . . . . . . . . . . . . . . . . . . . . . . . 43 113 9.1. Observations about applicability . . . . . . . . . . . . 44 114 10. Validation . . . . . . . . . . . . . . . . . . . . . . . . . 44 115 11. Security Considerations . . . . . . . . . . . . . . . . . . . 46 116 12. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . 46 117 13. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 47 118 14. Informative References . . . . . . . . . . . . . . . . . . . 47 119 Appendix A. Model Derivations . . . . . . . . . . . . . . . . . 51 120 A.1. Queueless Reno . . . . . . . . . . . . . . . . . . . . . 51 121 Appendix B. The effects of ACK scheduling . . . . . . . . . . . 52 122 Appendix C. Version Control . . . . . . . . . . . . . . . . . . 53 123 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 53 125 1. Introduction 127 Model Based Metrics (MBM) rely on peer-reviewed mathematical models 128 to specify a Targeted Suite of IP Diagnostic tests, designed to 129 assess whether common transport protocols can be expected to meet a 130 predetermined Target Transport Performance over an Internet path. 131 This note describes the modeling framework to derive the test 132 parameters for assessing an Internet path's ability to support a 133 predetermined Bulk Transport Capacity. 135 Each test in the Targeted IP Diagnostic Suite (TIDS) measures some 136 aspect of IP packet transfer needed to meet the Target Transport 137 Performance. For Bulk Transport Capacity the TIDS includes IP 138 diagnostic tests to verify that there is: sufficient IP capacity 139 (data rate); sufficient queue space at bottlenecks to absorb and 140 deliver typical transport bursts; and that the background packet loss 141 ratio is low enough not to interfere with congestion control; and 142 other properties described below. Unlike typical IPPM metrics which 143 yield measures of network properties, Model Based Metrics nominally 144 yield pass/fail evaluations of the ability of standard transport 145 protocols to meet the specific performance objective over some 146 network path. 148 In most cases, the IP diagnostic tests can be implemented by 149 combining existing IPPM metrics with additional controls for 150 generating test streams having a specified temporal structure (bursts 151 or standing queues caused by constant bit rate streams, etc.) and 152 statistical criteria for evaluating packet transfer. The temporal 153 structure of the test streams mimic transport protocol behavior over 154 the complete path; the statistical criteria models the transport 155 protocol's response to less than ideal IP packet transfer. In 156 control theory terms, the tests are "open loop". Note that running a 157 test requires the coordinated activity of sending and receiving 158 measurement points. 160 This note addresses Bulk Transport Capacity. It describes an 161 alternative to the approach presented in "A Framework for Defining 162 Empirical Bulk Transfer Capacity Metrics" [RFC3148]. Other Model 163 Based Metrics may cover other applications and transports, such as 164 VoIP over UDP and RTP, and new transport protocols. 166 This note assumes a traditional Reno TCP style self clocked, window 167 controlled transport protocol that uses packet loss and ECN CE marks 168 for congestion feedback. There are currently some experimental 169 protocols and congestion control algorithms that are rate based or 170 otherwise fall outside of these assumptions. In the future these new 171 protocols and algorithms may call for revised models. 173 The MBM approach, mapping Target Transport Performance to a Targeted 174 IP Diagnostic Suite (TIDS) of IP tests, solves some intrinsic 175 problems with using TCP or other throughput maximizing protocols for 176 measurement. In particular all throughput maximizing protocols (and 177 TCP congestion control in particular) cause some level of congestion 178 in order to detect when they have reached the available capacity 179 limitation of the network. This self inflicted congestion obscures 180 the network properties of interest and introduces non-linear dynamic 181 equilibrium behaviors that make any resulting measurements useless as 182 metrics because they have no predictive value for conditions or paths 183 different than that of the measurement itself. In order to prevent 184 these effects it is necessary to avoid the effects of TCP congestion 185 control in the measurement method. These issues are discussed at 186 length in Section 4. Readers whom are unfamiliar with basic 187 properties of TCP and TCP-like congestion control may find it easier 188 to start at Section 4 or Section 4.1. 190 A Targeted IP Diagnostic Suite does not have such difficulties. IP 191 diagnostics can be constructed such that they make strong statistical 192 statements about path properties that are independent of the 193 measurement details, such as vantage and choice of measurement 194 points. 196 1.1. Version Control 198 RFC Editor: Please remove this entire subsection prior to 199 publication. 201 REF Editor: The reference to draft-ietf-tcpm-rack is to attribute an 202 idea. This document should not block waiting for the completion of 203 that one. 205 Please send comments about this draft to ippm@ietf.org. See 206 http://goo.gl/02tkD for more information including: interim drafts, 207 an up to date todo list and information on contributing. 209 Formatted: Fri Sep 15 15:07:50 PDT 2017 211 Changes since -11 draft: 213 o (From IESG review comments.) 214 o Ben Campbell: Shorten the Abstract. 215 o Mirja Kuhlewind: Reduced redundancy. (See message) 216 o MK: Mention open loop in the introduction. 217 o MK: Spelled out ECN and reference RFC3168. 218 o MK: Added a paragraph to the introduction about assuming a 219 traditional self clocked, window controlled transport protocol. 220 o MK: Added language about initial window to the list at about 221 bursts at the end of section 4.1. 222 o MK: Network power is defined in the terminology section. 223 o MK: The introduction mention coordinated activity of both 224 endpoints. 225 o MK: The security section restates that some of the tests are not 226 intended for frequent monitoring tests as the high load can impact 227 other traffic negatively. 228 o MK: Restored "Informative References" section name. 229 o And a few minor nits. 231 Changes since -10 draft: 233 o A few more nits from various sources. 234 o (From IETF LC review comments.) 235 o David Mandelberg: design metrics to prevent DDOS. 236 o From Robert Sparks: 238 * Remove all legacy 2119 language. 239 * Fixed Xr notation inconsistency. 240 * Adjusted abstract: tests are only partially specified. 242 * Avoid rather than suppress the effects of congestion control 243 * Removed the unnecessary, excessively abstract and unclear 244 thought about IP vs TCP measurements. 245 * Changed "thwarted" to "not fulfilled". 246 * Qualified language about burst models. 247 * Replaced "infinitesimal" with other language. 248 * Added citations for the reordering strawman. 249 * Pointed out that pseudo CBR tests depend on self clock. 250 * Fixed some run on sentences. 251 o Update language to reflect RFC7567, AQM recommendations. 252 o Suggestion from Merry Mou (MIT) 254 Changes since -09 draft: 256 o Five last minute editing nits. 258 Changes since -08 draft: 260 o Language, spelling and usage nits. 261 o Expanded the abstract describe the models. 262 o Remove superfluous standards like language 263 o Remove superfluous "future technology" language. 264 o Interconnects -> network interconnections. 265 o Added more labels to Figure 1. 266 o Defined Bulk Transport. 267 o Clarified "implied bottleneck IP capacity" 268 o Clarified the history of the BTC metrics. 269 o Clarified stochastic vs non-stochastic test traffic generation. 270 o Reworked Fig 2 and 6.1 "Mimicking slowstart" 271 o Described the unsynchronized parallel stream failure case. 272 o Discussed how to measure devices that use virtual queues. 273 o Changed section 8.5.2 (Streaming Media) to be Passive 274 Measurements. 276 Changes since -07 draft: 278 o Sharpened the use of "statistical criteria" 279 o Sharpened the definition of test_window, and removed related 280 redundant text in several places 281 o Clarified "equilibrium" as "dynamic equilibrium, similar to 282 processes observed in chemistry" 283 o Properly explained "Heisenberg" as "observer effect" 284 o Added the observation from RFC 6576 that HW and SW congestion 285 control implementations do not generally give the same results. 286 o Noted that IP and application metrics differ as to how overhead is 287 handled. MBM is explicit about how it handles overhead. 288 o Clarified the language and added a new reference about the 289 problems caused by token bucket policers. 291 o Added an subsection in the example that comments on some of issues 292 that need to be mentioned in a future usage or applicability doc. 293 o Updated ippm-2680-bis to RFC7680 294 o Many terminology, punctuation and spelling nits. 296 Changes since -06 draft: 298 o More language nits: 300 * "Targeted IP Diagnostic Suite (TIDS)" replaces "Targeted 301 Diagnostic Suite (TDS)". 302 * "implied bottleneck IP capacity" replaces "implied bottleneck 303 IP rate". 304 * Updated to ECN CE Marks. 305 * Added "specified temporal structure" 306 * "test stream" replaces "test traffic" 307 * "packet transfer" replaces "packet delivery" 308 * Reworked discussion of slowstart, bursts and pacing. 309 * RFC 7567 replaces RFC 2309. 311 Changes since -05 draft: 313 o Wordsmithing on sections overhauled in -05 draft. 314 o Reorganized the document: 316 * Relocated subsection "Preconditions". 317 * Relocated subsection "New Requirements relative to RFC 2330". 318 o Addressed nits and not so nits by Ruediger Geib. (Thanks!) 319 o Substantially tightened the entire definitions section. 320 o Many terminology changes, to better conform to other docs : 322 * IP rate and IP capacity (following RFC 5136) replaces various 323 forms of link data rate. 324 * subpath replaces link. 325 * target_window_size replaces target_pipe_size. 326 * implied bottleneck IP rate replaces effective bottleneck link 327 rate. 328 * Packet delivery statistics replaces delivery statistics. 330 Changes since -04 draft: 332 o The introduction was heavily overhauled: split into a separate 333 introduction and overview. 334 o The new shorter introduction: 336 * Is a problem statement; 337 * This document provides a framework; 338 * That it replaces TCP measurement by IP tests; 339 * That the results are pass/fail. 340 o Added a diagram of the framework to the overview 341 o and introduces all of the elements of the framework. 342 o Renumbered sections, reducing the depth of some section numbers. 343 o Updated definitions to better agree with other documents: 345 * Reordered section 2 346 * Bulk [data] performance -> Bulk Transport Capacity, everywhere 347 including the title. 348 * loss rate and loss probability -> packet loss ratio 349 * end-to-end path -> complete path 350 * [end-to-end][target] performance -> Target Transport 351 Performance 352 * load test -> capacity test 354 2. Overview 356 This document describes a modeling framework for deriving a Targeted 357 IP Diagnostic Suite from a predetermined Target Transport 358 Performance. It is not a complete specification, and relies on other 359 standards documents to define important details such as packet Type-P 360 selection, sampling techniques, vantage selection, etc. We imagine 361 Fully Specified - Targeted IP Diagnostic Suites (FS-TIDS), that 362 define all of these details. We use Targeted IP Diagnostic Suite 363 (TIDS) to refer to the subset of such a specification that is in 364 scope for this document. This terminology is defined in Section 3. 366 Section 4 describes some key aspects of TCP behavior and what they 367 imply about the requirements for IP packet transfer. Most of the IP 368 diagnostic tests needed to confirm that the path meets these 369 properties can be built on existing IPPM metrics, with the addition 370 of statistical criteria for evaluating packet transfer and in a few 371 cases, new mechanisms to implement the required temporal structure. 372 (One group of tests, the standing queue tests described in 373 Section 8.2, don't correspond to existing IPPM metrics, but suitable 374 new IPPM metrics can be patterned after the existing definitions.) 376 Figure 1 shows the MBM modeling and measurement framework. The 377 Target Transport Performance, at the top of the figure, is determined 378 by the needs of the user or application, outside the scope of this 379 document. For Bulk Transport Capacity, the main performance 380 parameter of interest is the Target Data Rate. However, since TCP's 381 ability to compensate for less than ideal network conditions is 382 fundamentally affected by the Round Trip Time (RTT) and the Maximum 383 Transmission Unit (MTU) of the complete path, these parameters must 384 also be specified in advance based on knowledge about the intended 385 application setting. They may reflect a specific application over a 386 real path through the Internet or an idealized application and 387 hypothetical path representing a typical user community. Section 5 388 describes the common parameters and models derived from the Target 389 Transport Performance. 391 Target Transport Performance 392 (Target Data Rate, Target RTT and Target MTU) 393 | 394 ________V_________ 395 | mathematical | 396 | models | 397 | | 398 ------------------ 399 Traffic parameters | | Statistical criteria 400 | | 401 _______V____________V____Targeted_______ 402 | | * * * | Diagnostic Suite | 403 _____|_______V____________V________________ | 404 __|____________V____________V______________ | | 405 | IP diagnostic tests | | | 406 | | | | | | 407 | _____________V__ __V____________ | | | 408 | | traffic | | Delivery | | | | 409 | | pattern | | Evaluation | | | | 410 | | generation | | | | | | 411 | -------v-------- ------^-------- | | | 412 | | v test stream via ^ | | |-- 413 | | -->======================>-- | | | 414 | | subpath under test | |- 415 ----V----------------------------------V--- | 416 | | | | | | 417 V V V V V V 418 fail/inconclusive pass/fail/inconclusive 419 (traffic generation status) (test result) 421 Overall Modeling Framework 423 Figure 1 425 Mathematical TCP models are used to determine Traffic parameters and 426 subsequently to design traffic patterns that mimic TCP or other 427 transport protocol delivering bulk data and operating at the Target 428 Data Rate, MTU and RTT over a full range of conditions, including 429 flows that are bursty at multiple time scales. The traffic patterns 430 are generated based on the three Target parameters of complete path 431 and independent of the properties of individual subpaths using the 432 techniques described in Section 6. As much as possible the test 433 streams are generated deterministically (precomputed) to minimize the 434 extent to which test methodology, measurement points, measurement 435 vantage or path partitioning affect the details of the measurement 436 traffic. 438 Section 7 describes packet transfer statistics and methods to test 439 them against the statistical criteria provided by the mathematical 440 models. Since the statistical criteria typically apply to the 441 complete path (a composition of subpaths) [RFC6049], in situ testing 442 requires that the end-to-end statistical criteria be apportioned as 443 separate criteria for each subpath. Subpaths that are expected to be 444 bottlenecks would then be permitted to contribute a larger fraction 445 of the end-to-end packet loss budget. In compensation, subpaths that 446 are not expected to exhibit bottlenecks must be constrained to 447 contribute less packet loss. Thus the statistical criteria for each 448 subpath in each test of a TIDS is an apportioned share of the end-to- 449 end statistical criteria for the complete path which was determined 450 by the mathematical model. 452 Section 8 describes the suite of individual tests needed to verify 453 all of required IP delivery properties. A subpath passes if and only 454 if all of the individual IP diagnostic tests pass. Any subpath that 455 fails any test indicates that some users are likely to fail to attain 456 their Target Transport Performance under some conditions. In 457 addition to passing or failing, a test can be deemed to be 458 inconclusive for a number of reasons including: the precomputed 459 traffic pattern was not accurately generated; the measurement results 460 were not statistically significant; and others such as failing to 461 meet some required test preconditions. If all tests pass but some 462 are inconclusive, then the entire suite is deemed to be inconclusive. 464 In Section 9 we present an example TIDS that might be representative 465 of High Definition (HD) video, and illustrate how Model Based Metrics 466 can be used to address difficult measurement situations, such as 467 confirming that inter-carrier exchanges have sufficient performance 468 and capacity to deliver HD video between ISPs. 470 Since there is some uncertainty in the modeling process, Section 10 471 describes a validation procedure to diagnose and minimize false 472 positive and false negative results. 474 3. Terminology 476 Terms containing underscores (rather than spaces) appear in equations 477 and typically have algorithmic definitions. 479 General Terminology: 481 Target: A general term for any parameter specified by or derived 482 from the user's application or transport performance requirements. 484 Target Transport Performance: Application or transport performance 485 target values for the complete path. For Bulk Transport Capacity 486 defined in this note the Target Transport Performance includes the 487 Target Data Rate, Target RTT and Target MTU as described below. 488 Target Data Rate: The specified application data rate required for 489 an application's proper operation. Conventional Bulk Transport 490 Capacity (BTC) metrics are focused on the Target Data Rate, 491 however these metrics had little or no predictive value because 492 they do not consider the effects of the other two parameters of 493 the Target Transport Performance, the RTT and MTU of the complete 494 paths. 495 Target RTT (Round Trip Time): The specified baseline (minimum) RTT 496 of the longest complete path over which the user expects to be 497 able to meet the target performance. TCP and other transport 498 protocol's ability to compensate for path problems is generally 499 proportional to the number of round trips per second. The Target 500 RTT determines both key parameters of the traffic patterns (e.g. 501 burst sizes) and the thresholds on acceptable IP packet transfer 502 statistics. The Target RTT must be specified considering 503 appropriate packets sizes: MTU sized packets on the forward path, 504 ACK sized packets (typically header_overhead) on the return path. 505 Note that Target RTT is specified and not measured, MBM 506 measurements derived for a given target_RTT will be applicable to 507 any path with a smaller RTTs. 508 Target MTU (Maximum Transmission Unit): The specified maximum MTU 509 supported by the complete path the over which the application 510 expects to meet the target performance. In this document assume a 511 1500 Byte MTU unless otherwise specified. If some subpath has a 512 smaller MTU, then it becomes the Target MTU for the complete path, 513 and all model calculations and subpath tests must use the same 514 smaller MTU. 515 Targeted IP Diagnostic Suite (TIDS): A set of IP diagnostic tests 516 designed to determine if an otherwise ideal complete path 517 containing the subpath under test can sustain flows at a specific 518 target_data_rate using target_MTU sized packets when the RTT of 519 the complete path is target_RTT. 520 Fully Specified Targeted IP Diagnostic Suite (FS-TIDS): A TIDS 521 together with additional specification such as measurement packet 522 type ("type-p" [RFC2330]), etc. which are out of scope for this 523 document, but need to be drawn from other standards documents. 524 Bulk Transport Capacity: Bulk Transport Capacity Metrics evaluate an 525 Internet path's ability to carry bulk data, such as large files, 526 streaming (non-real time) video, and under some conditions, web 527 images and other content. Prior efforts to define BTC metrics 528 have been based on [RFC3148], which predates our understanding of 529 TCP and the requirements described in Section 4. In general "Bulk 530 Transport" indicates that performance is determined by the 531 interplay between the network, cross traffic and congestion 532 control in the transport protocol. It excludes situations where 533 performance is dominated by the RTT alone (e.g. transactions) or 534 bottlenecks elsewhere, such as in the application itself. 535 IP diagnostic tests: Measurements or diagnostics to determine if 536 packet transfer statistics meet some precomputed target. 537 traffic patterns: The temporal patterns or burstiness of traffic 538 generated by applications over transport protocols such as TCP. 539 There are several mechanisms that cause bursts at various time 540 scales as described in Section 4.1. Our goal here is to mimic the 541 range of common patterns (burst sizes and rates, etc), without 542 tying our applicability to specific applications, implementations 543 or technologies, which are sure to become stale. 544 Explicit Congestion Notification (ECN): See [RFC3168]. 545 packet transfer statistics: Raw, detailed or summary statistics 546 about packet transfer properties of the IP layer including packet 547 losses, ECN Congestion Experienced (CE) marks, reordering, or any 548 other properties that may be germane to transport performance. 549 packet loss ratio: As defined in [RFC7680]. 550 apportioned: To divide and allocate, for example budgeting packet 551 loss across multiple subpaths such that the losses will accumulate 552 to less than a specified end-to-end loss ratio. Apportioning 553 metrics is essentially the inverse of the process described in 554 [RFC5835]. 555 open loop: A control theory term used to describe a class of 556 techniques where systems that naturally exhibit circular 557 dependencies can be analyzed by suppressing some of the 558 dependencies, such that the resulting dependency graph is acyclic. 560 Terminology about paths, etc. See [RFC2330] and [RFC7398] for 561 existing terms and definitions. 563 data sender: Host sending data and receiving ACKs. 564 data receiver: Host receiving data and sending ACKs. 565 complete path: The end-to-end path from the data sender to the data 566 receiver. 567 subpath: A portion of the complete path. Note that there is no 568 requirement that subpaths be non-overlapping. A subpath can be a 569 small as a single device, link or interface. 570 measurement point: Measurement points as described in [RFC7398]. 571 test path: A path between two measurement points that includes a 572 subpath of the complete path under test. If the measurement 573 points are off path, the test path may include "test leads" 574 between the measurement points and the subpath. 575 dominant bottleneck: The bottleneck that generally determines most 576 of packet transfer statistics for the entire path. It typically 577 determines a flow's self clock timing, packet loss and ECN 578 Congestion Experienced (CE) marking rate, with other potential 579 bottlenecks having less effect on the packet transfer statistics. 580 See Section 4.1 on TCP properties. 581 front path: The subpath from the data sender to the dominant 582 bottleneck. 583 back path: The subpath from the dominant bottleneck to the receiver. 584 return path: The path taken by the ACKs from the data receiver to 585 the data sender. 586 cross traffic: Other, potentially interfering, traffic competing for 587 network resources (bandwidth and/or queue capacity). 589 Properties determined by the complete path and application. These 590 are described in more detail in Section 5.1. 592 Application Data Rate: General term for the data rate as seen by the 593 application above the transport layer in bytes per second. This 594 is the payload data rate, and explicitly excludes transport and 595 lower level headers (TCP/IP or other protocols), retransmissions 596 and other overhead that is not part to the total quantity of data 597 delivered to the application. 598 IP rate: The actual number of IP-layer bytes delivered through a 599 subpath, per unit time, including TCP and IP headers, retransmits 600 and other TCP/IP overhead. Follows from IP-type-P Link Usage 601 [RFC5136]. 602 IP capacity: The maximum number of IP-layer bytes that can be 603 transmitted through a subpath, per unit time, including TCP and IP 604 headers, retransmits and other TCP/IP overhead. Follows from IP- 605 type-P Link Capacity [RFC5136]. 606 bottleneck IP capacity: The IP capacity of the dominant bottleneck 607 in the forward path. All throughput maximizing protocols estimate 608 this capacity by observing the IP rate delivered through the 609 bottleneck. Most protocols derive their self clocks from the 610 timing of this data. See Section 4.1 and Appendix B for more 611 details. 612 implied bottleneck IP capacity: This is the bottleneck IP capacity 613 implied by the ACKs returning from the receiver. It is determined 614 by looking at how much application data the ACK stream at the 615 sender reports delivered to the data receiver per unit time at 616 various time scales. If the return path is thinning, batching or 617 otherwise altering the ACK timing the implied bottleneck IP 618 capacity over short time scales might be substantially larger than 619 the bottleneck IP capacity averaged over a full RTT. Since TCP 620 derives its clock from the data delivered through the bottleneck, 621 the front path must have sufficient buffering to absorb any data 622 bursts at the dimensions (size and IP rate) implied by the ACK 623 stream, which are potentially doubled during slowstart. If the 624 return path is not altering the ACK stream, then the implied 625 bottleneck IP capacity will be the same as the bottleneck IP 626 capacity. See Section 4.1 and Appendix B for more details. 628 sender interface rate: The IP rate which corresponds to the IP 629 capacity of the data sender's interface. Due to sender efficiency 630 algorithms including technologies such as TCP segmentation offload 631 (TSO), nearly all modern servers deliver data in bursts at full 632 interface link rate. Today 1 or 10 Gb/s are typical. 633 Header_overhead: The IP and TCP header sizes, which are the portion 634 of each MTU not available for carrying application payload. 635 Without loss of generality this is assumed to be the size for 636 returning acknowledgments (ACKs). For TCP, the Maximum Segment 637 Size (MSS) is the Target MTU minus the header_overhead. 639 Basic parameters common to models and subpath tests are defined here 640 are described in more detail in Section 5.2. Note that these are 641 mixed between application transport performance (excludes headers) 642 and IP performance (which include TCP headers and retransmissions as 643 part of the IP payload). 645 Network power: The observed data rate divided by the observed RTT. 646 Network power indicates how effectively a transport protocol is 647 filling a network. 648 Window [size]: The total quantity of data carried by packets in- 649 flight plus the data represented by ACKs circulating in the 650 network is referred to as the window. See Section 4.1. Sometimes 651 used with other qualifiers (congestion window, cwnd or receiver 652 window) to indicate which mechanism is controlling the window. 653 pipe size: A general term for number of packets needed in flight 654 (the window size) to exactly fill some network path or subpath. 655 It corresponds to the window size which maximizes network power. 656 Often used with additional qualifiers to specify which path, or 657 under what conditions, etc. 658 target_window_size: The average number of packets in flight (the 659 window size) needed to meet the Target Data Rate, for the 660 specified Target RTT, and MTU. It implies the scale of the bursts 661 that the network might experience. 662 run length: A general term for the observed, measured, or specified 663 number of packets that are (expected to be) delivered between 664 losses or ECN Congestion Experienced (CE) marks. Nominally one 665 over the sum of the loss and ECN CE marking probabilities, if 666 there are independently and identically distributed. 667 target_run_length: The target_run_length is an estimate of the 668 minimum number of non-congestion marked packets needed between 669 losses or ECN Congestion Experienced (CE) marks necessary to 670 attain the target_data_rate over a path with the specified 671 target_RTT and target_MTU, as computed by a mathematical model of 672 TCP congestion control. A reference calculation is shown in 673 Section 5.2 and alternatives in Appendix A 675 reference target_run_length: target_run_length computed precisely by 676 the method in Section 5.2. This is likely to be slightly more 677 conservative than required by modern TCP implementations. 679 Ancillary parameters used for some tests: 681 derating: Under some conditions the standard models are too 682 conservative. The modeling framework permits some latitude in 683 relaxing or "derating" some test parameters as described in 684 Section 5.3 in exchange for a more stringent TIDS validation 685 procedures, described in Section 10. Models can be derated by 686 including a multiplicative derating factor to make tests less 687 stringent. 688 subpath_IP_capacity: The IP capacity of a specific subpath. 689 test path: A subpath of a complete path under test. 690 test_path_RTT: The RTT observed between two measurement points using 691 packet sizes that are consistent with the transport protocol. 692 This is generally MTU sized packets of the forward path, 693 header_overhead sized packets on the return path. 694 test_path_pipe: The pipe size of a test path. Nominally the 695 test_path_RTT times the test path IP_capacity. 696 test_window: The smallest window sufficient to meet or exceed the 697 target_rate when operating with a pure self clock over a test 698 path. The test_window is typically given by 699 ceiling(target_data_rate*test_path_RTT/(target_MTU- 700 header_overhead)) but see the discussion in Appendix B about the 701 effects of channel scheduling on RTT. On some test paths the 702 test_window may need to be adjusted slightly to compensate for the 703 RTT being inflated by the devices that schedule packets. 705 The terminology below is used to define temporal patterns for test 706 stream. These patterns are designed to mimic TCP behavior, as 707 described in Section 4.1. 709 packet headway: Time interval between packets, specified from the 710 start of one to the start of the next. e.g. If packets are sent 711 with a 1 mS headway, there will be exactly 1000 packets per 712 second. 713 burst headway: Time interval between bursts, specified from the 714 start of the first packet one burst to the start of the first 715 packet of the next burst. e.g. If 4 packet bursts are sent with a 716 1 mS burst headway, there will be exactly 4000 packets per second. 717 paced single packets: Send individual packets at the specified rate 718 or packet headway. 719 paced bursts: Send bursts on a timer. Specify any 3 of: average 720 data rate, packet size, burst size (number of packets) and burst 721 headway (burst start to start). By default the bursts are assumed 722 to occur at full sender interface rate, such that the packet 723 headway within each burst is the minimum supported by the sender's 724 interface. Under some conditions it is useful to explicitly 725 specify the packet headway within each burst. 726 slowstart rate: Mimic TCP slowstart by sending 4 packet paced bursts 727 at an average data rate equal to twice the implied bottleneck IP 728 capacity (but not more than the sender interface rate). This is a 729 two level burst pattern described in more detail in Section 6.1. 730 If the implied bottleneck IP capacity is more than half of the 731 sender interface rate, slowstart rate becomes sender interface 732 rate. 733 slowstart burst: Mimic one round of TCP slowstart by sending a 734 specified number of packets packets in a two level burst pattern 735 that resembles slowstart. 736 repeated slowstart bursts: Repeat Slowstart bursts once per 737 target_RTT. For TCP each burst would be twice as large as the 738 prior burst, and the sequence would end at the first ECN CE mark 739 or lost packet. For measurement, all slowstart bursts would be 740 the same size (nominally target_window_size but other sizes might 741 be specified), and the ECN CE marks and lost packets are counted. 743 The tests described in this note can be grouped according to their 744 applicability. 746 Capacity tests: Capacity tests determine if a network subpath has 747 sufficient capacity to deliver the Target Transport Performance. 748 As long as the test stream is within the proper envelope for the 749 Target Transport Performance, the average packet losses or ECN 750 Congestion Experienced (CE) marks must be below the statistical 751 criteria computed by the model. As such, capacity tests reflect 752 parameters that can transition from passing to failing as a 753 consequence of cross traffic, additional presented load or the 754 actions of other network users. By definition, capacity tests 755 also consume significant network resources (data capacity and/or 756 queue buffer space), and the test schedules must be balanced by 757 their cost. 758 Monitoring tests: Monitoring tests are designed to capture the most 759 important aspects of a capacity test, but without presenting 760 excessive ongoing load themselves. As such they may miss some 761 details of the network's performance, but can serve as a useful 762 reduced-cost proxy for a capacity test, for example to support 763 continuous production network monitoring. 764 Engineering tests: Engineering tests evaluate how network algorithms 765 (such as AQM and channel allocation) interact with TCP-style self 766 clocked protocols and adaptive congestion control based on packet 767 loss and ECN Congestion Experienced (CE) marks. These tests are 768 likely to have complicated interactions with cross traffic and 769 under some conditions can be inversely sensitive to load. For 770 example a test to verify that an AQM algorithm causes ECN CE marks 771 or packet drops early enough to limit queue occupancy may 772 experience a false pass result in the presence of cross traffic. 773 It is important that engineering tests be performed under a wide 774 range of conditions, including both in situ and bench testing, and 775 over a wide variety of load conditions. Ongoing monitoring is 776 less likely to be useful for engineering tests, although sparse in 777 situ testing might be appropriate. 779 4. Background 781 At the time the "Framework for IP Performance Metrics" [RFC2330] was 782 published (1998), sound Bulk Transport Capacity (BTC) measurement was 783 known to be well beyond our capabilities. Even when Framework for 784 Empirical BTC Metrics [RFC3148] was published, we knew that we didn't 785 really understand the problem. Now, by hindsight we understand why 786 assessing BTC is such a hard problem: 788 o TCP is a control system with circular dependencies - everything 789 affects performance, including components that are explicitly not 790 part of the test (for example, the host processing power is not 791 in-scope of path performance tests). 792 o Congestion control is a dynamic equilibrium process, similar to 793 processes observed in chemistry and other fields. The network and 794 transport protocols find an operating point which balances between 795 opposing forces: the transport protocol pushing harder (raising 796 the data rate and/or window) while the network pushes back 797 (raising packet loss ratio, RTT and/or ECN CE marks). By design 798 TCP congestion control keeps raising the data rate until the 799 network gives some indication that its capacity has been exceeded 800 by dropping packets or adding ECN CE marks. If a TCP sender 801 accurately fills a path to its IP capacity, (e.g. the bottleneck 802 is 100% utilized), then packet losses and ECN CE marks are mostly 803 determined by the TCP sender and how aggressively it seeks 804 additional capacity, and not the network itself, since the network 805 must send exactly the signals that TCP needs to set its rate. 806 o TCP's ability to compensate for network impairments (such as loss, 807 delay and delay variation, outside of those caused by TCP itself) 808 is directly proportional to the number of send-ACK round trip 809 exchanges per second (i.e. inversely proportional to the RTT). As 810 a consequence an impaired subpath may pass a short RTT local test 811 even though it fails when the subpath is extended by an 812 effectively perfect network to some larger RTT. 813 o TCP has an extreme form of the Observer Effect (colloquially know 814 as the Heisenberg effect). Measurement and cross traffic interact 815 in unknown and ill defined ways. The situation is actually worse 816 than the traditional physics problem where you can at least 817 estimate bounds on the relative momentum of the measurement and 818 measured particles. For network measurement you can not in 819 general determine even the order of magnitude of the effect. It 820 is possible to construct measurement scenarios where the 821 measurement traffic starves real user traffic, yielding an overly 822 inflated measurement. The inverse is also possible: the user 823 traffic can fill the network, such that the measurement traffic 824 detects only minimal available capacity. You can not in general 825 determine which scenario might be in effect, so you can not gauge 826 the relative magnitude of the uncertainty introduced by 827 interactions with other network traffic. 828 o As a consequence of the properties listed above it is difficult, 829 if not impossible, for two independent implementations (HW or SW) 830 of TCP congestion control to produce equivalent performance 831 results [RFC6576] under the same network conditions, 833 These properties are a consequence of the dynamic equilibrium 834 behavior intrinsic to how all throughput maximizing protocols 835 interact with the Internet. These protocols rely on control systems 836 based on estimated network metrics to regulate the quantity of data 837 to send into the network. The packet sending characteristics in turn 838 alter the network properties estimated by the control system metrics, 839 such that there are circular dependencies between every transmission 840 characteristic and every estimated metric. Since some of these 841 dependencies are nonlinear, the entire system is nonlinear, and any 842 change anywhere causes a difficult to predict response in network 843 metrics. As a consequence Bulk Transport Capacity metrics have not 844 fulfilled the analytic framework envisioned in [RFC2330] 846 Model Based Metrics overcome these problems by making the measurement 847 system open loop: the packet transfer statistics (akin to the network 848 estimators) do not affect the traffic or traffic patterns (bursts), 849 which are computed on the basis of the Target Transport Performance. 850 A path or subpath meeting the Target Transfer Performance 851 requirements would exhibit packet transfer statistics and estimated 852 metrics that would not cause the control system to slow the traffic 853 below the Target Data Rate. 855 4.1. TCP properties 857 TCP and other self clocked protocols (e.g. SCTP) carry the vast 858 majority of all Internet data. Their dominant bulk data transport 859 behavior is to have an approximately fixed quantity of data and 860 acknowledgments (ACKs) circulating in the network. The data receiver 861 reports arriving data by returning ACKs to the data sender, the data 862 sender typically responds by sending approximately the same quantity 863 of data back into the network. The total quantity of data plus the 864 data represented by ACKs circulating in the network is referred to as 865 the window. The mandatory congestion control algorithms 866 incrementally adjust the window by sending slightly more or less data 867 in response to each ACK. The fundamentally important property of 868 this system is that it is self clocked: The data transmissions are a 869 reflection of the ACKs that were delivered by the network, the ACKs 870 are a reflection of the data arriving from the network. 872 A number of protocol features cause bursts of data, even in idealized 873 networks that can be modeled as simple queuing systems. 875 During slowstart the IP rate is doubled on each RTT by sending twice 876 as much data as was delivered to the receiver during the prior RTT. 877 Each returning ACK causes the sender to transmit twice the data the 878 ACK reported arriving at the receiver. For slowstart to be able to 879 fill the pipe, the network must be able to tolerate slowstart bursts 880 up to the full pipe size inflated by the anticipated window reduction 881 on the first loss or ECN CE mark. For example, with classic Reno 882 congestion control, an optimal slowstart has to end with a burst that 883 is twice the bottleneck rate for one RTT in duration. This burst 884 causes a queue which is equal to the pipe size (i.e. the window is 885 twice the pipe size) so when the window is halved in response to the 886 first packet loss, the new window will be the pipe size. 888 Note that if the bottleneck IP rate is less that half of the capacity 889 of the front path (which is almost always the case), the slowstart 890 bursts will not by themselves cause significant queues anywhere else 891 along the front path; they primarily exercise the queue at the 892 dominant bottleneck. 894 Several common efficiency algorithms also cause bursts. The self 895 clock is typically applied to groups of packets: the receiver's 896 delayed ACK algorithm generally sends only one ACK per two data 897 segments. Furthermore the modern senders use TCP segmentation 898 offload (TSO) to reduce CPU overhead. The sender's software stack 899 builds super sized TCP segments that the TSO hardware splits into MTU 900 sized segments on the wire. The net effect of TSO, delayed ACK and 901 other efficiency algorithms is to send bursts of segments at full 902 sender interface rate. 904 Note that these efficiency algorithms are almost always in effect, 905 including during slowstart, such that slowstart typically has a two 906 level burst structure. Section 6.1 describes slowstart in more 907 detail. 909 Additional sources of bursts include TCP's initial window [RFC6928], 910 application pauses, channel allocation mechanisms and network devices 911 that schedule ACKs. Appendix B describes these last two items. If 912 the application pauses (stops reading or writing data) for some 913 fraction of an RTT, many TCP implementations catch up to their 914 earlier window size by sending a burst of data at the full sender 915 interface rate. To fill a network with a realistic application, the 916 network has to be able to tolerate sender interface rate bursts large 917 enough to restore the prior window following application pauses. 919 Although the sender interface rate bursts are typically smaller than 920 the last burst of a slowstart, they are at a higher IP rate so they 921 potentially exercise queues at arbitrary points along the front path 922 from the data sender up to and including the queue at the dominant 923 bottleneck. It is known that these bursts can hurt network 924 performance, especially in conjunction with other queue pressure, 925 however we are not aware of any models for how frequent sender rate 926 bursts the network should be able to tolerate at various burst sizes. 928 In conclusion, to verify that a path can meet a Target Transport 929 Performance, it is necessary to independently confirm that the path 930 can tolerate bursts at the scales that can be caused by the above 931 mechanisms. Three cases are believed to be sufficient: 933 o Two level slowstart bursts sufficient to get connections started 934 properly. 935 o Ubiquitous sender interface rate bursts caused by efficiency 936 algorithms. We assume 4 packet bursts to be the most common case, 937 since it matches the effects of delayed ACK during slowstart. 938 These bursts should be assumed not to significantly affect packet 939 transfer statistics. 940 o Infrequent sender interface rate bursts that are the maximum of 941 the full target_window_size and the initial window size (10 942 segments in [RFC6928]). The Target_run_length may be derated for 943 these large fast bursts. 945 If a subpath can meet the required packet loss ratio for bursts at 946 all of these scales then it has sufficient buffering at all potential 947 bottlenecks to tolerate any of the bursts that are likely introduced 948 by TCP or other transport protocols. 950 4.2. Diagnostic Approach 952 A complete path of a given RTT and MTU, which are equal to or smaller 953 than the Target RTT and equal to or larger than the Target MTU 954 respectively, is expected to be able to attain a specified Bulk 955 Transport Capacity when all of the following conditions are met: 957 1. The IP capacity is above the Target Data Rate by sufficient 958 margin to cover all TCP/IP overheads. This can be confirmed by 959 the tests described in Section 8.1 or any number of IP capacity 960 tests adapted to implement MBM. 961 2. The observed packet transfer statistics are better than required 962 by a suitable TCP performance model (e.g. fewer packet losses or 963 ECN CE marks). See Section 8.1 or any number of low or fixed 964 rate packet loss tests outside of MBM. 965 3. There is sufficient buffering at the dominant bottleneck to 966 absorb a slowstart bursts large enough to get the flow out of 967 slowstart at a suitable window size. See Section 8.3. 968 4. There is sufficient buffering in the front path to absorb and 969 smooth sender interface rate bursts at all scales that are likely 970 to be generated by the application, any channel arbitration in 971 the ACK path or any other mechanisms. See Section 8.4. 972 5. When there is a slowly rising standing queue at the bottleneck 973 the onset of packet loss has to be at an appropriate point (time 974 or queue depth) and progressive [RFC7567]. See Section 8.2. 975 6. When there is a standing queue at a bottleneck for a shared media 976 subpath (e.g. half duplex), there must be a suitable bounds on 977 the interaction between ACKs and data, for example due to the 978 channel arbitration mechanism. See Section 8.2.4. 980 Note that conditions 1 through 4 require capacity tests for 981 validation, and thus may need to be monitored on an ongoing basis. 982 Conditions 5 and 6 require engineering tests, which are best 983 performed in controlled environments such as a bench test. They 984 won't generally fail due to load, but may fail in the field due to 985 configuration errors, etc. and should be spot checked. 987 A tool that can perform many of the tests is available from 988 [MBMSource]. 990 4.3. New requirements relative to RFC 2330 992 Model Based Metrics are designed to fulfill some additional 993 requirements that were not recognized at the time RFC 2330 was 994 written [RFC2330]. These missing requirements may have significantly 995 contributed to policy difficulties in the IP measurement space. Some 996 additional requirements are: 998 o IP metrics must be actionable by the ISP - they have to be 999 interpreted in terms of behaviors or properties at the IP or lower 1000 layers, that an ISP can test, repair and verify. 1001 o Metrics should be spatially composable, such that measures of 1002 concatenated paths should be predictable from subpaths. 1003 o Metrics must be vantage point invariant over a significant range 1004 of measurement point choices, including off path measurement 1005 points. The only requirements on MP selection should be that the 1006 RTT between the MPs is below some reasonable bound, and that the 1007 effects of the "test leads" connecting MPs to the subpath under 1008 test can be can be calibrated out of the measurements. The latter 1009 might be be accomplished if the test leads are effectively ideal 1010 or their properties can be deducted from the measurements between 1011 the MPs. While many of tests require that the test leads have at 1012 least as much IP capacity as the subpath under test, some do not, 1013 for example Background Packet Transfer Tests described in 1014 Section 8.1.3. 1015 o Metric measurements should be repeatable by multiple parties with 1016 no specialized access to MPs or diagnostic infrastructure. It 1017 should be possible for different parties to make the same 1018 measurement and observe the same results. In particular it is 1019 specifically important that both a consumer (or their delegate) 1020 and ISP be able to perform the same measurement and get the same 1021 result. Note that vantage independence is key to meeting this 1022 requirement. 1024 5. Common Models and Parameters 1026 5.1. Target End-to-end parameters 1028 The target end-to-end parameters are the Target Data Rate, Target RTT 1029 and Target MTU as defined in Section 3. These parameters are 1030 determined by the needs of the application or the ultimate end user 1031 and the complete Internet path over which the application is expected 1032 to operate. The target parameters are in units that make sense to 1033 upper layers: payload bytes delivered to the application, above TCP. 1034 They exclude overheads associated with TCP and IP headers, 1035 retransmits and other protocols (e.g. DNS). Note that IP-based 1036 network services include TCP headers and retransmissions as part of 1037 delivered payload, and this difference is recognized in calculations 1038 below (header_overhead). 1040 Other end-to-end parameters defined in Section 3 include the 1041 effective bottleneck data rate, the sender interface data rate and 1042 the TCP and IP header sizes. 1044 The target_data_rate must be smaller than all subpath IP capacities 1045 by enough headroom to carry the transport protocol overhead, 1046 explicitly including retransmissions and an allowance for 1047 fluctuations in TCP's actual data rate. Specifying a 1048 target_data_rate with insufficient headroom is likely to result in 1049 brittle measurements having little predictive value. 1051 Note that the target parameters can be specified for a hypothetical 1052 path, for example to construct TIDS designed for bench testing in the 1053 absence of a real application; or for a live in situ test of 1054 production infrastructure. 1056 The number of concurrent connections is explicitly not a parameter to 1057 this model. If a subpath requires multiple connections in order to 1058 meet the specified performance, that must be stated explicitly and 1059 the procedure described in Section 6.4 applies. 1061 5.2. Common Model Calculations 1063 The Target Transport Performance is used to derive the 1064 target_window_size and the reference target_run_length. 1066 The target_window_size, is the average window size in packets needed 1067 to meet the target_rate, for the specified target_RTT and target_MTU. 1068 It is given by: 1070 target_window_size = ceiling( target_rate * target_RTT / ( target_MTU 1071 - header_overhead ) ) 1073 Target_run_length is an estimate of the minimum required number of 1074 unmarked packets that must be delivered between losses or ECN 1075 Congestion Experienced (CE) marks, as computed by a mathematical 1076 model of TCP congestion control. The derivation here follows 1077 [MSMO97], and by design is quite conservative. 1079 Reference target_run_length is derived as follows: assume the 1080 subpath_IP_capacity is infinitesimally larger than the 1081 target_data_rate plus the required header_overhead. Then 1082 target_window_size also predicts the onset of queuing. A larger 1083 window will cause a standing queue at the bottleneck. 1085 Assume the transport protocol is using standard Reno style Additive 1086 Increase, Multiplicative Decrease (AIMD) congestion control [RFC5681] 1087 (but not Appropriate Byte Counting [RFC3465]) and the receiver is 1088 using standard delayed ACKs. Reno increases the window by one packet 1089 every pipe_size worth of ACKs. With delayed ACKs this takes 2 Round 1090 Trip Times per increase. To exactly fill the pipe, the spacing of 1091 losses must be no closer than when the peak of the AIMD sawtooth 1092 reached exactly twice the target_window_size. Otherwise, the 1093 multiplicative window reduction triggered by the loss would cause the 1094 network to be under-filled. Following [MSMO97] the number of packets 1095 between losses must be the area under the AIMD sawtooth. They must 1096 be no more frequent than every 1 in 1097 ((3/2)*target_window_size)*(2*target_window_size) packets, which 1098 simplifies to: 1100 target_run_length = 3*(target_window_size^2) 1102 Note that this calculation is very conservative and is based on a 1103 number of assumptions that may not apply. Appendix A discusses these 1104 assumptions and provides some alternative models. If a different 1105 model is used, a FS-TIDS must document the actual method for 1106 computing target_run_length and ratio between alternate 1107 target_run_length and the reference target_run_length calculated 1108 above, along with a discussion of the rationale for the underlying 1109 assumptions. 1111 These two parameters, target_window_size and target_run_length, 1112 directly imply most of the individual parameters for the tests in 1113 Section 8. 1115 5.3. Parameter Derating 1117 Since some aspects of the models are very conservative, the MBM 1118 framework permits some latitude in derating test parameters. Rather 1119 than trying to formalize more complicated models we permit some test 1120 parameters to be relaxed as long as they meet some additional 1121 procedural constraints: 1123 o The FS-TIDS must document and justify the actual method used to 1124 compute the derated metric parameters. 1125 o The validation procedures described in Section 10 must be used to 1126 demonstrate the feasibility of meeting the Target Transport 1127 Performance with infrastructure that just barely passes the 1128 derated tests. 1129 o The validation process for a FS-TIDS itself must be documented is 1130 such a way that other researchers can duplicate the validation 1131 experiments. 1133 Except as noted, all tests below assume no derating. Tests where 1134 there is not currently a well established model for the required 1135 parameters explicitly include derating as a way to indicate 1136 flexibility in the parameters. 1138 5.4. Test Preconditions 1140 Many tests have preconditions which are required to assure their 1141 validity. Examples include: the presence or non-presence of cross 1142 traffic on specific subpaths; negotiating ECN; and appropriate 1143 preamble packet stream to testing to put reactive network elements 1144 into the proper states [RFC7312]. If preconditions are not properly 1145 satisfied for some reason, the tests should be considered to be 1146 inconclusive. In general it is useful to preserve diagnostic 1147 information as to why the preconditions were not met, and any test 1148 data that was collected even if it is not useful for the intended 1149 test. Such diagnostic information and partial test data may be 1150 useful for improving the test or test procedures themselves. 1152 It is important to preserve the record that a test was scheduled, 1153 because otherwise precondition enforcement mechanisms can introduce 1154 sampling bias. For example, canceling tests due to cross traffic on 1155 subscriber access links might introduce sampling bias in tests of the 1156 rest of the network by reducing the number of tests during peak 1157 network load. 1159 Test preconditions and failure actions must be specified in a FS- 1160 TIDS. 1162 6. Generating test streams 1164 Many important properties of Model Based Metrics, such as vantage 1165 independence, are a consequence of using test streams that have 1166 temporal structures that mimic TCP or other transport protocols 1167 running over a complete path. As described in Section 4.1, self 1168 clocked protocols naturally have burst structures related to the RTT 1169 and pipe size of the complete path. These bursts naturally get 1170 larger (contain more packets) as either the Target RTT or Target Data 1171 Rate get larger, or the Target MTU gets smaller. An implication of 1172 these relationships is that test streams generated by running self 1173 clocked protocols over short subpaths may not adequately exercise the 1174 queuing at any bottleneck to determine if the subpath can support the 1175 full Target Transport Performance over the complete path. 1177 Failing to authentically mimic TCP's temporal structure is part of 1178 the reason why simple performance tools such as iPerf, netperf, nc, 1179 etc have the reputation of yielding false pass results over short 1180 test paths, even when some subpath has a flaw. 1182 The definitions in Section 3 are sufficient for most test streams. 1183 We describe the slowstart and standing queue test streams in more 1184 detail. 1186 In conventional measurement practice stochastic processes are used to 1187 eliminate many unintended correlations and sample biases. However 1188 MBM tests are designed to explicitly mimic temporal correlations 1189 caused by network or protocol elements themselves. Some portions of 1190 these systems, such as traffic arrival (test scheduling) are 1191 naturally stochastic. Other behaviors, such as back-to-back packet 1192 transmissions, are dominated by implementation specific deterministic 1193 effects. Although these behaviors always contain non-deterministic 1194 elements and might be modeled stochastically, these details typically 1195 do not contribute significantly to the overall system behavior. 1196 Furthermore, it is known that real protocols are subject to failures 1197 caused by network property estimators suffering from bias due to 1198 correlation in their own traffic. For example TCP's RTT estimator 1199 used to determine the Retransmit Time Out (RTO), can be fooled by 1200 periodic cross traffic or start-stop applications. For these reasons 1201 many details of the test streams are specified deterministically. 1203 It may prove useful to introduce fine grained noise sources into the 1204 models used for generating test streams in an update of Model Based 1205 Metrics, but the complexity is not warranted at the time this 1206 document was written. 1208 6.1. Mimicking slowstart 1210 TCP slowstart has a two level burst structure as shown in Figure 2. 1211 The fine time structure is caused by efficiency algorithms that 1212 deliberately batch work (CPU, channel allocation, etc) to better 1213 amortize certain network and host overheads. ACKs passing through 1214 the return path typically cause the sender to transmit small bursts 1215 of data at full sender interface rate. For example TCP Segmentation 1216 Offload (TSO) and Delayed Acknowledgment both contribute to this 1217 effect. During slowstart these bursts are at the same headway as the 1218 returning ACKs, but are typically twice as large (e.g. having twice 1219 as much data) as the ACK reported was delivered to the receiver. Due 1220 to variations in delayed ACK and algorithms such as Appropriate Byte 1221 Counting [RFC3465], different pairs of senders and receivers produce 1222 slightly different burst patterns. Without loss of generality, we 1223 assume each ACK causes 4 packet sender interface rate bursts at an 1224 average headway equal to the ACK headway, and corresponding to 1225 sending at an average rate equal to twice the effective bottleneck IP 1226 rate. Each slowstart burst consists of a series of 4 packet sender 1227 interface rate bursts such that the total number of packets is the 1228 current window size (as of the last packet in the burst). 1230 The coarse time structure is due to each RTT being a reflection of 1231 the prior RTT. For real transport protocols, each slowstart burst is 1232 twice as large (twice the window) as the previous burst but is spread 1233 out in time by the network bottleneck, such that each successive RTT 1234 exhibits the same effective bottleneck IP rate. The slowstart phase 1235 ends on the first lost packet or ECN mark, which is intended to 1236 happen after successive slowstart bursts merge in time: the next 1237 burst starts before the bottleneck queue is fully drained and the 1238 prior burst is complete. 1240 For diagnostic tests described below we preserve the fine time 1241 structure but manipulate the coarse structure of the slowstart bursts 1242 (burst size and headway) to measure the ability of the dominant 1243 bottleneck to absorb and smooth slowstart bursts. 1245 Note that a stream of repeated slowstart bursts has three different 1246 average rates, depending on the averaging time interval. At the 1247 finest time scale (a few packet times at the sender interface) the 1248 peak of the average IP rate is the same as the sender interface rate; 1249 at a medium timescale (a few ACK times at the dominant bottleneck) 1250 the peak of the average IP rate is twice the implied bottleneck IP 1251 capacity; and at time scales longer than the target_RTT and when the 1252 burst size is equal to the target_window_size, the average rate is 1253 equal to the target_data_rate. This pattern corresponds to repeating 1254 the last RTT of TCP slowstart when delayed ACK and sender side byte 1255 counting are present but without the limits specified in Appropriate 1256 Byte Counting [RFC3465]. 1258 time ==> ( - equals one packet) 1260 Fine time structure of the packet stream: 1262 ---- ---- ---- ---- ---- 1264 |<>| sender interface rate bursts (typically 3 or 4 packets) 1265 |<===>| burst headway (from the ACK headway) 1267 \____repeating sender______/ 1268 rate bursts 1270 Coarse (RTT level) time structure of the packet stream: 1272 ---- ---- ---- ---- ---- ---- ---- ... 1274 |<========================>| slowstart burst size (from the window) 1275 |<==============================================>| slowstart headway 1276 (from the RTT) 1277 \__________________________/ \_________ ... 1278 one slowstart burst Repeated slowstart bursts 1280 Multiple levels of Slowstart Bursts 1282 Figure 2 1284 6.2. Constant window pseudo CBR 1286 Implement pseudo constant bit rate by running a standard self clocked 1287 protocol such as TCP with a fixed window size. If that window size 1288 is test_window, the data rate will be slightly above the target_rate. 1290 Since the test_window is constrained to be an integer number of 1291 packets, for small RTTs or low data rates there may not be 1292 sufficiently precise control over the data rate. Rounding the 1293 test_window up (as defined above) is likely to result in data rates 1294 that are higher than the target rate, but reducing the window by one 1295 packet may result in data rates that are too small. Also cross 1296 traffic potentially raises the RTT, implicitly reducing the rate. 1298 Cross traffic that raises the RTT nearly always makes the test more 1299 strenuous (more demanding for the network path). 1301 Note that Constant window pseudo CBR (and Scanned window pseudo CBR 1302 in the next section) both rely on a self clock which is at least 1303 partially derived from the properties of the subnet under test. This 1304 introduces the possibility that the subnet under test exhibits 1305 behaviors such as extreme RTT fluctuations that prevent these 1306 algorithms from accurately controlling data rates. 1308 A FS-TIDS specifying a constant window CBR test must explicitly 1309 indicate under what conditions errors in the data rate cause tests to 1310 be inconclusive. Conventional paced measurement traffic may be more 1311 appropriate for these environments. 1313 6.3. Scanned window pseudo CBR 1315 Scanned window pseudo CBR is similar to the constant window CBR 1316 described above, except the window is scanned across a range of sizes 1317 designed to include two key events, the onset of queuing and the 1318 onset of packet loss or ECN CE marks. The window is scanned by 1319 incrementing it by one packet every 2*target_window_size delivered 1320 packets. This mimics the additive increase phase of standard Reno 1321 TCP congestion avoidance when delayed ACKs are in effect. Normally 1322 the window increases separated by intervals slightly longer than 1323 twice the target_RTT. 1325 There are two ways to implement this test: one built by applying a 1326 window clamp to standard congestion control in a standard protocol 1327 such as TCP and the other built by stiffening a non-standard 1328 transport protocol. When standard congestion control is in effect, 1329 any losses or ECN CE marks cause the transport to revert to a window 1330 smaller than the clamp such that the scanning clamp loses control the 1331 window size. The NPAD pathdiag tool is an example of this class of 1332 algorithms [Pathdiag]. 1334 Alternatively a non-standard congestion control algorithm can respond 1335 to losses by transmitting extra data, such that it maintains the 1336 specified window size independent of losses or ECN CE marks. Such a 1337 stiffened transport explicitly violates mandatory Internet congestion 1338 control [RFC5681] and is not suitable for in situ testing. It is 1339 only appropriate for engineering testing under laboratory conditions. 1340 The Windowed Ping tool implements such a test [WPING]. The tool 1341 described in the paper has been updated.[mpingSource] 1343 The test procedures in Section 8.2 describe how to the partition the 1344 scans into regions and how to interpret the results. 1346 6.4. Concurrent or channelized testing 1348 The procedures described in this document are only directly 1349 applicable to single stream measurement, e.g. one TCP connection or 1350 measurement stream. In an ideal world, we would disallow all 1351 performance claims based multiple concurrent streams, but this is not 1352 practical due to at least two issues. First, many very high rate 1353 link technologies are channelized and at last partially pin the flow 1354 to channel mapping to minimize packet reordering within flows. 1355 Second, TCP itself has scaling limits. Although the former problem 1356 might be overcome through different design decisions, the later 1357 problem is more deeply rooted. 1359 All congestion control algorithms that are philosophically aligned 1360 with the standard [RFC5681] (e.g. claim some level of TCP 1361 compatibility, friendliness or fairness) have scaling limits, in the 1362 sense that as a long fast network (LFN) with a fixed RTT and MTU gets 1363 faster, these congestion control algorithms get less accurate and as 1364 a consequence have difficulty filling the network [CCscaling]. These 1365 properties are a consequence of the original Reno AIMD congestion 1366 control design and the requirement in [RFC5681] that all transport 1367 protocols have similar responses to congestion. 1369 There are a number of reasons to want to specify performance in terms 1370 of multiple concurrent flows, however this approach is not 1371 recommended for data rates below several megabits per second, which 1372 can be attained with run lengths under 10000 packets on many paths. 1373 Since the required run length goes as the square of the data rate, at 1374 higher rates the run lengths can be unreasonably large, and multiple 1375 flows might be the only feasible approach. 1377 If multiple flows are deemed necessary to meet aggregate performance 1378 targets then this must be stated in both the design of the TIDS and 1379 in any claims about network performance. The IP diagnostic tests 1380 must be performed concurrently with the specified number of 1381 connections. For the tests that use bursty test streams, the bursts 1382 should be synchronized across streams unless there is a priori 1383 knowledge that the applications have some explicit mechanism to 1384 stagger their own bursts. In the absences of an explicit mechanism 1385 to stagger bursts many network and application artifacts will 1386 sometimes implicitly synchronize bursts. A test that does not 1387 control burst synchronization may be prone to false pass results for 1388 some applications. 1390 7. Interpreting the Results 1392 7.1. Test outcomes 1394 To perform an exhaustive test of a complete network path, each test 1395 of the TIDS is applied to each subpath of the complete path. If any 1396 subpath fails any test then a standard transport protocol running 1397 over the complete path can also be expected to fail to attain the 1398 Target Transport Performance under some conditions. 1400 In addition to passing or failing, a test can be deemed to be 1401 inconclusive for a number of reasons. Proper instrumentation and 1402 treatment of inconclusive outcomes is critical to the accuracy and 1403 robustness of Model Based Metrics. Tests can be inconclusive if the 1404 precomputed traffic pattern or data rates were not accurately 1405 generated; the measurement results were not statistically 1406 significant; and others causes such as failing to meet some required 1407 preconditions for the test. See Section 5.4 1409 For example consider a test that implements Constant Window Pseudo 1410 CBR (Section 6.2) by adding rate controls and detailed IP packet 1411 transfer instrumentation to TCP (e.g. [RFC4898]). TCP includes 1412 built in control systems which might interfere with the sending data 1413 rate. If such a test meets the required packet transfer statistics 1414 (e.g. run length) while failing to attain the specified data rate it 1415 must be treated as an inconclusive result, because we can not a 1416 priori determine if the reduced data rate was caused by a TCP problem 1417 or a network problem, or if the reduced data rate had a material 1418 effect on the observed packet transfer statistics. 1420 Note that for capacity tests, if the observed packet transfer 1421 statistics meet the statistical criteria for failing (accepting 1422 hypnosis H1 in Section 7.2), the test can can be considered to have 1423 failed because it doesn't really matter that the test didn't attain 1424 the required data rate. 1426 The really important new properties of MBM, such as vantage 1427 independence, are a direct consequence of opening the control loops 1428 in the protocols, such that the test stream does not depend on 1429 network conditions or IP packets received. Any mechanism that 1430 introduces feedback between the path's measurements and the test 1431 stream generation is at risk of introducing nonlinearities that spoil 1432 these properties. Any exceptional event that indicates that such 1433 feedback has happened should cause the test to be considered 1434 inconclusive. 1436 One way to view inconclusive tests is that they reflect situations 1437 where a test outcome is ambiguous between limitations of the network 1438 and some unknown limitation of the IP diagnostic test itself, which 1439 may have been caused by some uncontrolled feedback from the network. 1441 Note that procedures that attempt to search the target parameter 1442 space to find the limits on some parameter such as target_data_rate 1443 are at risk of breaking the location independent properties of Model 1444 Based Metrics, if any part of the boundary between passing and 1445 inconclusive or failing results is sensitive to RTT (which is 1446 normally the case). For example the maximum data rate for a marginal 1447 link (e.g. exhibiting excess errors) is likely to be sensitive to 1448 the test_path_RTT. The maximum observed data rate over the test path 1449 has very little value for predicting the maximum rate over a 1450 different path. 1452 One of the goals for evolving TIDS designs will be to keep sharpening 1453 distinction between inconclusive, passing and failing tests. The 1454 criteria for for passing, failing and inconclusive tests must be 1455 explicitly stated for every test in the TIDS or FS-TIDS. 1457 One of the goals of evolving the testing process, procedures, tools 1458 and measurement point selection should be to minimize the number of 1459 inconclusive tests. 1461 It may be useful to keep raw packet transfer statistics and ancillary 1462 metrics [RFC3148] for deeper study of the behavior of the network 1463 path and to measure the tools themselves. Raw packet transfer 1464 statistics can help to drive tool evolution. Under some conditions 1465 it might be possible to re-evaluate the raw data for satisfying 1466 alternate Target Transport Performance. However it is important to 1467 guard against sampling bias and other implicit feedback which can 1468 cause false results and exhibit measurement point vantage 1469 sensitivity. Simply applying different delivery criteria based on a 1470 different Target Transport Performance is insufficient if the test 1471 traffic patterns (bursts, etc.) does not match the alternate Target 1472 Transport Performance. 1474 7.2. Statistical criteria for estimating run_length 1476 When evaluating the observed run_length, we need to determine 1477 appropriate packet stream sizes and acceptable error levels for 1478 efficient measurement. In practice, can we compare the empirically 1479 estimated packet loss and ECN Congestion Experienced (CE) marking 1480 ratios with the targets as the sample size grows? How large a sample 1481 is needed to say that the measurements of packet transfer indicate a 1482 particular run length is present? 1484 The generalized measurement can be described as recursive testing: 1485 send packets (individually or in patterns) and observe the packet 1486 transfer performance (packet loss ratio or other metric, any marking 1487 we define). 1489 As each packet is sent and measured, we have an ongoing estimate of 1490 the performance in terms of the ratio of packet loss or ECN CE mark 1491 to total packets (i.e. an empirical probability). We continue to 1492 send until conditions support a conclusion or a maximum sending limit 1493 has been reached. 1495 We have a target_mark_probability, 1 mark per target_run_length, 1496 where a "mark" is defined as a lost packet, a packet with ECN CE 1497 mark, or other signal. This constitutes the null Hypothesis: 1499 H0: no more than one mark in target_run_length = 1500 3*(target_window_size)^2 packets 1502 and we can stop sending packets if on-going measurements support 1503 accepting H0 with the specified Type I error = alpha (= 0.05 for 1504 example). 1506 We also have an alternative Hypothesis to evaluate: if performance is 1507 significantly lower than the target_mark_probability. Based on 1508 analysis of typical values and practical limits on measurement 1509 duration, we choose four times the H0 probability: 1511 H1: one or more marks in (target_run_length/4) packets 1513 and we can stop sending packets if measurements support rejecting H0 1514 with the specified Type II error = beta (= 0.05 for example), thus 1515 preferring the alternate hypothesis H1. 1517 H0 and H1 constitute the Success and Failure outcomes described 1518 elsewhere in the memo, and while the ongoing measurements do not 1519 support either hypothesis the current status of measurements is 1520 inconclusive. 1522 The problem above is formulated to match the Sequential Probability 1523 Ratio Test (SPRT) [Wald45] and [Montgomery90]. Note that as 1524 originally framed the events under consideration were all 1525 manufacturing defects. In networking, ECN CE marks and lost packets 1526 are not defects but signals, indicating that the transport protocol 1527 should slow down. 1529 The Sequential Probability Ratio Test also starts with a pair of 1530 hypothesis specified as above: 1532 H0: p0 = one defect in target_run_length 1533 H1: p1 = one defect in target_run_length/4 1534 As packets are sent and measurements collected, the tester evaluates 1535 the cumulative defect count against two boundaries representing H0 1536 Acceptance or Rejection (and acceptance of H1): 1538 Acceptance line: Xa = -h1 + s*n 1539 Rejection line: Xr = h2 + s*n 1541 where n increases linearly for each packet sent and 1543 h1 = { log((1-alpha)/beta) }/k 1544 h2 = { log((1-beta)/alpha) }/k 1545 k = log{ (p1(1-p0)) / (p0(1-p1)) } 1546 s = [ log{ (1-p0)/(1-p1) } ]/k 1548 for p0 and p1 as defined in the null and alternative Hypotheses 1549 statements above, and alpha and beta as the Type I and Type II 1550 errors. 1552 The SPRT specifies simple stopping rules: 1554 o Xa < defect_count(n) < Xr: continue testing 1555 o defect_count(n) <= Xa: Accept H0 1556 o defect_count(n) >= Xr: Accept H1 1558 The calculations above are implemented in the R-tool for Statistical 1559 Analysis [Rtool] , in the add-on package for Cross-Validation via 1560 Sequential Testing (CVST) [CVST]. 1562 Using the equations above, we can calculate the minimum number of 1563 packets (n) needed to accept H0 when x defects are observed. For 1564 example, when x = 0: 1566 Xa = 0 = -h1 + s*n 1567 and n = h1 / s 1569 Note that the derivations in [Wald45] and [Montgomery90] differ. 1570 Montgomery's simplified derivation of SPRT may assume a Bernoulli 1571 processes, where the packet loss probabilities are independent and 1572 identically distributed, making the SPRT more accessible. Wald's 1573 seminal paper showed that this assumption is not necessary. It helps 1574 to remember that the goal of SPRT is not to estimate the value of the 1575 packet loss rate, but only whether or not the packet loss ratio is 1576 likely low enough (when we accept the H0 null hypothesis) yielding 1577 success; or too high (when we accept the H1 alternate hypothesis) 1578 yielding failure. 1580 7.3. Reordering Tolerance 1582 All tests must be instrumented for packet level reordering [RFC4737]. 1583 However, there is no consensus for how much reordering should be 1584 acceptable. Over the last two decades the general trend has been to 1585 make protocols and applications more tolerant to reordering (see for 1586 example [RFC4015]), in response to the gradual increase in reordering 1587 in the network. This increase has been due to the deployment of 1588 technologies such as multithreaded routing lookups and Equal Cost 1589 MultiPath (ECMP) routing. These techniques increase parallelism in 1590 network and are critical to enabling overall Internet growth to 1591 exceed Moore's Law. 1593 Note that transport retransmission strategies can trade off 1594 reordering tolerance vs how quickly they can repair losses vs 1595 overhead from spurious retransmissions. In advance of new 1596 retransmission strategies we propose the following strawman: 1597 Transport protocols should be able to adapt to reordering as long as 1598 the reordering extent is not more than the maximum of one quarter 1599 window or 1 mS, whichever is larger. (These values come from 1600 experience prototyping Early Retransmit [RFC5827] and related 1601 algorithms. They agree with the values being proposed for "RACK: a 1602 time-based fast loss detection algorithm" [I-D.ietf-tcpm-rack].) 1603 Within this limit on reorder extent, there should be no bound on 1604 reordering density. 1606 By implication, recording which is less than these bounds should not 1607 be treated as a network impairment. However [RFC4737] still applies: 1608 reordering should be instrumented and the maximum reordering that can 1609 be properly characterized by the test (because of the bound on 1610 history buffers) should be recorded with the measurement results. 1612 Reordering tolerance and diagnostic limitations, such as the size of 1613 the history buffer used to diagnose packets that are way out-of- 1614 order, must be specified in a FSTIDS. 1616 8. IP Diagnostic Tests 1618 The IP diagnostic tests below are organized according to the 1619 technique used to generate the test stream as described in Section 6. 1620 All of the results are evaluated in accordance with Section 7, 1621 possibly with additional test specific critera. 1623 We also introduce some combined tests which are more efficient when 1624 networks are expected to pass, but conflate diagnostic signatures 1625 when they fail. 1627 8.1. Basic Data Rate and Packet Transfer Tests 1629 We propose several versions of the basic data rate and packet 1630 transfer statistics test that differ in how the data rate is 1631 controlled. The data can be paced on a timer, or window controlled 1632 (and self clocked). The first two tests implicitly confirm that 1633 sub_path has sufficient raw capacity to carry the target_data_rate. 1634 They are recommended for relatively infrequent testing, such as an 1635 installation or periodic auditing process. The third, background 1636 packet transfer statistics, is a low rate test designed for ongoing 1637 monitoring for changes in subpath quality. 1639 8.1.1. Delivery Statistics at Paced Full Data Rate 1641 Confirm that the observed run length is at least the 1642 target_run_length while relying on timer to send data at the 1643 target_rate using the procedure described in in Section 6.1 with a 1644 burst size of 1 (single packets) or 2 (packet pairs). 1646 The test is considered to be inconclusive if the packet transmission 1647 can not be accurately controlled for any reason. 1649 RFC 6673 [RFC6673] is appropriate for measuring packet transfer 1650 statistics at full data rate. 1652 8.1.2. Delivery Statistics at Full Data Windowed Rate 1654 Confirm that the observed run length is at least the 1655 target_run_length while sending at an average rate approximately 1656 equal to the target_data_rate, by controlling (or clamping) the 1657 window size of a conventional transport protocol to test_window. 1659 Since losses and ECN CE marks cause transport protocols to reduce 1660 their data rates, this test is expected to be less precise about 1661 controlling its data rate. It should not be considered inconclusive 1662 as long as at least some of the round trips reached the full 1663 target_data_rate without incurring losses or ECN CE marks. To pass 1664 this test the network must deliver target_window_size packets in 1665 target_RTT time without any losses or ECN CE marks at least once per 1666 two target_window_size round trips, in addition to meeting the run 1667 length statistical test. 1669 8.1.3. Background Packet Transfer Statistics Tests 1671 The background run length is a low rate version of the target target 1672 rate test above, designed for ongoing lightweight monitoring for 1673 changes in the observed subpath run length without disrupting users. 1674 It should be used in conjunction with one of the above full rate 1675 tests because it does not confirm that the subpath can support raw 1676 data rate. 1678 RFC 6673 [RFC6673] is appropriate for measuring background packet 1679 transfer statistics. 1681 8.2. Standing Queue Tests 1683 These engineering tests confirm that the bottleneck is well behaved 1684 across the onset of packet loss, which typically follows after the 1685 onset of queuing. Well behaved generally means lossless for 1686 transient queues, but once the queue has been sustained for a 1687 sufficient period of time (or reaches a sufficient queue depth) there 1688 should be a small number of losses or ECN CE marks to signal to the 1689 transport protocol that it should reduce its window or data rate. 1690 Losses that are too early can prevent the transport from averaging at 1691 the target_data_rate. Losses that are too late indicate that the 1692 queue might not have an appropriate AQM [RFC7567] and as a 1693 consequence subject to bufferbloat [wikiBloat]. Queues without AQM 1694 have the potential to inflict excess delays on all flows sharing the 1695 bottleneck. Excess losses (more than half of the window) at the 1696 onset of loss make loss recovery problematic for the transport 1697 protocol. Non-linear, erratic or excessive RTT increases suggest 1698 poor interactions between the channel acquisition algorithms and the 1699 transport self clock. All of the tests in this section use the same 1700 basic scanning algorithm, described here, but score the link or 1701 subpath on the basis of how well it avoids each of these problems. 1703 Some network technologies rely on virtual queues or other techniques 1704 to meter traffic without adding any queuing delay, in which case the 1705 data rate will vary with the window size all the way up to the onset 1706 of load induced packet loss or ECN CE marks. For these technologies, 1707 the discussion of queuing in Section 6.3 does not apply, but it is 1708 still necessary to confirm that the onset of losses or ECN CE marks 1709 be at an appropriate point and progressive. If the network 1710 bottleneck does not introduce significant queuing delay, modify the 1711 procedure described in Section 6.3 to start the scan at a window 1712 equal to or slightly smaller than the test_window. 1714 Use the procedure in Section 6.3 to sweep the window across the onset 1715 of queuing and the onset of loss. The tests below all assume that 1716 the scan emulates standard additive increase and delayed ACK by 1717 incrementing the window by one packet for every 2*target_window_size 1718 packets delivered. A scan can typically be divided into three 1719 regions: below the onset of queuing, a standing queue, and at or 1720 beyond the onset of loss. 1722 Below the onset of queuing the RTT is typically fairly constant, and 1723 the data rate varies in proportion to the window size. Once the data 1724 rate reaches the subpath IP rate, the data rate becomes fairly 1725 constant, and the RTT increases in proportion to the increase in 1726 window size. The precise transition across the start of queuing can 1727 be identified by the maximum network power, defined to be the ratio 1728 data rate over the RTT. The network power can be computed at each 1729 window size, and the window with the maximum is taken as the start of 1730 the queuing region. 1732 If there is random background loss (e.g. bit errors, etc), precise 1733 determination of the onset of queue induced packet loss may require 1734 multiple scans. Above the onset of queuing loss, all transport 1735 protocols are expected to experience periodic losses determined by 1736 the interaction between the congestion control and AQM algorithms. 1737 For standard congestion control algorithms the periodic losses are 1738 likely to be relatively widely spaced and the details are typically 1739 dominated by the behavior of the transport protocol itself. For the 1740 stiffened transport protocols case (with non-standard, aggressive 1741 congestion control algorithms) the details of periodic losses will be 1742 dominated by how the window increase function responds to loss. 1744 8.2.1. Congestion Avoidance 1746 A subpath passes the congestion avoidance standing queue test if more 1747 than target_run_length packets are delivered between the onset of 1748 queuing (as determined by the window with the maximum network power 1749 as described above) and the first loss or ECN CE mark. If this test 1750 is implemented using a standard congestion control algorithm with a 1751 clamp, it can be performed in situ in the production internet as a 1752 capacity test. For an example of such a test see [Pathdiag]. 1754 For technologies that do not have conventional queues, use the 1755 test_window in place of the onset of queuing. i.e. A subpath passes 1756 the congestion avoidance standing queue test if more than 1757 target_run_length packets are delivered between start of the scan at 1758 test_window and the first loss or ECN CE mark. 1760 8.2.2. Bufferbloat 1762 This test confirms that there is some mechanism to limit buffer 1763 occupancy (e.g. that prevents bufferbloat). Note that this is not 1764 strictly a requirement for single stream bulk transport capacity, 1765 however if there is no mechanism to limit buffer queue occupancy then 1766 a single stream with sufficient data to deliver is likely to cause 1767 the problems described in [RFC7567], and [wikiBloat]. This may cause 1768 only minor symptoms for the dominant flow, but has the potential to 1769 make the subpath unusable for other flows and applications. 1771 Pass if the onset of loss occurs before a standing queue has 1772 introduced more delay than than twice target_RTT, or other well 1773 defined and specified limit. Note that there is not yet a model for 1774 how much standing queue is acceptable. The factor of two chosen here 1775 reflects a rule of thumb. In conjunction with the previous test, 1776 this test implies that the first loss should occur at a queuing delay 1777 which is between one and two times the target_RTT. 1779 Specified RTT limits that are larger than twice the target_RTT must 1780 be fully justified in the FS-TIDS. 1782 8.2.3. Non excessive loss 1784 This test confirms that the onset of loss is not excessive. Pass if 1785 losses are equal or less than the increase in the cross traffic plus 1786 the test stream window increase since the previous RTT. This could 1787 be restated as non-decreasing total throughput of the subpath at the 1788 onset of loss. (Note that when there is a transient drop in subpath 1789 throughput and there is not already a standing queue, a subpath that 1790 passes other queue tests in this document will have sufficient queue 1791 space to hold one full RTT worth of data). 1793 Note that token bucket policers will not pass this test, which is as 1794 intended. TCP often stumbles badly if more than a small fraction of 1795 the packets are dropped in one RTT. Many TCP implementations will 1796 require a timeout and slowstart to recover their self clock. Even if 1797 they can recover from the massive losses the sudden change in 1798 available capacity at the bottleneck wastes serving and front path 1799 capacity until TCP can adapt to the new rate [Policing]. 1801 8.2.4. Duplex Self Interference 1803 This engineering test confirms a bound on the interactions between 1804 the forward data path and the ACK return path when they share a half 1805 duplex link. 1807 Some historical half duplex technologies had the property that each 1808 direction held the channel until it completely drained its queue. 1809 When a self clocked transport protocol, such as TCP, has data and 1810 ACKs passing in opposite directions through such a link, the behavior 1811 often reverts to stop-and-wait. Each additional packet added to the 1812 window raises the observed RTT by two packet times, once as the 1813 additional packet passes through the data path, and once for the 1814 additional delay incurred by the ACK waiting on the return path. 1816 The duplex self interference test fails if the RTT rises by more than 1817 a fixed bound above the expected queuing time computed from the 1818 excess window divided by the subpath IP Capacity. This bound must be 1819 smaller than target_RTT/2 to avoid reverting to stop and wait 1820 behavior. (e.g. Data packets and ACKs both have to be released at 1821 least twice per RTT.) 1823 8.3. Slowstart tests 1825 These tests mimic slowstart: data is sent at twice the effective 1826 bottleneck rate to exercise the queue at the dominant bottleneck. 1828 8.3.1. Full Window slowstart test 1830 This is a capacity test to confirm that slowstart is not likely to 1831 exit prematurely. Send slowstart bursts that are target_window_size 1832 total packets. 1834 Accumulate packet transfer statistics as described in Section 7.2 to 1835 score the outcome. Pass if it is statistically significant that the 1836 observed number of good packets delivered between losses or ECN CE 1837 marks is larger than the target_run_length. Fail if it is 1838 statistically significant that the observed interval between losses 1839 or ECN CE marks is smaller than the target_run_length. 1841 It is deemed inconclusive if the elapsed time to send the data burst 1842 is not less than half of the time to receive the ACKs. (i.e. It is 1843 acceptable to send data too fast, but sending it slower than twice 1844 the actual bottleneck rate as indicated by the ACKs is deemed 1845 inconclusive). The headway for the slowstart bursts should be the 1846 target_RTT. 1848 Note that these are the same parameters as the Sender Full Window 1849 burst test, except the burst rate is at slowstart rate, rather than 1850 sender interface rate. 1852 8.3.2. Slowstart AQM test 1854 Do a continuous slowstart (send data continuously at twice the 1855 implied IP bottleneck capacity), until the first loss, stop, allow 1856 the network to drain and repeat, gathering statistics on how many 1857 packets were delivered before the loss, the pattern of losses, 1858 maximum observed RTT and window size. Justify the results. There is 1859 not currently sufficient theory justifying requiring any particular 1860 result, however design decisions that affect the outcome of this 1861 tests also affect how the network balances between long and short 1862 flows (the "mice vs elephants" problem). The queue sojourn time for 1863 the first packet delivered after the first loss should be at least 1864 one half of the target_RTT. 1866 This is an engineering test: It should be performed on a quiescent 1867 network or testbed, since cross traffic has the potential to change 1868 the results in ill defined ways. 1870 8.4. Sender Rate Burst tests 1872 These tests determine how well the network can deliver bursts sent at 1873 sender's interface rate. Note that this test most heavily exercises 1874 the front path, and is likely to include infrastructure may be out of 1875 scope for an access ISP, even though the bursts might be caused by 1876 ACK compression, thinning or channel arbitration in the access ISP. 1877 See Appendix B. 1879 Also, there are a several details about sender interface rate bursts 1880 that are not fully defined here. These details, such as the assumed 1881 sender interface rate, should be explicitly stated is a FS-TIDS. 1883 Current standards permit TCP to send full window bursts following an 1884 application pause. (Congestion Window Validation [RFC2861] and 1885 updates to support Rate-Limited Traffic [RFC7661], are not required). 1886 Since full window bursts are consistent with standard behavior, it is 1887 desirable that the network be able to deliver such bursts, otherwise 1888 application pauses will cause unwarranted losses. Note that the AIMD 1889 sawtooth requires a peak window that is twice target_window_size, so 1890 the worst case burst may be 2*target_window_size. 1892 It is also understood in the application and serving community that 1893 interface rate bursts have a cost to the network that has to be 1894 balanced against other costs in the servers themselves. For example 1895 TCP Segmentation Offload (TSO) reduces server CPU in exchange for 1896 larger network bursts, which increase the stress on network buffer 1897 memory. Some newer TCP implementations can pace traffic at scale 1898 [TSO_pacing][TSO_fq_pacing]. It remains to be determined if and how 1899 quickly these changes will be deployed. 1901 There is not yet theory to unify these costs or to provide a 1902 framework for trying to optimize global efficiency. We do not yet 1903 have a model for how much server rate bursts should be tolerated by 1904 the network. Some bursts must be tolerated by the network, but it is 1905 probably unreasonable to expect the network to be able to efficiently 1906 deliver all data as a series of bursts. 1908 For this reason, this is the only test for which we encourage 1909 derating. A TIDS could include a table of pairs of derating 1910 parameters: burst sizes and how much each burst size is permitted to 1911 reduce the run length, relative to to the target_run_length. 1913 8.5. Combined and Implicit Tests 1915 Combined tests efficiently confirm multiple network properties in a 1916 single test, possibly as a side effect of normal content delivery. 1917 They require less measurement traffic than other testing strategies 1918 at the cost of conflating diagnostic signatures when they fail. 1919 These are by far the most efficient for monitoring networks that are 1920 nominally expected to pass all tests. 1922 8.5.1. Sustained Bursts Test 1924 The sustained burst test implements a combined worst case version of 1925 all of the capacity tests above. It is simply: 1927 Send target_window_size bursts of packets at server interface rate 1928 with target_RTT burst headway (burst start to next burst start). 1929 Verify that the observed packet transfer statistics meets the 1930 target_run_length. 1932 Key observations: 1934 o The subpath under test is expected to go idle for some fraction of 1935 the time, determined by the difference between the time to drain 1936 the queue at the subpath_IP_capacity, and the target_RTT. If the 1937 queue does not drain completely it may be an indication that the 1938 the subpath has insufficient IP capacity or that there is some 1939 other problem with the test (e.g. inconclusive). 1940 o The burst sensitivity can be derated by sending smaller bursts 1941 more frequently. E.g. send target_window_size*derate packet 1942 bursts every target_RTT*derate, where "derate" is less than one. 1943 o When not derated, this test is the most strenuous capacity test. 1944 o A subpath that passes this test is likely to be able to sustain 1945 higher rates (close to subpath_IP_capacity) for paths with RTTs 1946 significantly smaller than the target_RTT. 1947 o This test can be implemented with instrumented TCP [RFC4898], 1948 using a specialized measurement application at one end [MBMSource] 1949 and a minimal service at the other end [RFC0863] [RFC0864]. 1950 o This test is efficient to implement, since it does not require 1951 per-packet timers, and can make use of TSO in modern NIC hardware. 1952 o If a subpath is known to pass the Standing Queue engineering tests 1953 (particularly that it has a progressive onset of loss at an 1954 appropriate queue depth), then the Sustained Burst Test is 1955 sufficient to assure that the subpath under test will not impair 1956 Bulk Transport Capacity at the target performance under all 1957 conditions. See Section 8.2 for a discussion of the standing 1958 queue tests. 1960 Note that this test is clearly independent of the subpath RTT, or 1961 other details of the measurement infrastructure, as long as the 1962 measurement infrastructure can accurately and reliably deliver the 1963 required bursts to the subpath under test. 1965 8.5.2. Passive Measurements 1967 Any non-throughput maximizing application, such as fixed rate 1968 streaming media, can be used to implement passive or hybrid (defined 1969 in [RFC7799]) versions of Model Based Metrics with some additional 1970 instrumentation and possibly a traffic shaper or other controls in 1971 the servers. The essential requirement is that the data transmission 1972 be constrained such that even with arbitrary application pauses and 1973 bursts, the data rate and burst sizes stay within the envelope 1974 defined by the individual tests described above. 1976 If the application's serving data rate can be constrained to be less 1977 than or equal to the target_data_rate and the serving_RTT (the RTT 1978 between the sender and client) is less than the target_RTT, this 1979 constraint is most easily implemented by clamping the transport 1980 window size to serving_window_clamp, set to the test_window, computed 1981 for the actual serving path. 1983 Under the above constraints the serving_window_clamp will limit the 1984 both the serving data rate and burst sizes to be no larger than the 1985 procedures in Section 8.1.2 and Section 8.4 or Section 8.5.1. Since 1986 the serving RTT is smaller than the target_RTT, the worst case bursts 1987 that might be generated under these conditions will be smaller than 1988 called for by Section 8.4 and the sender rate burst sizes are 1989 implicitly derated by the serving_window_clamp divided by the 1990 target_window_size at the very least. (Depending on the application 1991 behavior, the data might be significantly smoother than specified by 1992 any of the burst tests.) 1994 In an alternative implementation the data rate and bursts might be 1995 explicitly controlled by a programmable traffic shaper or pacing at 1996 the sender. This would provide better control over transmissions but 1997 is more complicated to implement, although the required technology is 1998 available [TSO_pacing][TSO_fq_pacing]. 2000 Note that these techniques can be applied to any content delivery 2001 that can operated at a constrained data rate to inhibit TCP 2002 equilibrium behavior. 2004 Furthermore note that Dynamic Adaptive Streaming over HTTP (DASH) is 2005 generally in conflict with passive Model Based Metrics measurement, 2006 because it is a rate maximizing protocol. It can still meet the 2007 requirement here if the rate can be capped, for example by knowing a 2008 priori the maximum rate needed to deliver a particular piece of 2009 content. 2011 9. An Example 2013 In this section we illustrate a TIDS designed to confirm that an 2014 access ISP can reliably deliver HD video from multiple content 2015 providers to all of their customers. With modern codecs, minimal HD 2016 video (720p) generally fits in 2.5 Mb/s. Due to their geographical 2017 size, network topology and modem characteristics the ISP determines 2018 that most content is within a 50 mS RTT of their users (This example 2019 RTT is a sufficient to cover the propagation delay to continental 2020 Europe or either US coast with low delay modems or somewhat smaller 2021 geographical regions if the modems require additional delay to 2022 implement advanced compression and error recovery). 2024 2.5 Mb/s over a 50 ms path 2026 +----------------------+-------+---------+ 2027 | End-to-End Parameter | value | units | 2028 +----------------------+-------+---------+ 2029 | target_rate | 2.5 | Mb/s | 2030 | target_RTT | 50 | ms | 2031 | target_MTU | 1500 | bytes | 2032 | header_overhead | 64 | bytes | 2033 | | | | 2034 | target_window_size | 11 | packets | 2035 | target_run_length | 363 | packets | 2036 +----------------------+-------+---------+ 2038 Table 1 2040 Table 1 shows the default TCP model with no derating, and as such is 2041 quite conservative. The simplest TIDS would be to use the sustained 2042 burst test, described in Section 8.5.1. Such a test would send 11 2043 packet bursts every 50mS, and confirming that there was no more than 2044 1 packet loss per 33 bursts (363 total packets in 1.650 seconds). 2046 Since this number represents is the entire end-to-end loss budget, 2047 independent subpath tests could be implemented by apportioning the 2048 packet loss ratio across subpaths. For example 50% of the losses 2049 might be allocated to the access or last mile link to the user, 40% 2050 to the network interconnections with other ISPs and 1% to each 2051 internal hop (assuming no more than 10 internal hops). Then all of 2052 the subpaths can be tested independently, and the spatial composition 2053 of passing subpaths would be expected to be within the end-to-end 2054 loss budget. 2056 9.1. Observations about applicability 2058 Guidance on deploying and using MBM belong in a future document. 2059 However this example illustrates some the issues that may need to be 2060 considered. 2062 Note that another ISP, with different geographical coverage, topology 2063 or modem technology may need to assume a different target_RTT, and as 2064 a consequence different target_window_size and target_run_length, 2065 even for the same target_data rate. One of the implications of this 2066 is that infrastructure shared by multiple ISPs, such as inter- 2067 exchange points (IXPs) and other interconnects may need to be 2068 evaluated on the basis of the most stringent target_window_size and 2069 target_run_length of any participating ISP. One way to do this might 2070 be to choose target parameters for evaluating such shared 2071 infrastructure on the basis of a hypothetical reference path that 2072 does not necessarily match any actual paths. 2074 Testing interconnects has generally been problematic: conventional 2075 performance tests run between measurement points adjacent to either 2076 side of the interconnect are not generally useful. Unconstrained TCP 2077 tests, such as iPerf [iPerf] are usually overly aggressive due to the 2078 small RTT (often less than 1 mS). With a short RTT these tools are 2079 likely to report inflated data rates because on a short RTT these 2080 tools can tolerate very high packet loss ratios and can push other 2081 cross traffic off of the network. As a consequence these 2082 measurements are useless for predicting actual user performance over 2083 longer paths, and may themselves be quite disruptive. Model Based 2084 Metrics solves this problem. The interconnect can be evaluated with 2085 the same TIDS as other subpaths. Continuing our example, if the 2086 interconnect is apportioned 40% of the losses, 11 packet bursts sent 2087 every 50mS should have fewer than one loss per 82 bursts (902 2088 packets). 2090 10. Validation 2092 Since some aspects of the models are likely to be too conservative, 2093 Section 5.2 permits alternate protocol models and Section 5.3 permits 2094 test parameter derating. If either of these techniques are used, we 2095 require demonstrations that such a TIDS can robustly detect subpaths 2096 that will prevent authentic applications using state-of-the-art 2097 protocol implementations from meeting the specified Target Transport 2098 Performance. This correctness criteria is potentially difficult to 2099 prove, because it implicitly requires validating a TIDS against all 2100 possible paths and subpaths. The procedures described here are still 2101 experimental. 2103 We suggest two approaches, both of which should be applied: first, 2104 publish a fully open description of the TIDS, including what 2105 assumptions were used and and how it was derived, such that the 2106 research community can evaluate the design decisions, test them and 2107 comment on their applicability; and second, demonstrate that 2108 applications do meet the Target Transport Performance when running 2109 over a network testbed which has the tightest possible constraints 2110 that still allow the tests in the TIDS to pass. 2112 This procedure resembles an epsilon-delta proof in calculus. 2113 Construct a test network such that all of the individual tests of the 2114 TIDS pass by only small (infinitesimal) margins, and demonstrate that 2115 a variety of authentic applications running over real TCP 2116 implementations (or other protocols as appropriate) meets the Target 2117 Transport Performance over such a network. The workloads should 2118 include multiple types of streaming media and transaction oriented 2119 short flows (e.g. synthetic web traffic). 2121 For example, for the HD streaming video TIDS described in Section 9, 2122 the IP capacity should be exactly the header_overhead above 2.5 Mb/s, 2123 the per packet random background loss ratio should be 1/363, for a 2124 run length of 363 packets, the bottleneck queue should be 11 packets 2125 and the front path should have just enough buffering to withstand 11 2126 packet interface rate bursts. We want every one of the TIDS tests to 2127 fail if we slightly increase the relevant test parameter, so for 2128 example sending a 12 packet burst should cause excess (possibly 2129 deterministic) packet drops at the dominant queue at the bottleneck. 2130 This network has the tightest possible constraints that can be 2131 expected to pass the TIDS, yet it should be possible for a real 2132 application using a stock TCP implementation in the vendor's default 2133 configuration to attain 2.5 Mb/s over an 50 mS path. 2135 The most difficult part of setting up such a testbed is arranging for 2136 it to have the tightest possible constraints that still allow it to 2137 pass the individual tests. Two approaches are suggested: 2138 constraining (configuring) the network devices not to use all 2139 available resources (e.g. by limiting available buffer space or data 2140 rate); and pre-loading subpaths with cross traffic. Note that is it 2141 important that a single tightly constrained environment just barely 2142 passes all tests, otherwise there is a chance that TCP can exploit 2143 extra latitude in some parameters (such as data rate) to partially 2144 compensate for constraints in other parameters (queue space, or vice- 2145 versa). 2147 To the extent that a TIDS is used to inform public dialog it should 2148 be fully publicly documented, including the details of the tests, 2149 what assumptions were used and how it was derived. All of the 2150 details of the validation experiment should also be published with 2151 sufficient detail for the experiments to be replicated by other 2152 researchers. All components should either be open source of fully 2153 described proprietary implementations that are available to the 2154 research community. 2156 11. Security Considerations 2158 Measurement is often used to inform business and policy decisions, 2159 and as a consequence is potentially subject to manipulation. Model 2160 Based Metrics are expected to be a huge step forward because 2161 equivalent measurements can be performed from multiple vantage 2162 points, such that performance claims can be independently validated 2163 by multiple parties. 2165 Much of the acrimony in the Net Neutrality debate is due to the 2166 historical lack of any effective vantage independent tools to 2167 characterize network performance. Traditional methods for measuring 2168 Bulk Transport Capacity are sensitive to RTT and as a consequence 2169 often yield very different results when run local to an ISP or 2170 interconnect and when run over a customer's complete path. Neither 2171 the ISP nor customer can repeat the others measurements, leading to 2172 high levels of distrust and acrimony. Model Based Metrics are 2173 expected to greatly improve this situation. 2175 Note that in situ measurements sometimes requires sending synthetic 2176 measurement traffic between arbitrary locations in the network, and 2177 as such are potentially attractive platforms for launching DDOS 2178 attacks. All active measurement tools and protocols must be designed 2179 to minimize the opportunities for these misuses. See the discussion 2180 in section 7 of [RFC7594]. 2182 Some of the tests described in the note are not intended for frequent 2183 network monitoring since they have the potential to cause high 2184 network loads and might adversely affect other traffic. 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. Informative 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 [RFC3168] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition 2227 of Explicit Congestion Notification (ECN) to IP", 2228 RFC 3168, DOI 10.17487/RFC3168, September 2001, 2229 . 2231 [RFC3465] Allman, M., "TCP Congestion Control with Appropriate Byte 2232 Counting (ABC)", RFC 3465, February 2003. 2234 [RFC4015] Ludwig, R. and A. Gurtov, "The Eifel Response Algorithm 2235 for TCP", RFC 4015, February 2005. 2237 [RFC4737] Morton, A., Ciavattone, L., Ramachandran, G., Shalunov, 2238 S., and J. Perser, "Packet Reordering Metrics", RFC 4737, 2239 November 2006. 2241 [RFC4898] Mathis, M., Heffner, J., and R. Raghunarayan, "TCP 2242 Extended Statistics MIB", RFC 4898, May 2007. 2244 [RFC5136] Chimento, P. and J. Ishac, "Defining Network Capacity", 2245 RFC 5136, February 2008. 2247 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion 2248 Control", RFC 5681, September 2009. 2250 [RFC5827] Allman, M., Avrachenkov, K., Ayesta, U., Blanton, J., and 2251 P. Hurtig, "Early Retransmit for TCP and Stream Control 2252 Transmission Protocol (SCTP)", RFC 5827, 2253 DOI 10.17487/RFC5827, May 2010, 2254 . 2256 [RFC5835] Morton, A. and S. Van den Berghe, "Framework for Metric 2257 Composition", RFC 5835, April 2010. 2259 [RFC6049] Morton, A. and E. Stephan, "Spatial Composition of 2260 Metrics", RFC 6049, January 2011. 2262 [RFC6576] Geib, R., Ed., Morton, A., Fardid, R., and A. Steinmitz, 2263 "IP Performance Metrics (IPPM) Standard Advancement 2264 Testing", BCP 176, RFC 6576, DOI 10.17487/RFC6576, March 2265 2012, . 2267 [RFC6673] Morton, A., "Round-Trip Packet Loss Metrics", RFC 6673, 2268 August 2012. 2270 [RFC6928] Chu, J., Dukkipati, N., Cheng, Y., and M. Mathis, 2271 "Increasing TCP's Initial Window", RFC 6928, 2272 DOI 10.17487/RFC6928, April 2013, 2273 . 2275 [RFC7312] Fabini, J. and A. Morton, "Advanced Stream and Sampling 2276 Framework for IP Performance Metrics (IPPM)", RFC 7312, 2277 August 2014. 2279 [RFC7398] Bagnulo, M., Burbridge, T., Crawford, S., Eardley, P., and 2280 A. Morton, "A Reference Path and Measurement Points for 2281 Large-Scale Measurement of Broadband Performance", 2282 RFC 7398, February 2015. 2284 [RFC7567] Baker, F., Ed. and G. Fairhurst, Ed., "IETF 2285 Recommendations Regarding Active Queue Management", 2286 BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015, 2287 . 2289 [RFC7594] Eardley, P., Morton, A., Bagnulo, M., Burbridge, T., 2290 Aitken, P., and A. Akhter, "A Framework for Large-Scale 2291 Measurement of Broadband Performance (LMAP)", RFC 7594, 2292 DOI 10.17487/RFC7594, September 2015, 2293 . 2295 [RFC7661] Fairhurst, G., Sathiaseelan, A., and R. Secchi, "Updating 2296 TCP to Support Rate-Limited Traffic", RFC 7661, 2297 DOI 10.17487/RFC7661, October 2015, 2298 . 2300 [RFC7680] Almes, G., Kalidindi, S., Zekauskas, M., and A. Morton, 2301 Ed., "A One-Way Loss Metric for IP Performance Metrics 2302 (IPPM)", STD 82, RFC 7680, DOI 10.17487/RFC7680, January 2303 2016, . 2305 [RFC7799] Morton, A., "Active and Passive Metrics and Methods (with 2306 Hybrid Types In-Between)", RFC 7799, DOI 10.17487/RFC7799, 2307 May 2016, . 2309 [I-D.ietf-tcpm-rack] 2310 Cheng, Y., Cardwell, N., and N. Dukkipati, "RACK: a time- 2311 based fast loss detection algorithm for TCP", draft-ietf- 2312 tcpm-rack-02 (work in progress), March 2017. 2314 [MSMO97] Mathis, M., Semke, J., Mahdavi, J., and T. Ott, "The 2315 Macroscopic Behavior of the TCP Congestion Avoidance 2316 Algorithm", Computer Communications Review volume 27, 2317 number3, July 1997. 2319 [WPING] Mathis, M., "Windowed Ping: An IP Level Performance 2320 Diagnostic", INET 94, June 1994. 2322 [mpingSource] 2323 Fan, X., Mathis, M., and D. Hamon, "Git Repository for 2324 mping: An IP Level Performance Diagnostic", Sept 2013, 2325 . 2327 [MBMSource] 2328 Hamon, D., Stuart, S., and H. Chen, "Git Repository for 2329 Model Based Metrics", Sept 2013, . 2332 [Pathdiag] 2333 Mathis, M., Heffner, J., O'Neil, P., and P. Siemsen, 2334 "Pathdiag: Automated TCP Diagnosis", Passive and Active 2335 Measurement , June 2008. 2337 [iPerf] Wikipedia Contributors, , "iPerf", Wikipedia, The Free 2338 Encyclopedia , cited March 2015, 2339 . 2342 [Wald45] Wald, A., "Sequential Tests of Statistical Hypotheses", 2343 The Annals of Mathematical Statistics, Vol. 16, No. 2, pp. 2344 117-186, Published by: Institute of Mathematical 2345 Statistics, Stable URL: 2346 http://www.jstor.org/stable/2235829, June 1945. 2348 [Montgomery90] 2349 Montgomery, D., "Introduction to Statistical Quality 2350 Control - 2nd ed.", ISBN 0-471-51988-X, 1990. 2352 [Rtool] R Development Core Team, , "R: A language and environment 2353 for statistical computing. R Foundation for Statistical 2354 Computing, Vienna, Austria. ISBN 3-900051-07-0, URL 2355 http://www.R-project.org/", , 2011. 2357 [CVST] Krueger, T. and M. Braun, "R package: Fast Cross- 2358 Validation via Sequential Testing", version 0.1, 11 2012. 2360 [AFD] Pan, R., Breslau, L., Prabhakar, B., and S. Shenker, 2361 "Approximate fairness through differential dropping", 2362 SIGCOMM Comput. Commun. Rev. 33, 2, April 2003. 2364 [wikiBloat] 2365 Wikipedia, , "Bufferbloat", http://en.wikipedia.org/ 2366 w/ index.php?title=Bufferbloat&oldid=608805474, March 2367 2015. 2369 [CCscaling] 2370 Fernando, F., Doyle, J., and S. Steven, "Scalable laws for 2371 stable network congestion control", Proceedings of 2372 Conference on Decision and 2373 Control, http://www.ee.ucla.edu/~paganini, December 2001. 2375 [TSO_pacing] 2376 Corbet, J., "TSO sizing and the FQ scheduler", 2377 LWN.net https://lwn.net/Articles/564978/, Aug 2013. 2379 [TSO_fq_pacing] 2380 Dumazet, E. and Y. Chen, "TSO, fair queuing, pacing: 2381 three's a charm", Proceedings of IETF 88, TCPM WG 2382 https://www.ietf.org/proceedings/88/slides/slides-88-tcpm- 2383 9.pdf, Nov 2013. 2385 [Policing] 2386 Flach, T., Papageorge, P., Terzis, A., Pedrosa, L., Cheng, 2387 Y., Karim, T., Katz-Bassett, E., and R. Govindan, "An 2388 Internet-Wide Analysis of Traffic Policing", ACM SIGCOMM , 2389 August 2016. 2391 Appendix A. Model Derivations 2393 The reference target_run_length described in Section 5.2 is based on 2394 very conservative assumptions: that all excess data in flight 2395 (window) above the target_window_size contributes to a standing queue 2396 that raises the RTT, and that classic Reno congestion control with 2397 delayed ACKs are in effect. In this section we provide two 2398 alternative calculations using different assumptions. 2400 It may seem out of place to allow such latitude in a measurement 2401 method, but this section provides offsetting requirements. 2403 The estimates provided by these models make the most sense if network 2404 performance is viewed logarithmically. In the operational Internet, 2405 data rates span more than 8 orders of magnitude, RTT spans more than 2406 3 orders of magnitude, and packet loss ratio spans at least 8 orders 2407 of magnitude if not more. When viewed logarithmically (as in 2408 decibels), these correspond to 80 dB of dynamic range. On an 80 dB 2409 scale, a 3 dB error is less than 4% of the scale, even though it 2410 represents a factor of 2 in untransformed parameter. 2412 This document gives a lot of latitude for calculating 2413 target_run_length, however people designing a TIDS should consider 2414 the effect of their choices on the ongoing tussle about the relevance 2415 of "TCP friendliness" as an appropriate model for Internet capacity 2416 allocation. Choosing a target_run_length that is substantially 2417 smaller than the reference target_run_length specified in Section 5.2 2418 strengthens the argument that it may be appropriate to abandon "TCP 2419 friendliness" as the Internet fairness model. This gives developers 2420 incentive and permission to develop even more aggressive applications 2421 and protocols, for example by increasing the number of connections 2422 that they open concurrently. 2424 A.1. Queueless Reno 2426 In Section 5.2 models were derived based on the assumption that the 2427 subpath IP rate matches the target rate plus overhead, such that the 2428 excess window needed for the AIMD sawtooth causes a fluctuating queue 2429 at the bottleneck. 2431 An alternate situation would be a bottleneck where there is no 2432 significant queue and losses are caused by some mechanism that does 2433 not involve extra delay, for example by the use of a virtual queue as 2434 done in Approximate Fair Dropping [AFD]. A flow controlled by such a 2435 bottleneck would have a constant RTT and a data rate that fluctuates 2436 in a sawtooth due to AIMD congestion control. Assume the losses are 2437 being controlled to make the average data rate meet some goal which 2438 is equal or greater than the target_rate. The necessary run length 2439 to meet the target_rate can be computed as follows: 2441 For some value of Wmin, the window will sweep from Wmin packets to 2442 2*Wmin packets in 2*Wmin RTT (due to delayed ACK). Unlike the 2443 queuing case where Wmin = target_window_size, we want the average of 2444 Wmin and 2*Wmin to be the target_window_size, so the average data 2445 rate is the target rate. Thus we want Wmin = 2446 (2/3)*target_window_size. 2448 Between losses each sawtooth delivers (1/2)(Wmin+2*Wmin)(2Wmin) 2449 packets in 2*Wmin round trip times. 2451 Substituting these together we get: 2453 target_run_length = (4/3)(target_window_size^2) 2455 Note that this is 44% of the reference_run_length computed earlier. 2456 This makes sense because under the assumptions in Section 5.2 the 2457 AMID sawtooth caused a queue at the bottleneck, which raised the 2458 effective RTT by 50%. 2460 Appendix B. The effects of ACK scheduling 2462 For many network technologies simple queuing models don't apply: the 2463 network schedules, thins or otherwise alters the timing of ACKs and 2464 data, generally to raise the efficiency of the channel allocation 2465 algorithms when confronted with relatively widely spaced small ACKs. 2466 These efficiency strategies are ubiquitous for half duplex, wireless 2467 and broadcast media. 2469 Altering the ACK stream by holding or thinning ACKs typically has two 2470 consequences: it raises the implied bottleneck IP capacity, making 2471 the fine grained slowstart bursts either faster or larger and it 2472 raises the effective RTT by the average time that the ACKs and data 2473 are delayed. The first effect can be partially mitigated by re- 2474 clocking ACKs once they are beyond the bottleneck on the return path 2475 to the sender, however this further raises the effective RTT. 2477 The most extreme example of this sort of behavior would be a half 2478 duplex channel that is not released as long as the endpoint currently 2479 holding the channel has more traffic (data or ACKs) to send. Such 2480 environments cause self clocked protocols under full load to revert 2481 to extremely inefficient stop and wait behavior. The channel 2482 constrains the protocol to send an entire window of data as a single 2483 contiguous burst on the forward path, followed by the entire window 2484 of ACKs on the return path. 2486 If a particular return path contains a subpath or device that alters 2487 the timing of the ACK stream, then the entire front path from the 2488 sender up to the bottleneck must be tested at the burst parameters 2489 implied by the ACK scheduling algorithm. The most important 2490 parameter is the Implied Bottleneck IP Capacity, which is the average 2491 rate at which the ACKs advance snd.una. Note that thinning the ACK 2492 stream (relying on the cumulative nature of seg.ack to permit 2493 discarding some ACKs) causes most TCP implementations to send 2494 interface rate bursts to offset the longer times between ACKs in 2495 order to maintain the average data rate. 2497 Note that due to ubiquitous self clocking in Internet protocols, ill 2498 conceived channel allocation mechanisms are likely to increases the 2499 queuing stress on the front path because they cause larger full 2500 sender rate data bursts. 2502 Holding data or ACKs for channel allocation or other reasons (such as 2503 forward error correction) always raises the effective RTT relative to 2504 the minimum delay for the path. Therefore it may be necessary to 2505 replace target_RTT in the calculation in Section 5.2 by an 2506 effective_RTT, which includes the target_RTT plus a term to account 2507 for the extra delays introduced by these mechanisms. 2509 Appendix C. Version Control 2511 This section to be removed prior to publication. 2513 Formatted: Thu Apr 7 18:12:37 PDT 2016 2515 Authors' Addresses 2517 Matt Mathis 2518 Google, Inc 2519 1600 Amphitheater Parkway 2520 Mountain View, California 94043 2521 USA 2523 Email: mattmathis@google.com 2525 Al Morton 2526 AT&T Labs 2527 200 Laurel Avenue South 2528 Middletown, NJ 07748 2529 USA 2531 Phone: +1 732 420 1571 2532 Email: acmorton@att.com