idnits 2.17.1 draft-ietf-ippm-model-based-metrics-08.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 1446 has weird spacing: '... and n = h1...' -- The document date (July 8, 2016) is 2849 days in the past. Is this intentional? Checking references for intended status: Experimental ---------------------------------------------------------------------------- -- Obsolete informational reference (is this intentional?): RFC 2861 (Obsoleted by RFC 7661) Summary: 0 errors (**), 0 flaws (~~), 2 warnings (==), 2 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 IP Performance Working Group M. Mathis 3 Internet-Draft Google, Inc 4 Intended status: Experimental A. Morton 5 Expires: January 9, 2017 AT&T Labs 6 July 8, 2016 8 Model Based Metrics for Bulk Transport Capacity 9 draft-ietf-ippm-model-based-metrics-08.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 key infrastructure, such as interconnects or even 18 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 For Bulk Transport Capacity, the IP diagnostics are built on test 23 streams that mimic TCP over the complete path and statistical 24 criteria for evaluating the packet transfer statistics of those 25 streams. The temporal structure of the test stream (bursts, etc) 26 mimic TCP or other transport protocol carrying bulk data over a long 27 path but are constructed to be independent of the details of the 28 subpath under test, end systems or applications. Likewise the 29 success criteria evaluates the packet transfer statistics of the 30 subpath against criteria determined by protocol performance models 31 applied to the Target Transport Performance of the complete path. 32 The success criteria also does not depend on the details of the 33 subpath, end systems or application. 35 Model Based Metrics exhibit several important new properties not 36 present in other Bulk Transport Capacity Metrics, including the 37 ability to reason about concatenated or overlapping subpaths. The 38 results are vantage independent which is critical for supporting 39 independent validation of tests by comparing results from multiple 40 measurement points. 42 This document does not define the IP diagnostic tests, but provides a 43 framework for designing suites of IP diagnostic tests that are 44 tailored to confirming that infrastructure can meet the predetermined 45 Target Transport Performance. 47 Status of This Memo 49 This Internet-Draft is submitted in full conformance with the 50 provisions of BCP 78 and BCP 79. 52 Internet-Drafts are working documents of the Internet Engineering 53 Task Force (IETF). Note that other groups may also distribute 54 working documents as Internet-Drafts. The list of current Internet- 55 Drafts is at http://datatracker.ietf.org/drafts/current/. 57 Internet-Drafts are draft documents valid for a maximum of six months 58 and may be updated, replaced, or obsoleted by other documents at any 59 time. It is inappropriate to use Internet-Drafts as reference 60 material or to cite them other than as "work in progress." 62 This Internet-Draft will expire on January 9, 2017. 64 Copyright Notice 66 Copyright (c) 2016 IETF Trust and the persons identified as the 67 document authors. All rights reserved. 69 This document is subject to BCP 78 and the IETF Trust's Legal 70 Provisions Relating to IETF Documents 71 (http://trustee.ietf.org/license-info) in effect on the date of 72 publication of this document. Please review these documents 73 carefully, as they describe your rights and restrictions with respect 74 to this document. Code Components extracted from this document must 75 include Simplified BSD License text as described in Section 4.e of 76 the Trust Legal Provisions and are provided without warranty as 77 described in the Simplified BSD License. 79 Table of Contents 81 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3 82 1.1. Version Control . . . . . . . . . . . . . . . . . . . . . 5 83 2. Overview . . . . . . . . . . . . . . . . . . . . . . . . . . 7 84 3. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 9 85 4. Background . . . . . . . . . . . . . . . . . . . . . . . . . 15 86 4.1. TCP properties . . . . . . . . . . . . . . . . . . . . . 17 87 4.2. Diagnostic Approach . . . . . . . . . . . . . . . . . . . 19 88 4.3. New requirements relative to RFC 2330 . . . . . . . . . . 20 89 5. Common Models and Parameters . . . . . . . . . . . . . . . . 20 90 5.1. Target End-to-end parameters . . . . . . . . . . . . . . 21 91 5.2. Common Model Calculations . . . . . . . . . . . . . . . . 21 92 5.3. Parameter Derating . . . . . . . . . . . . . . . . . . . 22 93 5.4. Test Preconditions . . . . . . . . . . . . . . . . . . . 23 94 6. Generating test streams . . . . . . . . . . . . . . . . . . . 23 95 6.1. Mimicking slowstart . . . . . . . . . . . . . . . . . . . 24 96 6.2. Constant window pseudo CBR . . . . . . . . . . . . . . . 25 97 6.3. Scanned window pseudo CBR . . . . . . . . . . . . . . . . 26 98 6.4. Concurrent or channelized testing . . . . . . . . . . . . 27 99 7. Interpreting the Results . . . . . . . . . . . . . . . . . . 27 100 7.1. Test outcomes . . . . . . . . . . . . . . . . . . . . . . 27 101 7.2. Statistical criteria for estimating run_length . . . . . 29 102 7.3. Reordering Tolerance . . . . . . . . . . . . . . . . . . 31 103 8. IP Diagnostic Tests . . . . . . . . . . . . . . . . . . . . . 32 104 8.1. Basic Data Rate and Packet Transfer Tests . . . . . . . . 32 105 8.1.1. Delivery Statistics at Paced Full Data Rate . . . . . 33 106 8.1.2. Delivery Statistics at Full Data Windowed Rate . . . 33 107 8.1.3. Background Packet Transfer Statistics Tests . . . . . 33 108 8.2. Standing Queue Tests . . . . . . . . . . . . . . . . . . 34 109 8.2.1. Congestion Avoidance . . . . . . . . . . . . . . . . 35 110 8.2.2. Bufferbloat . . . . . . . . . . . . . . . . . . . . . 35 111 8.2.3. Non excessive loss . . . . . . . . . . . . . . . . . 36 112 8.2.4. Duplex Self Interference . . . . . . . . . . . . . . 36 113 8.3. Slowstart tests . . . . . . . . . . . . . . . . . . . . . 37 114 8.3.1. Full Window slowstart test . . . . . . . . . . . . . 37 115 8.3.2. Slowstart AQM test . . . . . . . . . . . . . . . . . 37 116 8.4. Sender Rate Burst tests . . . . . . . . . . . . . . . . . 38 117 8.5. Combined and Implicit Tests . . . . . . . . . . . . . . . 39 118 8.5.1. Sustained Bursts Test . . . . . . . . . . . . . . . . 39 119 8.5.2. Streaming Media . . . . . . . . . . . . . . . . . . . 40 120 9. An Example . . . . . . . . . . . . . . . . . . . . . . . . . 40 121 9.1. Observations about applicability . . . . . . . . . . . . 41 122 10. Validation . . . . . . . . . . . . . . . . . . . . . . . . . 42 123 11. Security Considerations . . . . . . . . . . . . . . . . . . . 43 124 12. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 44 125 13. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 44 126 14. References . . . . . . . . . . . . . . . . . . . . . . . . . 44 127 14.1. Normative References . . . . . . . . . . . . . . . . . . 44 128 14.2. Informative References . . . . . . . . . . . . . . . . . 44 129 Appendix A. Model Derivations . . . . . . . . . . . . . . . . . 48 130 A.1. Queueless Reno . . . . . . . . . . . . . . . . . . . . . 48 131 Appendix B. The effects of ACK scheduling . . . . . . . . . . . 49 132 Appendix C. Version Control . . . . . . . . . . . . . . . . . . 50 133 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 50 135 1. Introduction 137 Model Based Metrics (MBM) rely on peer-reviewed mathematical models 138 to specify a Targeted Suite of IP Diagnostic tests, designed to 139 assess whether common transport protocols can be expected to meet a 140 predetermined Target Transport Performance over an Internet path. 141 This note describes the modeling framework to derive the test 142 parameters for assessing an Internet path's ability to support a 143 predetermined Bulk Transport Capacity. 145 Each test in the Targeted IP Diagnostic Suite (TIDS) measures some 146 aspect of IP packet transfer needed to meet the Target Transport 147 Performance. For Bulk Transport Capacity the TIDS includes IP 148 diagnostic tests to verify that there is: sufficient IP capacity 149 (data rate); sufficient queue space at bottlenecks to absorb and 150 deliver typical transport bursts; and that the background packet loss 151 ratio is low enough not to interfere with congestion control; and 152 other properties described below. Unlike typical IPPM metrics which 153 yield measures of network properties, Model Based Metrics nominally 154 yield pass/fail evaluations of the ability of standard transport 155 protocols to meet the specific performance objective over some 156 network path. 158 In most cases, the IP diagnostic tests can be implemented by 159 combining existing IPPM metrics with additional controls for 160 generating test streams having a specified temporal structure (bursts 161 or standing queues caused by constant bit rate streams, etc.) and 162 statistical criteria for evaluating packet transfer. The temporal 163 structure of the test streams mimic transport protocol behavior over 164 the complete path, the statistical criteria models the transport 165 protocol's response to less than ideal IP packet transfer. 167 This note addresses Bulk Transport Capacity. It describes an 168 alternative to the approach presented in "A Framework for Defining 169 Empirical Bulk Transfer Capacity Metrics" [RFC3148]. In the future, 170 other Model Based Metrics may cover other applications and 171 transports, such as VoIP over UDP and RTP, and new transport 172 protocols. 174 The MBM approach, mapping Target Transport Performance to a Targeted 175 IP Diagnostic Suite (TIDS) of IP tests, solves some intrinsic 176 problems with using TCP or other throughput maximizing protocols for 177 measurement. In particular all throughput maximizing protocols (and 178 TCP congestion control in particular) cause some level of congestion 179 in order to detect when they have reached the available capacity 180 limitation of the network. This self inflicted congestion obscures 181 the network properties of interest and introduces non-linear dynamic 182 equilibrium behaviors that make any resulting measurements useless as 183 metrics because they have no predictive value for conditions or paths 184 different than that of the measurement itself. In order to prevent 185 these effects it is necessary to suppress the effects of TCP 186 congestion control in the measurement method. These issues are 187 discussed at length in Section 4. 189 A Targeted IP Diagnostic Suite does not have such difficulties. IP 190 diagnostics can be constructed such that they make strong statistical 191 statements about path properties that are independent of the 192 measurement details, such as vantage and choice of measurement 193 points. Model Based Metrics are designed to bridge the gap between 194 empirical IP measurements and expected TCP performance for multiple 195 standardized versions of TCP. 197 1.1. Version Control 199 RFC Editor: Please remove this entire subsection prior to 200 publication. 202 Please send comments about this draft to ippm@ietf.org. See 203 http://goo.gl/02tkD for more information including: interim drafts, 204 an up to date todo list and information on contributing. 206 Formatted: Fri Jul 8 16:00:10 PDT 2016 208 Changes since -07 draft: 210 o Sharpened the use of "statistical criteria" 211 o Sharpened the definition of test_window, and removed related 212 redundant text in several places 213 o Clarified "equilibrium" as "dynamic equilibrium, similar to 214 processes observed in chemistry" 215 o Properly explained "Heisenberg" as "observer effect" 216 o Added the observation from RFC 6576 that HW and SW congestion 217 control implementations do not generally give the same results. 218 o Noted that IP and application metrics differ as to how overhead is 219 handled. MBM is explicit about how it handles overhead. 220 o Clarified the language and added a new reference about the 221 problems caused by token bucket policers. 222 o Added an subsection in the example that comments on some of issues 223 that need to be mentioned in a future usage or applicability doc. 224 o Updated ippm-2680-bis to RFC7680 225 o Many terminology, punctuation and spelling nits. 227 Changes since -06 draft: 229 o More language nits: 231 * "Targeted IP Diagnostic Suite (TIDS)" replaces "Targeted 232 Diagnostic Suite (TDS)". 233 * "implied bottleneck IP capacity" replaces "implied bottleneck 234 IP rate". 235 * Updated to ECN CE Marks. 236 * Added "specified temporal structure" 237 * "test stream" replaces "test traffic" 238 * "packet transfer" replaces "packet delivery" 239 * Reworked discussion of slowstart, bursts and pacing. 240 * RFC 7567 replaces RFC 2309. 242 Changes since -05 draft: 244 o Wordsmithing on sections overhauled in -05 draft. 245 o Reorganized the document: 247 * Relocated subsection "Preconditions". 248 * Relocated subsection "New Requirements relative to RFC 2330". 249 o Addressed nits and not so nits by Ruediger Geib. (Thanks!) 250 o Substantially tightened the entire definitions section. 251 o Many terminology changes, to better conform to other docs : 253 * IP rate and IP capacity (following RFC 5136) replaces various 254 forms of link data rate. 255 * subpath replaces link. 256 * target_window_size replaces target_pipe_size. 257 * implied bottleneck IP rate replaces effective bottleneck link 258 rate. 259 * Packet delivery statistics replaces delivery statistics. 261 Changes since -04 draft: 263 o The introduction was heavily overhauled: split into a separate 264 introduction and overview. 265 o The new shorter introduction: 267 * Is a problem statement; 268 * This document provides a framework; 269 * That it replaces TCP measurement by IP tests; 270 * That the results are pass/fail. 271 o Added a diagram of the framework to the overview 272 o and introduces all of the elements of the framework. 273 o Renumbered sections, reducing the depth of some section numbers. 274 o Updated definitions to better agree with other documents: 276 * Reordered section 2 277 * Bulk [data] performance -> Bulk Transport Capacity, everywhere 278 including the title. 279 * loss rate and loss probability -> packet loss ratio 280 * end-to-end path -> complete path 281 * [end-to-end][target] performance -> Target Transport 282 Performance 283 * load test -> capacity test 285 2. Overview 287 This document describes a modeling framework for deriving a Targeted 288 IP Diagnostic Suite from a predetermined Target Transport 289 Performance. It is not a complete specification, and relies on other 290 standards documents to define important details such as packet Type-P 291 selection, sampling techniques, vantage selection, etc. We imagine 292 Fully Specified - Targeted IP Diagnostic Suites (FS-TIDS), that 293 define all of these details. We use Targeted IP Diagnostic Suite 294 (TIDS) to refer to the subset of such a specification that is in 295 scope for this document. This terminology is defined in Section 3. 297 Section 4 describes some key aspects of TCP behavior and what they 298 imply about the requirements for IP packet transfer. Most of the IP 299 diagnostic tests needed to confirm that the path meets these 300 properties can be built on existing IPPM metrics, with the addition 301 of statistical criteria for evaluating packet transfer and in a few 302 cases, new mechanisms to implement the required temporal structure. 303 (One group of tests, the standing queue tests described in 304 Section 8.2, don't correspond to existing IPPM metrics, but suitable 305 metrics can be patterned after the existing definitions.) 307 Figure 1 shows the MBM modeling and measurement framework. The 308 Target Transport Performance, at the top of the figure, is determined 309 by the needs of the user or application, outside the scope of this 310 document. For Bulk Transport Capacity, the main performance 311 parameter of interest is the Target Data Rate. However, since TCP's 312 ability to compensate for less than ideal network conditions is 313 fundamentally affected by the Round Trip Time (RTT) and the Maximum 314 Transmission Unit (MTU) of the complete path, these parameters must 315 also be specified in advance based on knowledge about the intended 316 application setting. They may reflect a specific application over a 317 real path through the Internet or an idealized application and 318 hypothetical path representing a typical user community. Section 5 319 describes the common parameters and models derived from the Target 320 Transport Performance. 322 Target Transport Performance 323 (Target Data Rate, Target RTT and Target MTU) 324 | 325 ________V_________ 326 | mathematical | 327 | models | 328 | | 329 ------------------ 330 Traffic parameters | | Statistical criteria 331 | | 332 _______V____________V____Targeted_______ 333 | | * * * | Diagnostic Suite | 334 _____|_______V____________V________________ | 335 __|____________V____________V______________ | | 336 | IP diagnostic tests | | | 337 | | | | | | 338 | _____________V__ __V____________ | | | 339 | | traffic | | Delivery | | | | 340 | | pattern | | Evaluation | | | | 341 | | generation | | | | | | 342 | -------v-------- ------^-------- | | | 343 | | v test stream via ^ | | |-- 344 | | -->======================>-- | | | 345 | | subpath under test | |- 346 ----V----------------------------------V--- | 347 | | | | | | 348 V V V V V V 349 fail/inconclusive pass/fail/inconclusive 351 Overall Modeling Framework 353 Figure 1 355 The mathematical models are used to determine Traffic parameters and 356 subsequently to design traffic patterns that mimic TCP or other 357 transport protocol delivering bulk data and operating at the Target 358 Data Rate, MTU and RTT over a full range of conditions, including 359 flows that are bursty at multiple time scales. The traffic patterns 360 are generated based on the three Target parameters of complete path 361 and independent of the properties of individual subpaths using the 362 techniques described in Section 6. As much as possible the test 363 stream is generated deterministically (precomputed) to minimize the 364 extent to which test methodology, measurement points, measurement 365 vantage or path partitioning affect the details of the measurement 366 traffic. 368 Section 7 describes packet transfer statistics and methods to test 369 them against the statistical criteria provided by the mathematical 370 models. Since the statistical criteria are typically for the 371 complete path (a composition of subpaths) [RFC6049], in situ testing 372 requires that the end-to-end statistical criteria be apportioned as 373 separate criteria for each subpath. Subpaths that are expected to be 374 bottlenecks would then be permitted to contribute a larger fraction 375 of the end-to-end packet loss budget. In compensation, non- 376 bottlenecked subpaths have to be constrained to contribute less 377 packet loss. Thus the statistical criteria for each subpath in each 378 test of a TIDS is an apportioned share of the end-to-end statistical 379 criteria for the complete path which was determined by the 380 mathematical model. 382 Section 8 describes the suite of individual tests needed to verify 383 all of required IP delivery properties. A subpath passes if and only 384 if all of the individual IP diagnostic tests pass. Any subpath that 385 fails any test indicates that some users are likely to fail to attain 386 their Target Transport Performance under some conditions. In 387 addition to passing or failing, a test can be deemed to be 388 inconclusive for a number of reasons including: the precomputed 389 traffic pattern was not accurately generated; the measurement results 390 were not statistically significant; and others such as failing to 391 meet some required test preconditions. If all tests pass but some 392 are inconclusive, then the entire suite is deemed to be inconclusive. 394 In Section 9 we present an example TIDS that might be representative 395 of HD video, and illustrate how Model Based Metrics can be used to 396 address difficult measurement situations, such as confirming that 397 inter-carrier exchanges have sufficient performance and capacity to 398 deliver HD video between ISPs. 400 Since there is some uncertainty in the modeling process, Section 10 401 describes a validation procedure to diagnose and minimize false 402 positive and false negative results. 404 3. Terminology 406 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 407 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 408 document are to be interpreted as described in [RFC2119]. 410 Note that terms containing underscores (rather than spaces) appear in 411 equations in the modeling sections. In some cases both forms are 412 used for aesthetic reasons, they do not have different meanings. 414 General Terminology: 416 Target: A general term for any parameter specified by or derived 417 from the user's application or transport performance requirements. 419 Target Transport Performance: Application or transport performance 420 target values for the complete path. For Bulk Transport Capacity 421 defined in this note the Target Transport Performance includes the 422 Target Data Rate, Target RTT and Target MTU as described below. 423 Target Data Rate: The specified application data rate required for 424 an application's proper operation. Conventional BTC metrics are 425 focused on the Target Data Rate, however these metrics had little 426 or no predictive value because they do not consider the effects of 427 the other two parameters of the Target Transport Performance, the 428 RTT and MTU of the complete paths. 429 Target RTT (Round Trip Time): The specified baseline (minimum) RTT 430 of the longest complete path over which the user expects to be 431 able meet the target performance. TCP and other transport 432 protocol's ability to compensate for path problems is generally 433 proportional to the number of round trips per second. The Target 434 RTT determines both key parameters of the traffic patterns (e.g. 435 burst sizes) and the thresholds on acceptable IP packet transfer 436 statistics. The Target RTT must be specified considering 437 appropriate packets sizes: MTU sized packets on the forward path, 438 ACK sized packets (typically header_overhead) on the return path. 439 Note that Target RTT is specified and not measured, MBM 440 measurements derived for a given target_RTT will be applicable to 441 any path with a smaller RTTs. 442 Target MTU (Maximum Transmission Unit): The specified maximum MTU 443 supported by the complete path the over which the application 444 expects to meet the target performance. Assume 1500 Byte MTU 445 unless otherwise specified. If some subpath has a smaller MTU, 446 then it becomes the Target MTU for the complete path, and all 447 model calculations and subpath tests must use the same smaller 448 MTU. 449 Targeted IP Diagnostic Suite (TIDS): A set of IP diagnostic tests 450 designed to determine if an otherwise ideal complete path 451 containing the subpath under test can sustain flows at a specific 452 target_data_rate using target_MTU sized packets when the RTT of 453 the complete path is target_RTT. 454 Fully Specified Targeted IP Diagnostic Suite (FS-TIDS): A TIDS 455 together with additional specification such as "type-p", etc which 456 are out of scope for this document, but need to be drawn from 457 other standards documents. 458 Bulk Transport Capacity: Bulk Transport Capacity Metrics evaluate an 459 Internet path's ability to carry bulk data, such as large files, 460 streaming (non-real time) video, and under some conditions, web 461 images and other content. Prior efforts to define BTC metrics 462 have been based on [RFC3148], which predates our understanding of 463 TCP and the requirements described in Section 4 464 IP diagnostic tests: Measurements or diagnostics to determine if 465 packet transfer statistics meet some precomputed target. 467 traffic patterns: The temporal patterns or burstiness of traffic 468 generated by applications over transport protocols such as TCP. 469 There are several mechanisms that cause bursts at various time 470 scales as described in Section 4.1. Our goal here is to mimic the 471 range of common patterns (burst sizes and rates, etc), without 472 tying our applicability to specific applications, implementations 473 or technologies, which are sure to become stale. 474 packet transfer statistics: Raw, detailed or summary statistics 475 about packet transfer properties of the IP layer including packet 476 losses, ECN Congestion Experienced (CE) marks, reordering, or any 477 other properties that may be germane to transport performance. 478 packet loss ratio: As defined in [RFC7680]. 479 apportioned: To divide and allocate, for example budgeting packet 480 loss across multiple subpaths such that the losses will accumulate 481 to less than a specified end-to-end loss ratio. Apportioning 482 metrics is essentially the inverse of the process described in 483 [RFC5835]. 484 open loop: A control theory term used to describe a class of 485 techniques where systems that naturally exhibit circular 486 dependencies can be analyzed by suppressing some of the 487 dependencies, such that the resulting dependency graph is acyclic. 489 Terminology about paths, etc. See [RFC2330] and [RFC7398]. 491 data sender: Host sending data and receiving ACKs. 492 data receiver: Host receiving data and sending ACKs. 493 complete path: The end-to-end path from the data sender to the data 494 receiver. 495 subpath: A portion of the complete path. Note that there is no 496 requirement that subpaths be non-overlapping. A subpath can be a 497 small as a single device, link or interface. 498 measurement point: Measurement points as described in [RFC7398]. 499 test path: A path between two measurement points that includes a 500 subpath of the complete path under test. If the measurement 501 points are off path, the test path may include "test leads" 502 between the measurement points and the subpath. 503 dominant bottleneck: The bottleneck that generally determines most 504 of packet transfer statistics for the entire path. It typically 505 determines a flow's self clock timing, packet loss and ECN 506 Congestion Experienced (CE) marking rate, with other potential 507 bottlenecks having less effect on the packet transfer statistics. 508 See Section 4.1 on TCP properties. 509 front path: The subpath from the data sender to the dominant 510 bottleneck. 511 back path: The subpath from the dominant bottleneck to the receiver. 512 return path: The path taken by the ACKs from the data receiver to 513 the data sender. 515 cross traffic: Other, potentially interfering, traffic competing for 516 network resources (bandwidth and/or queue capacity). 518 Properties determined by the complete path and application. These 519 are described in more detail in Section 5.1. 521 Application Data Rate: General term for the data rate as seen by the 522 application above the transport layer in bytes per second. This 523 is the payload data rate, and explicitly excludes transport and 524 lower level headers (TCP/IP or other protocols), retransmissions 525 and other overhead that is not part to the total quantity of data 526 delivered to the application. 527 IP rate: The actual number of IP-layer bytes delivered through a 528 subpath, per unit time, including TCP and IP headers, retransmits 529 and other TCP/IP overhead. Follows from IP-type-P Link Usage 530 [RFC5136]. 531 IP capacity: The maximum number of IP-layer bytes that can be 532 transmitted through a subpath, per unit time, including TCP and IP 533 headers, retransmits and other TCP/IP overhead. Follows from IP- 534 type-P Link Capacity [RFC5136]. 535 bottleneck IP capacity: The IP capacity of the dominant bottleneck 536 in the forward path. All throughput maximizing protocols estimate 537 this capacity by observing the IP rate delivered through the 538 bottleneck. Most protocols derive their self clocks from the 539 timing of this data. See Section 4.1 and Appendix B for more 540 details. 541 implied bottleneck IP capacity: This is the bottleneck IP capacity 542 implied by the ACKs returning from the receiver. It is determined 543 by looking at how much application data the ACK stream at the 544 sender reports delivered to the data receiver per unit time at 545 various time scales. If the return path is thinning, batching or 546 otherwise altering the ACK timing the implied bottleneck IP 547 capacity over short time scales might be substantially larger than 548 the bottleneck IP capacity averaged over a full RTT. Since TCP 549 derives its clock from the data delivered through the bottleneck 550 the front path must have sufficient buffering to absorb any data 551 bursts at the dimensions (duration and IP rate) implied by the ACK 552 stream, potentially doubled during slowstart. If the return path 553 is not altering the ACK stream, then the implied bottleneck IP 554 capacity will be the same as the bottleneck IP capacity. See 555 Section 4.1 and Appendix B for more details. 556 sender interface rate: The IP rate which corresponds to the IP 557 capacity of the data sender's interface. Due to sender efficiency 558 algorithms including technologies such as TCP segmentation offload 559 (TSO), nearly all moderns servers deliver data in bursts at full 560 interface link rate. Today 1 or 10 Gb/s are typical. 561 Header_overhead: The IP and TCP header sizes, which are the portion 562 of each MTU not available for carrying application payload. 564 Without loss of generality this is assumed to be the size for 565 returning acknowledgments (ACKs). For TCP, the Maximum Segment 566 Size (MSS) is the Target MTU minus the header_overhead. 568 Basic parameters common to models and subpath tests are defined here 569 are described in more detail in Section 5.2. Note that these are 570 mixed between application transport performance (excludes headers) 571 and IP performance (which include TCP headers and retransmissions as 572 part of the IP payload). 574 Window [size]: The total quantity of data plus the data represented 575 by ACKs circulating in the network is referred to as the window. 576 See Section 4.1. Sometimes used with other qualifiers (congestion 577 window, cwnd or receiver window) to indicate which mechanism is 578 controlling the window. 579 pipe size: A general term for number of packets needed in flight 580 (the window size) to exactly fill some network path or subpath. 581 It corresponds to the window size which maximizes network power, 582 the observed data rate divided by the observed RTT. Often used 583 with additional qualifiers to specify which path, or under what 584 conditions, etc. 585 target_window_size: The average number of packets in flight (the 586 window size) needed to meet the Target Data Rate, for the 587 specified Target RTT, and MTU. It implies the scale of the bursts 588 that the network might experience. 589 run length: A general term for the observed, measured, or specified 590 number of packets that are (expected to be) delivered between 591 losses or ECN Congestion Experienced (CE) marks. Nominally one 592 over the sum of the loss and ECN CE marking probabilities, if 593 there are independently and identically distributed. 594 target_run_length: The target_run_length is an estimate of the 595 minimum number of non-congestion marked packets needed between 596 losses or ECN Congestion Experienced (CE) marks necessary to 597 attain the target_data_rate over a path with the specified 598 target_RTT and target_MTU, as computed by a mathematical model of 599 TCP congestion control. A reference calculation is shown in 600 Section 5.2 and alternatives in Appendix A 601 reference target_run_length: target_run_length computed precisely by 602 the method in Section 5.2. This is likely to be slightly more 603 conservative than required by modern TCP implementations. 605 Ancillary parameters used for some tests: 607 derating: Under some conditions the standard models are too 608 conservative. The modeling framework permits some latitude in 609 relaxing or "derating" some test parameters as described in 610 Section 5.3 in exchange for a more stringent TIDS validation 611 procedures, described in Section 10. 613 subpath_IP_capacity: The IP capacity of a specific subpath. 614 test path: A subpath of a complete path under test. 615 test_path_RTT: The RTT observed between two measurement points using 616 packet sizes that are consistent with the transport protocol. 617 This is generally MTU sized packets of the forward path, 618 header_overhead sized packets on the return path. 619 test_path_pipe: The pipe size of a test path. Nominally the test 620 path RTT times the test path IP_capacity. 621 test_window: The smallest window sufficient to meet or exceeded the 622 target_rate when operating with a pure self clock over a test 623 path. The test_window is typically given by 624 ceiling(target_data_rate*test_path_RTT/(target_MTU- 625 header_overhead)) but see the discussion in Appendix B about the 626 effects of channel scheduling on RTT. On some test paths the 627 test_window may need to be adjusted slightly to compensate for the 628 RTT being inflated by the devices that schedule packets. 630 The terminology below is used to define temporal patterns for test 631 stream. These patterns are designed to mimic TCP behavior, as 632 described in Section 4.1. 634 packet headway: Time interval between packets, specified from the 635 start of one to the start of the next. e.g. If packets are sent 636 with a 1 mS headway, there will be exactly 1000 packets per 637 second. 638 burst headway: Time interval between bursts, specified from the 639 start of the first packet one burst to the start of the first 640 packet of the next burst. e.g. If 4 packet bursts are sent with a 641 1 mS burst headway, there will be exactly 4000 packets per second. 642 paced single packets: Send individual packets at the specified rate 643 or packet headway. 644 paced bursts: Send bursts on a timer. Specify any 3 of: average 645 data rate, packet size, burst size (number of packets) and burst 646 headway (burst start to start). By default the bursts are assumed 647 full sender interface rate, such that the packet headway within 648 each burst is the minimum supported by the sender's interface. 649 Under some conditions it is useful to explicitly specify the 650 packet headway within each burst. 651 slowstart rate: Mimic TCP slowstart by sending 4 packet paced bursts 652 at an average data rate equal to twice the implied bottleneck IP 653 capacity (but not more than the sender interface rate). This is a 654 two level burst pattern described in more detail in Section 6.1. 655 If the implied bottleneck IP capacity is more than half of the 656 sender interface rate, slowstart rate becomes sender interface 657 rate. 658 slowstart burst: Mimic one round of TCP slowstart by sending a 659 specified number of packets packets in a two level burst pattern 660 that resembles slowstart. 662 repeated slowstart bursts: Repeat Slowstart bursts once per 663 target_RTT. For TCP each burst would be twice as large as the 664 prior burst, and the sequence would end at the first ECN CE mark 665 or lost packet. For measurement, all slowstart bursts would be 666 the same size (nominally target_window_size but other sizes might 667 be specified), and the ECN CE marks and lost packets are counted. 669 The tests described in this note can be grouped according to their 670 applicability. 672 Capacity tests: Capacity tests determine if a network subpath has 673 sufficient capacity to deliver the Target Transport Performance. 674 As long as the test stream is within the proper envelope for the 675 Target Transport Performance, the average packet losses or ECN 676 Congestion Experienced (CE) marks must be below the statistical 677 criteria computed by the model. As such, capacity tests reflect 678 parameters that can transition from passing to failing as a 679 consequence of cross traffic, additional presented load or the 680 actions of other network users. By definition, capacity tests 681 also consume significant network resources (data capacity and/or 682 queue buffer space), and the test schedules must be balanced by 683 their cost. 684 Monitoring tests: Monitoring tests are designed to capture the most 685 important aspects of a capacity test, but without presenting 686 excessive ongoing load themselves. As such they may miss some 687 details of the network's performance, but can serve as a useful 688 reduced-cost proxy for a capacity test, for example to support 689 ongoing monitoring. 690 Engineering tests: Engineering tests evaluate how network algorithms 691 (such as AQM and channel allocation) interact with TCP-style self 692 clocked protocols and adaptive congestion control based on packet 693 loss and ECN Congestion Experienced (CE) marks. These tests are 694 likely to have complicated interactions with cross traffic and 695 under some conditions can be inversely sensitive to load. For 696 example a test to verify that an AQM algorithm causes ECN CE marks 697 or packet drops early enough to limit queue occupancy may 698 experience a false pass result in the presence of cross traffic. 699 It is important that engineering tests be performed under a wide 700 range of conditions, including both in situ and bench testing, and 701 over a wide variety of load conditions. Ongoing monitoring is 702 less likely to be useful for engineering tests, although sparse in 703 situ testing might be appropriate. 705 4. Background 707 At the time the IPPM WG was chartered, sound Bulk Transport Capacity 708 (BTC) measurement was known to be well beyond our capabilities. Even 709 at the time that Framework for Empirical BTC Metrics [RFC3148] was 710 written we knew that we didn't fully understand the problem. Now, by 711 hindsight we understand why assessing BTC is such a hard problem: 713 o TCP is a control system with circular dependencies - everything 714 affects performance, including components that are explicitly not 715 part of the test (for example, the host processing power is not 716 in-scope of path performance tests). 717 o Congestion control is a dynamic equilibrium process, similar to 718 processes observed in chemistry and other fields. The network and 719 transport protocols find an operating point which balances between 720 opposing forces: the transport protocol pushing harder (raising 721 the data rate and or window) while the network pushes back 722 (raising packet loss ratio, RTT and/or ECN CE marks). By design 723 TCP congestion control keeps raising the data rate until the 724 network gives some indication that its capacity has been exceeded 725 by dropping packets or ECN CE marks. If a TCP sender accurately 726 fills a path to its IP capacity, (e.g. the bottleneck is 100% 727 utilized), then packet losses and ECN CE marks are mostly 728 determined by the TCP sender and how aggressively it seeks 729 additional capacity, and not the network itself, since the network 730 must send exactly the signals that TCP needs to set its rate. 731 o TCP's ability to compensate for network impairments (such as loss, 732 delay and delay variation, outside of those caused by TCP itself) 733 is directly proportional to the number of send-ACK round trip 734 exchanges per second (i.e. inversely proportional to the RTT). As 735 a consequence an impaired subpath may pass a short RTT local test 736 even though it fails when the subpath is extended by an 737 effectively perfect network to some larger RTT. 738 o TCP has an extreme form of the Observer Effect (colloquially know 739 as the Heisenberg effect). Measurement and cross traffic interact 740 in unknown and ill defined ways. The situation is actually worse 741 than the traditional physics problem where you can at least 742 estimate bounds on the relative momentum of the measurement and 743 measured particles. For network measurement you can not in 744 general determine even the order of magnitude of the effect. It 745 is possible to construct measurement scenarios where the 746 measurement traffic starves real user traffic, yielding an overly 747 inflated measurement. The inverse is also possible: the user 748 traffic can fill the network, such that the measurement traffic 749 detects only minimal available capacity. You can not in general 750 determine which scenario might be in effect, so you can not gauge 751 the relative magnitude of the uncertainty introduced by 752 interactions with other network traffic. 753 o It is difficult, if not impossible, for two independent 754 implementations (HW or SW) of TCP congestion control to produce 755 equivalent performance results [RFC6576] under the same network 756 conditions, as an outcome of the other properties listed. 758 These properties are a consequence of the dynamic equilibrium 759 behavior intrinsic to how all throughput maximizing protocols 760 interact with the Internet. These protocols rely on control systems 761 based on estimated network metrics to regulate the quantity of data 762 sent into the network. The packet sending characteristics in turn 763 alter the network properties estimated by the control system metrics, 764 such that there are circular dependencies between every transmission 765 characteristic and every estimated metric. Since some of these 766 dependencies are nonlinear, the entire system is nonlinear, and any 767 change causes a response in packet sending characteristics or 768 estimated network metrics that is difficult to predict. 770 Model Based Metrics overcome these problems by making the measurement 771 system open loop: the packet transfer statistics (akin to the network 772 estimators) do not affect the traffic or traffic patterns (bursts), 773 which are computed on the basis of the Target Transport Performance. 774 A path or subpath meeting the Target Transfer Performance 775 requirements would exhibit packet transfer statistics and estimated 776 metrics that would not cause the control system to slow the traffic 777 below the Target Data Rate. 779 4.1. TCP properties 781 TCP and SCTP are self clocked protocols that carry the vast majority 782 of all Internet data. Their dominant behavior is to have an 783 approximately fixed quantity of data and acknowledgments (ACKs) 784 circulating in the network. The data receiver reports arriving data 785 by returning ACKs to the data sender, the data sender typically 786 responds by sending exactly the same quantity of data back into the 787 network. The total quantity of data plus the data represented by 788 ACKs circulating in the network is referred to as the window. The 789 mandatory congestion control algorithms incrementally adjust the 790 window by sending slightly more or less data in response to each ACK. 791 The fundamentally important property of this system is that it is 792 self clocked: The data transmissions are a reflection of the ACKs 793 that were delivered by the network, the ACKs are a reflection of the 794 data arriving from the network. 796 A number of protocol features cause bursts of data, even in idealized 797 networks that can be modeled as simple queuing systems. 799 During slowstart the IP rate is doubled on each RTT by sending twice 800 as much data as was delivered to the receiver during the prior RTT. 801 Each returning ACK causes the sender to transmit twice the data the 802 ACK reported arriving at the receiver. For slowstart to be able to 803 fill the pipe, the network must be able to tolerate slowstart bursts 804 up to the full pipe size inflated by the anticipated window reduction 805 on the first loss or ECN CE mark. For example, with classic Reno 806 congestion control, an optimal slowstart has to end with a burst that 807 is twice the bottleneck rate for one RTT in duration. This burst 808 causes a queue which is equal to the pipe size (i.e. the window is 809 twice the pipe size) so when the window is halved in response to the 810 first packet loss, the new window will be the pipe size. 812 Note that if the bottleneck IP rate is less that half of the capacity 813 of the front path (which is almost always the case), the slowstart 814 bursts will not by themselves cause significant queues anywhere else 815 along the front path; they primarily exercise the queue at the 816 dominant bottleneck. 818 Several common efficiency algorithms also cause bursts. The self 819 clock is typically applied to groups of packets: the receiver's 820 delayed ACK algorithm generally sends only one ACK per two data 821 segments. Furthermore the modern senders use TCP segmentation 822 offload (TSO) to reduce CPU overhead. The sender's software stack 823 builds super sized TCP segments that the TSO hardware splits into MTU 824 sized segments on the wire. The net effect of TSO, delayed ACK and 825 other efficiency algorithms is to send bursts of segments at full 826 sender interface rate. 828 Note that these efficiency algorithms are almost always in effect, 829 including during slowstart, such that slowstart typically has a two 830 level burst structure. Section 6.1 describes slowstart in more 831 detail. 833 Additional sources of bursts include TCP's initial window [RFC6928], 834 application pauses, channel allocation mechanisms and network devices 835 that schedule ACKs. Appendix B describes these last two items. If 836 the application pauses (stops reading or writing data) for some 837 fraction of an RTT, many TCP implementations catch up to their 838 earlier window size by sending a burst of data at the full sender 839 interface rate. To fill a network with a realistic application, the 840 network has to be able to tolerate sender interface rate bursts large 841 enough to restore the prior window following application pauses. 843 Although the sender interface rate bursts are typically smaller than 844 the last burst of a slowstart, they are at a higher IP rate so they 845 potentially exercise queues at arbitrary points along the front path 846 from the data sender up to and including the queue at the dominant 847 bottleneck. There is no model for how frequent or what sizes of 848 sender rate bursts the network should tolerate. 850 In conclusion, to verify that a path can meet a Target Transport 851 Performance, it is necessary to independently confirm that the path 852 can tolerate bursts at the scales that can be caused by the above 853 mechanisms. Three cases are believed to be sufficient: 855 o Two level slowstart bursts sufficient to get connections started 856 properly. 857 o Ubiquitous sender interface rate bursts caused by efficiency 858 algorithms. We assume 4 packet bursts to be the most common case, 859 since it matches the effects of delayed ACK during slowstart. 860 These bursts should be assumed not to significantly affect packet 861 transfer statistics. 862 o Infrequent sender interface rate bursts that are full 863 target_window_size. Target_run_length may be derated for these 864 large fast bursts. 866 If a subpath can meet the required packet loss ratio for bursts at 867 all of these scales then it has sufficient buffering at all potential 868 bottlenecks to tolerate any of the bursts that are likely introduced 869 by TCP or other transport protocols. 871 4.2. Diagnostic Approach 873 A complete path of a given RTT and MTU, which are equal to or smaller 874 than the Target RTT and equal to or larger than the Target MTU 875 respectively, is expected to be able to attain a specified Bulk 876 Transport Capacity when all of the following conditions are met: 878 1. The IP capacity is above the Target Data Rate by sufficient 879 margin to cover all TCP/IP overheads. This can be confirmed by 880 the tests described in Section 8.1 or any number of IP capacity 881 tests adapted to implement MBM. 882 2. The observed packet transfer statistics are better than required 883 by a suitable TCP performance model (e.g. fewer packet losses or 884 ECN CE marks). See Section 8.1 or any number of low rate packet 885 loss tests outside of MBM. 886 3. There is sufficient buffering at the dominant bottleneck to 887 absorb a slowstart bursts large enough to get the flow out of 888 slowstart at a suitable window size. See Section 8.3. 889 4. There is sufficient buffering in the front path to absorb and 890 smooth sender interface rate bursts at all scales that are likely 891 to be generated by the application, any channel arbitration in 892 the ACK path or any other mechanisms. See Section 8.4. 893 5. When there is a slowly rising standing queue at the bottleneck 894 the onset of packet loss has to be at an appropriate point (time 895 or queue depth) and progressive [RFC7567]. See Section 8.2. 896 6. When there is a standing queue at a bottleneck for a shared media 897 subpath (e.g. half duplex), there must be a suitable bounds on 898 the interaction between ACKs and data, for example due to the 899 channel arbitration mechanism. See Section 8.2.4. 901 Note that conditions 1 through 4 require capacity tests for 902 validation, and thus may need to be monitored on an ongoing basis. 904 Conditions 5 and 6 require engineering tests, which are best 905 performed in controlled environments such as a bench test. They 906 won't generally fail due to load, but may fail in the field due to 907 configuration errors, etc. and should be spot checked. 909 We are developing a tool that can perform many of the tests described 910 here [MBMSource]. 912 4.3. New requirements relative to RFC 2330 914 Model Based Metrics are designed to fulfill some additional 915 requirements that were not recognized at the time RFC 2330 was 916 written [RFC2330]. These missing requirements may have significantly 917 contributed to policy difficulties in the IP measurement space. Some 918 additional requirements are: 920 o IP metrics must be actionable by the ISP - they have to be 921 interpreted in terms of behaviors or properties at the IP or lower 922 layers, that an ISP can test, repair and verify. 923 o Metrics should be spatially composable, such that measures of 924 concatenated paths should be predictable from subpaths. 925 o Metrics must be vantage point invariant over a significant range 926 of measurement point choices, including off path measurement 927 points. The only requirements on MP selection should be that the 928 RTT between the MPs is below some reasonable bound, and that the 929 effects of the "test leads" connecting MPs to the subpath under 930 test can be can be calibrated out of the measurements. The latter 931 might be be accomplished if the test leads are effectively ideal 932 or their properties can be deducted from the measurements between 933 the MPs. While many of tests require that the test leads have at 934 least as much IP capacity as the subpath under test, some do not, 935 for example Background Packet Transfer Tests described in 936 Section 8.1.3. 937 o Metric measurements should be repeatable by multiple parties with 938 no specialized access to MPs or diagnostic infrastructure. It 939 should be possible for different parties to make the same 940 measurement and observe the same results. In particular it is 941 specifically important that both a consumer (or their delegate) 942 and ISP be able to perform the same measurement and get the same 943 result. Note that vantage independence is key to meeting this 944 requirement. 946 5. Common Models and Parameters 947 5.1. Target End-to-end parameters 949 The target end-to-end parameters are the Target Data Rate, Target RTT 950 and Target MTU as defined in Section 3. These parameters are 951 determined by the needs of the application or the ultimate end user 952 and the complete Internet path over which the application is expected 953 to operate. The target parameters are in units that make sense to 954 upper layers: payload bytes delivered to the application, above TCP. 955 They exclude overheads associated with TCP and IP headers, 956 retransmits and other protocols (e.g. DNS). Note that IP-based 957 network services include TCP headers and retransmissions as part of 958 delivered payload, and this difference is recognized in calculations 959 below (header_overhead). 961 Other end-to-end parameters defined in Section 3 include the 962 effective bottleneck data rate, the sender interface data rate and 963 the TCP and IP header sizes. 965 The target_data_rate must be smaller than all subpath IP capacities 966 by enough headroom to carry the transport protocol overhead, 967 explicitly including retransmissions and an allowance for 968 fluctuations in TCP's actual data rate. Specifying a 969 target_data_rate with insufficient headroom is likely to result in 970 brittle measurements having little predictive value. 972 Note that the target parameters can be specified for a hypothetical 973 path, for example to construct TIDS designed for bench testing in the 974 absence of a real application; or for a live in situ test of 975 production infrastructure. 977 The number of concurrent connections is explicitly not a parameter to 978 this model. If a subpath requires multiple connections in order to 979 meet the specified performance, that must be stated explicitly and 980 the procedure described in Section 6.4 applies. 982 5.2. Common Model Calculations 984 The Target Transport Performance is used to derive the 985 target_window_size and the reference target_run_length. 987 The target_window_size, is the average window size in packets needed 988 to meet the target_rate, for the specified target_RTT and target_MTU. 989 It is given by: 991 target_window_size = ceiling( target_rate * target_RTT / ( target_MTU 992 - header_overhead ) ) 993 Target_run_length is an estimate of the minimum required number of 994 unmarked packets that must be delivered between losses or ECN 995 Congestion Experienced (CE) marks, as computed by a mathematical 996 model of TCP congestion control. The derivation here follows 997 [MSMO97], and by design is quite conservative. 999 Reference target_run_length is derived as follows: assume the 1000 subpath_IP_capacity is infinitesimally larger than the 1001 target_data_rate plus the required header_overhead. Then 1002 target_window_size also predicts the onset of queueing. A larger 1003 window will cause a standing queue at the bottleneck. 1005 Assume the transport protocol is using standard Reno style Additive 1006 Increase, Multiplicative Decrease (AIMD) congestion control [RFC5681] 1007 (but not Appropriate Byte Counting [RFC3465]) and the receiver is 1008 using standard delayed ACKs. Reno increases the window by one packet 1009 every pipe_size worth of ACKs. With delayed ACKs this takes 2 Round 1010 Trip Times per increase. To exactly fill the pipe, losses must be no 1011 closer than when the peak of the AIMD sawtooth reached exactly twice 1012 the target_window_size otherwise the multiplicative window reduction 1013 triggered by the loss would cause the network to be underfilled. 1014 Following [MSMO97] the number of packets between losses must be the 1015 area under the AIMD sawtooth. They must be no more frequent than 1016 every 1 in ((3/2)*target_window_size)*(2*target_window_size) packets, 1017 which simplifies to: 1019 target_run_length = 3*(target_window_size^2) 1021 Note that this calculation is very conservative and is based on a 1022 number of assumptions that may not apply. Appendix A discusses these 1023 assumptions and provides some alternative models. If a different 1024 model is used, a fully specified TIDS or FSTIDS MUST document the 1025 actual method for computing target_run_length and ratio between 1026 alternate target_run_length and the reference target_run_length 1027 calculated above, along with a discussion of the rationale for the 1028 underlying assumptions. 1030 These two parameters, target_window_size and target_run_length, 1031 directly imply most of the individual parameters for the tests in 1032 Section 8. 1034 5.3. Parameter Derating 1036 Since some aspects of the models are very conservative, the MBM 1037 framework permits some latitude in derating test parameters. Rather 1038 than trying to formalize more complicated models we permit some test 1039 parameters to be relaxed as long as they meet some additional 1040 procedural constraints: 1042 o The TIDS or FSTIDS MUST document and justify the actual method 1043 used to compute the derated metric parameters. 1044 o The validation procedures described in Section 10 must be used to 1045 demonstrate the feasibility of meeting the Target Transport 1046 Performance with infrastructure that infinitesimally passes the 1047 derated tests. 1048 o The validation process for a FSTIDS itself must be documented is 1049 such a way that other researchers can duplicate the validation 1050 experiments. 1052 Except as noted, all tests below assume no derating. Tests where 1053 there is not currently a well established model for the required 1054 parameters explicitly include derating as a way to indicate 1055 flexibility in the parameters. 1057 5.4. Test Preconditions 1059 Many tests have preconditions which are required to assure their 1060 validity. Examples include: the presence or non-presence of cross 1061 traffic on specific subpaths; negotiating ECN; and appropriate 1062 preamble packet stream to testing to put reactive network elements 1063 into the proper states [RFC7312]. If preconditions are not properly 1064 satisfied for some reason, the tests should be considered to be 1065 inconclusive. In general it is useful to preserve diagnostic 1066 information as to why the preconditions were not met, and any test 1067 data that was collected even if it is not useful for the intended 1068 test. Such diagnostic information and partial test data may be 1069 useful for improving the test in the future. 1071 It is important to preserve the record that a test was scheduled, 1072 because otherwise precondition enforcement mechanisms can introduce 1073 sampling bias. For example, canceling tests due to cross traffic on 1074 subscriber access links might introduce sampling bias in tests of the 1075 rest of the network by reducing the number of tests during peak 1076 network load. 1078 Test preconditions and failure actions MUST be specified in a FSTIDS. 1080 6. Generating test streams 1082 Many important properties of Model Based Metrics, such as vantage 1083 independence, are a consequence of using test streams that have 1084 temporal structures that mimic TCP or other transport protocols 1085 running over a complete path. As described in Section 4.1, self 1086 clocked protocols naturally have burst structures related to the RTT 1087 and pipe size of the complete path. These bursts naturally get 1088 larger (contain more packets) as either the Target RTT or Target Data 1089 Rate get larger, or the Target MTU gets smaller. An implication of 1090 these relationships is that test streams generated by running self 1091 clocked protocols over short subpaths may not adequately exercise the 1092 queuing at any bottleneck to determine if the subpath can support the 1093 full Target Transport Performance over the complete path. 1095 Failing to authentically mimic TCP's temporal structure is part the 1096 reason why simple performance tools such as iPerf, netperf, nc, etc 1097 have the reputation of yielding false pass results over short test 1098 paths, even when some subpath has a flaw. 1100 The definitions in Section 3 are sufficient for most test streams. 1101 We describe the slowstart and standing queue test streams in more 1102 detail. 1104 In conventional measurement practice stochastic processes are used to 1105 eliminate many unintended correlations and sample biases. However 1106 MBM tests are designed to explicitly mimic temporal correlations 1107 caused by network or protocol elements themselves and are intended to 1108 accurately reflect implementation behavior. Some portion of the 1109 system, such as traffic arrival (test scheduling) are naturally 1110 stochastic. Other details, such as protocol processing times, are 1111 technically non-deterministic and might be modeled stochastically, 1112 but are only a tiny part of the overall behavior which is dominated 1113 by implementation specific deterministic effects. Furthermore, it is 1114 known that sampling bias is a real problem for some protocol 1115 implementations. For example TCP's RTT estimator used to determine 1116 the Retransmit Time Out (RTO), can be fooled by periodic cross 1117 traffic or start-stop applications. 1119 At some point in the future it may make sense to introduce fine 1120 grained noise sources into the models used for generating test 1121 streams, but they are not warranted at this time. 1123 6.1. Mimicking slowstart 1125 TCP slowstart has a two level burst structure as shown in Figure 2. 1126 The fine structure is caused by the interaction between the ACK clock 1127 and TCP efficiency algorithms. Each ACK passing through the return 1128 path triggers a small data burst. These bursts are typically full 1129 sender interface rate, with the same headway as the returning ACKs, 1130 but having twice as much data as the ACK reported was delivered to 1131 the receiver. Due to variations in delayed ACK and algorithms such 1132 as Appropriate Byte Counting [RFC3465], different pairs of senders 1133 and receivers produce different burst patterns. Without loss of 1134 generality, we assume each ACK causes 4 packet bursts at an average 1135 headway equal to the ACK headway, and corresponding to sending at an 1136 average rate equal to twice the effective bottleneck IP rate. This 1137 fine structure defines one slowstart rate burst. 1139 For a transport protocol the slowstart bursts are repeated every 1140 target_RTT. Each slowstart burst is twice as large as the previous 1141 burst, and slowstart ends on the first lost packet or ECN mark. For 1142 diagnostic tests described below we preserve the fine structure but 1143 manipulate the burst size and headway to measure the ability of the 1144 dominant bottleneck to absorb and smooth slowstart bursts. 1146 Note that a stream of repeated slowstart bursts has three different 1147 average rates, depending on the averaging interval. At the finest 1148 time scale (a few packet times at the sender interface) the peak of 1149 the average IP rate is the same as the sender interface rate; at a 1150 medium timescale (a few packet times at the dominant bottleneck) the 1151 peak of the average IP rate is twice the implied bottleneck IP 1152 capacity; and at time scales longer than the target_RTT and when the 1153 burst size is equal to the target_window_size the average rate is 1154 equal to the target_data_rate. This pattern corresponds to repeating 1155 the last RTT of TCP slowstart when delayed ACK and sender side byte 1156 counting are present but without the limits specified in Appropriate 1157 Byte Counting [RFC3465]. 1159 time ==> ( - equals one packet) 1160 Packet stream: 1162 ---- ---- ---- ---- ---- ---- ---- ... 1164 |<>| sender interface rate bursts (typically 3 or 4 packets) 1165 |<===>| burst headway (determined by ACK headway) 1166 |<========================>| slowstart burst size(from the window) 1167 |<==============================================>| slowstart headway 1168 \____________ _____________/ \______ __ ... 1169 V V 1170 One slowstart burst Repeated slowstart bursts 1172 Multiple levels of Slowstart Bursts 1174 Figure 2 1176 6.2. Constant window pseudo CBR 1178 Implement pseudo constant bit rate by running a standard self clocked 1179 protocol such as TCP with a fixed window size. If that window size 1180 is test_window, the data rate will be slightly above the target_rate. 1182 Since the test_window is constrained to be an integer number of 1183 packets, for small RTTs or low data rates there may not be 1184 sufficiently precise control over the data rate. Rounding the 1185 test_window up (the default) is likely to result in data rates that 1186 are higher than the target rate, but reducing the window by one 1187 packet may result in data rates that are too small. Also cross 1188 traffic potentially raises the RTT, implicitly reducing the rate. 1189 Cross traffic that raises the RTT nearly always makes the test more 1190 strenuous. A FS-TIDS specifying a constant window CBR tests MUST 1191 explicitly indicate under what conditions errors in the data rate 1192 cause tests to inconclusive. 1194 Since constant window pseudo CBR testing is sensitive to RTT 1195 fluctuations it will be less accurate at controlling the data rate in 1196 environments with fluctuating delays. Conventional paced measurement 1197 traffic may be more appropriate for these environments. 1199 6.3. Scanned window pseudo CBR 1201 Scanned window pseudo CBR is similar to the constant window CBR 1202 described above, except the window is scanned across a range of sizes 1203 designed to include two key events, the onset of queuing and the 1204 onset of packet loss or ECN CE marks. The window is scanned by 1205 incrementing it by one packet every 2*target_window_size delivered 1206 packets. This mimics the additive increase phase of standard Reno 1207 TCP congestion avoidance when delayed ACKs are in effect. Normally 1208 the window increases separated by intervals slightly longer than 1209 twice the target_RTT. 1211 There are two ways to implement this test: one built by applying a 1212 window clamp to standard congestion control in a standard protocol 1213 such as TCP and the other built by stiffening a non-standard 1214 transport protocol. When standard congestion control is in effect, 1215 any losses or ECN CE marks cause the transport to revert to a window 1216 smaller than the clamp such that the scanning clamp loses control the 1217 window size. The NPAD pathdiag tool is an example of this class of 1218 algorithms [Pathdiag]. 1220 Alternatively a non-standard congestion control algorithm can respond 1221 to losses by transmitting extra data, such that it maintains the 1222 specified window size independent of losses or ECN CE marks. Such a 1223 stiffened transport explicitly violates mandatory Internet congestion 1224 control [RFC5681] and is not suitable for in situ testing. It is 1225 only appropriate for engineering testing under laboratory conditions. 1226 The Windowed Ping tool implements such a test [WPING]. The tool 1227 described in the paper has been updated.[mpingSource] 1229 The test procedures in Section 8.2 describe how to the partition the 1230 scans into regions and how to interpret the results. 1232 6.4. Concurrent or channelized testing 1234 The procedures described in this document are only directly 1235 applicable to single stream measurement, e.g. one TCP connection or 1236 measurement stream. In an ideal world, we would disallow all 1237 performance claims based multiple concurrent streams, but this is not 1238 practical due to at least two different issues. First, many very 1239 high rate link technologies are channelized and at last partially pin 1240 the flow to channel mapping to minimize packet reordering within 1241 flows. Second, TCP itself has scaling limits. Although the former 1242 problem might be overcome through different design decisions, the 1243 later problem is more deeply rooted. 1245 All congestion control algorithms that are philosophically aligned 1246 with the standard [RFC5681] (e.g. claim some level of TCP 1247 compatibility, friendliness or fairness) have scaling limits, in the 1248 sense that as a long fast network (LFN) with a fixed RTT and MTU gets 1249 faster, these congestion control algorithms get less accurate and as 1250 a consequence have difficulty filling the network [CCscaling]. These 1251 properties are a consequence of the original Reno AIMD congestion 1252 control design and the requirement in [RFC5681] that all transport 1253 protocols have similar responses to congestion. 1255 There are a number of reasons to want to specify performance in term 1256 of multiple concurrent flows, however this approach is not 1257 recommended for data rates below several megabits per second, which 1258 can be attained with run lengths under 10000 packets on many paths. 1259 Since the required run length goes as the square of the data rate, at 1260 higher rates the run lengths can be unreasonably large, and multiple 1261 flows might be the only feasible approach. 1263 If multiple flows are deemed necessary to meet aggregate performance 1264 targets then this MUST be stated both the design of the TIDS and in 1265 any claims about network performance. The IP diagnostic tests MUST 1266 be performed concurrently with the specified number of connections. 1267 For the tests that use bursty test streams, the bursts should be 1268 synchronized across streams. 1270 7. Interpreting the Results 1272 7.1. Test outcomes 1274 To perform an exhaustive test of a complete network path, each test 1275 of the TIDS is applied to each subpath of the complete path. If any 1276 subpath fails any test then a standard transport protocol running 1277 over the complete path can also be expected to fail to attain the 1278 Target Transport Performance under some conditions. 1280 In addition to passing or failing, a test can be deemed to be 1281 inconclusive for a number of reasons. Proper instrumentation and 1282 treatment of inconclusive outcomes is critical to the accuracy and 1283 robustness of Model Based Metrics. Tests can be inconclusive if the 1284 precomputed traffic pattern or data rates were not accurately 1285 generated; the measurement results were not statistically 1286 significant; and others causes such as failing to meet some required 1287 preconditions for the test. See Section 5.4 1289 For example consider a test that implements Constant Window Pseudo 1290 CBR (Section 6.2) by adding rate controls and detailed IP packet 1291 transfer instrumentation to TCP (e.g. [RFC4898]). TCP includes 1292 built in control systems which might interfere with the sending data 1293 rate. If such a test meets the required packet transfer statistics 1294 (e.g. run length) while failing to attain the specified data rate it 1295 must be treated as an inconclusive result, because we can not a 1296 priori determine if the reduced data rate was caused by a TCP problem 1297 or a network problem, or if the reduced data rate had a material 1298 effect on the observed packet transfer statistics. 1300 Note that for capacity tests, if the observed packet transfer 1301 statistics meet the statistical criteria for failing (accepting 1302 hypnosis H1 in Section 7.2), the test can can be considered to have 1303 failed because it doesn't really matter that the test didn't attain 1304 the required data rate. 1306 The really important new properties of MBM, such as vantage 1307 independence, are a direct consequence of opening the control loops 1308 in the protocols, such that the test stream does not depend on 1309 network conditions or IP packets received. Any mechanism that 1310 introduces feedback between the path's measurements and the test 1311 stream generation is at risk of introducing non-linearities that 1312 spoil these properties. Any exceptional event that indicates that 1313 such feedback has happened should cause the test to be considered 1314 inconclusive. 1316 One way to view inconclusive tests is that they reflect situations 1317 where a test outcome is ambiguous between limitations of the network 1318 and some unknown limitation of the IP diagnostic test itself, which 1319 may have been caused by some uncontrolled feedback from the network. 1321 Note that procedures that attempt to search the target parameter 1322 space to find the limits on some parameter such as target_data_rate 1323 are at risk of breaking the location independent properties of Model 1324 Based Metrics, if any part of the boundary between passing and 1325 inconclusive or failing results is sensitive to RTT (which is 1326 normally the case). For example the maximum data rate for a 1327 margional link (e.g. exhibiting excess errors) is likely to be 1328 sensitive to the test path RTT. The maximum observed data rate over 1329 the test path has very little predictive value for the maximum rate 1330 over a different path. 1332 One of the goals for evolving TIDS designs will be to keep sharpening 1333 distinction between inconclusive, passing and failing tests. The 1334 criteria for for passing, failing and inconclusive tests MUST be 1335 explicitly stated for every test in the TIDS or FSTIDS. 1337 One of the goals of evolving the testing process, procedures, tools 1338 and measurement point selection should be to minimize the number of 1339 inconclusive tests. 1341 It may be useful to keep raw packet transfer statistics and ancillary 1342 metrics [RFC3148] for deeper study of the behavior of the network 1343 path and to measure the tools themselves. Raw packet transfer 1344 statistics can help to drive tool evolution. Under some conditions 1345 it might be possible to re-evaluate the raw data for satisfying 1346 alternate Target Transport Performance. However it is important to 1347 guard against sampling bias and other implicit feedback which can 1348 cause false results and exhibit measurement point vantage 1349 sensitivity. Simply applying different delivery criteria based on a 1350 different Target Transport Performance is insufficient if the test 1351 traffic patterns (bursts, etc.) does not match the alternate Target 1352 Transport Performance. 1354 7.2. Statistical criteria for estimating run_length 1356 When evaluating the observed run_length, we need to determine 1357 appropriate packet stream sizes and acceptable error levels for 1358 efficient measurement. In practice, can we compare the empirically 1359 estimated packet loss and ECN Congestion Experienced (CE) marking 1360 ratios with the targets as the sample size grows? How large a sample 1361 is needed to say that the measurements of packet transfer indicate a 1362 particular run length is present? 1364 The generalized measurement can be described as recursive testing: 1365 send packets (individually or in patterns) and observe the packet 1366 delivery performance (packet loss ratio or other metric, any marking 1367 we define). 1369 As each packet is sent and measured, we have an ongoing estimate of 1370 the performance in terms of the ratio of packet loss or ECN CE mark 1371 to total packets (i.e. an empirical probability). We continue to 1372 send until conditions support a conclusion or a maximum sending limit 1373 has been reached. 1375 We have a target_mark_probability, 1 mark per target_run_length, 1376 where a "mark" is defined as a lost packet, a packet with ECN CE 1377 mark, or other signal. This constitutes the null Hypothesis: 1379 H0: no more than one mark in target_run_length = 1380 3*(target_window_size)^2 packets 1382 and we can stop sending packets if on-going measurements support 1383 accepting H0 with the specified Type I error = alpha (= 0.05 for 1384 example). 1386 We also have an alternative Hypothesis to evaluate: if performance is 1387 significantly lower than the target_mark_probability. Based on 1388 analysis of typical values and practical limits on measurement 1389 duration, we choose four times the H0 probability: 1391 H1: one or more marks in (target_run_length/4) packets 1393 and we can stop sending packets if measurements support rejecting H0 1394 with the specified Type II error = beta (= 0.05 for example), thus 1395 preferring the alternate hypothesis H1. 1397 H0 and H1 constitute the Success and Failure outcomes described 1398 elsewhere in the memo, and while the ongoing measurements do not 1399 support either hypothesis the current status of measurements is 1400 inconclusive. 1402 The problem above is formulated to match the Sequential Probability 1403 Ratio Test (SPRT) [StatQC]. Note that as originally framed the 1404 events under consideration were all manufacturing defects. In 1405 networking, ECN CE marks and lost packets are not defects but 1406 signals, indicating that the transport protocol should slow down. 1408 The Sequential Probability Ratio Test also starts with a pair of 1409 hypothesis specified as above: 1411 H0: p0 = one defect in target_run_length 1412 H1: p1 = one defect in target_run_length/4 1414 As packets are sent and measurements collected, the tester evaluates 1415 the cumulative defect count against two boundaries representing H0 1416 Acceptance or Rejection (and acceptance of H1): 1418 Acceptance line: Xa = -h1 + s*n 1419 Rejection line: Xr = h2 + s*n 1421 where n increases linearly for each packet sent and 1422 h1 = { log((1-alpha)/beta) }/k 1423 h2 = { log((1-beta)/alpha) }/k 1424 k = log{ (p1(1-p0)) / (p0(1-p1)) } 1425 s = [ log{ (1-p0)/(1-p1) } ]/k 1427 for p0 and p1 as defined in the null and alternative Hypotheses 1428 statements above, and alpha and beta as the Type I and Type II 1429 errors. 1431 The SPRT specifies simple stopping rules: 1433 o Xa < defect_count(n) < Xb: continue testing 1434 o defect_count(n) <= Xa: Accept H0 1435 o defect_count(n) >= Xb: Accept H1 1437 The calculations above are implemented in the R-tool for Statistical 1438 Analysis [Rtool] , in the add-on package for Cross-Validation via 1439 Sequential Testing (CVST) [CVST]. 1441 Using the equations above, we can calculate the minimum number of 1442 packets (n) needed to accept H0 when x defects are observed. For 1443 example, when x = 0: 1445 Xa = 0 = -h1 + s*n 1446 and n = h1 / s 1448 7.3. Reordering Tolerance 1450 All tests MUST be instrumented for packet level reordering [RFC4737]. 1451 However, there is no consensus for how much reordering should be 1452 acceptable. Over the last two decades the general trend has been to 1453 make protocols and applications more tolerant to reordering (see for 1454 example [RFC4015]), in response to the gradual increase in reordering 1455 in the network. This increase has been due to the deployment of 1456 technologies such as multi threaded routing lookups and Equal Cost 1457 MultiPath (ECMP) routing. These techniques increase parallelism in 1458 network and are critical to enabling overall Internet growth to 1459 exceed Moore's Law. 1461 Note that transport retransmission strategies can trade off 1462 reordering tolerance vs how quickly they can repair losses vs 1463 overhead from spurious retransmissions. In advance of new 1464 retransmission strategies we propose the following strawman: 1465 Transport protocols should be able to adapt to reordering as long as 1466 the reordering extent is not more than the maximum of one quarter 1467 window or 1 mS, whichever is larger. Within this limit on reorder 1468 extent, there should be no bound on reordering density. 1470 By implication, recording which is less than these bounds should not 1471 be treated as a network impairment. However [RFC4737] still applies: 1472 reordering should be instrumented and the maximum reordering that can 1473 be properly characterized by the test (because of the bound on 1474 history buffers) should be recorded with the measurement results. 1476 Reordering tolerance and diagnostic limitations, such as the size of 1477 the history buffer used to diagnose packets that are way out-of- 1478 order, MUST be specified in a FSTIDS. 1480 8. IP Diagnostic Tests 1482 The IP diagnostic tests below are organized by traffic pattern: basic 1483 data rate and packet transfer statistics, standing queues, slowstart 1484 bursts, and sender rate bursts. We also introduce some combined 1485 tests which are more efficient when networks are expected to pass, 1486 but conflate diagnostic signatures when they fail. 1488 There are a number of test details which are not fully defined here. 1489 They must be fully specified in a FS-TIDS. From a standardization 1490 perspective, this lack of specificity will weaken this version of 1491 Model Based Metrics, however it is anticipated that this weakness is 1492 than offset by the extent to which MBM suppresses the problems caused 1493 by using transport protocols for measurement. e.g. non-specific MBM 1494 metrics are likely to have better repeatability than many existing 1495 BTC like metrics. Once we have good field experience, the missing 1496 details can be fully specified. 1498 8.1. Basic Data Rate and Packet Transfer Tests 1500 We propose several versions of the basic data rate and packet 1501 transfer statistics test. All measure the number of packets 1502 delivered between losses or ECN Congestion Experienced (CE) marks, 1503 using a data stream that is rate controlled at approximately the 1504 target_data_rate. 1506 The tests below differ in how the data rate is controlled. The data 1507 can be paced on a timer, or window controlled (and self clocked). 1508 The first two tests implicitly confirm that sub_path has sufficient 1509 raw capacity to carry the target_data_rate. They are recommended for 1510 relatively infrequent testing, such as an installation or periodic 1511 auditing process. The third, background packet transfer statistics, 1512 is a low rate test designed for ongoing monitoring for changes in 1513 subpath quality. 1515 All rely on the data receiver accumulating packet transfer statistics 1516 as described in Section 7.2 to score the outcome: 1518 Pass: it is statistically significant that the observed interval 1519 between losses or ECN CE marks is larger than the target_run_length. 1521 Fail: it is statistically significant that the observed interval 1522 between losses or ECN CE marks is smaller than the target_run_length. 1524 A test is considered to be inconclusive if it failed to generate the 1525 data rate as specified below, meet the qualifications defined in 1526 Section 5.4 or neither run length statistical hypothesis was 1527 confirmed in the allotted test duration. 1529 8.1.1. Delivery Statistics at Paced Full Data Rate 1531 Confirm that the observed run length is at least the 1532 target_run_length while relying on timer to send data at the 1533 target_rate using the procedure described in in Section 6.1 with a 1534 burst size of 1 (single packets) or 2 (packet pairs). 1536 The test is considered to be inconclusive if the packet transmission 1537 can not be accurately controlled for any reason. 1539 RFC 6673 [RFC6673] is appropriate for measuring packet transfer 1540 statistics at full data rate. 1542 8.1.2. Delivery Statistics at Full Data Windowed Rate 1544 Confirm that the observed run length is at least the 1545 target_run_length while sending at an average rate approximately 1546 equal to the target_data_rate, by controlling (or clamping) the 1547 window size of a conventional transport protocol to test_window. 1549 Since losses and ECN CE marks cause transport protocols to reduce 1550 their data rates, this test is expected to be less precise about 1551 controlling its data rate. It should not be considered inconclusive 1552 as long as at least some of the round trips reached the full 1553 target_data_rate without incurring losses or ECN CE marks. To pass 1554 this test the network MUST deliver target_window_size packets in 1555 target_RTT time without any losses or ECN CE marks at least once per 1556 two target_window_size round trips, in addition to meeting the run 1557 length statistical test. 1559 8.1.3. Background Packet Transfer Statistics Tests 1561 The background run length is a low rate version of the target target 1562 rate test above, designed for ongoing lightweight monitoring for 1563 changes in the observed subpath run length without disrupting users. 1564 It should be used in conjunction with one of the above full rate 1565 tests because it does not confirm that the subpath can support raw 1566 data rate. 1568 RFC 6673 [RFC6673] is appropriate for measuring background packet 1569 transfer statistics. 1571 8.2. Standing Queue Tests 1573 These engineering tests confirm that the bottleneck is well behaved 1574 across the onset of packet loss, which typically follows after the 1575 onset of queueing. Well behaved generally means lossless for 1576 transient queues, but once the queue has been sustained for a 1577 sufficient period of time (or reaches a sufficient queue depth) there 1578 should be a small number of losses or ECN CE marks to signal to the 1579 transport protocol that it should reduce its window. Losses that are 1580 too early can prevent the transport from averaging at the 1581 target_data_rate. Losses that are too late indicate that the queue 1582 might be subject to bufferbloat [wikiBloat] and inflict excess 1583 queuing delays on all flows sharing the bottleneck queue. Excess 1584 losses (more than half of the window) at the onset of congestion make 1585 loss recovery problematic for the transport protocol. Non-linear, 1586 erratic or excessive RTT increases suggest poor interactions between 1587 the channel acquisition algorithms and the transport self clock. All 1588 of the tests in this section use the same basic scanning algorithm, 1589 described here, but score the link or subpath on the basis of how 1590 well it avoids each of these problems. 1592 For some technologies the data might not be subject to increasing 1593 delays, in which case the data rate will vary with the window size 1594 all the way up to the onset of load induced packet loss or ECN CE 1595 marks. For theses technologies, the discussion of queueing does not 1596 apply, but it is still required that the onset of losses or ECN CE 1597 marks be at an appropriate point and progressive. Start the scan at 1598 a window equal to or slightly below the test_window. 1600 Use the procedure in Section 6.3 to sweep the window across the onset 1601 of queueing and the onset of loss. The tests below all assume that 1602 the scan emulates standard additive increase and delayed ACK by 1603 incrementing the window by one packet for every 2*target_window_size 1604 packets delivered. A scan can typically be divided into three 1605 regions: below the onset of queueing, a standing queue, and at or 1606 beyond the onset of loss. 1608 Below the onset of queueing the RTT is typically fairly constant, and 1609 the data rate varies in proportion to the window size. Once the data 1610 rate reaches the subpath IP rate, the data rate becomes fairly 1611 constant, and the RTT increases in proportion to the increase in 1612 window size. The precise transition across the start of queueing can 1613 be identified by the maximum network power, defined to be the ratio 1614 data rate over the RTT. The network power can be computed at each 1615 window size, and the window with the maximum is taken as the start of 1616 the queueing region. 1618 If there is random background loss (e.g. bit errors, etc), precise 1619 determination of the onset of queue induced packet loss may require 1620 multiple scans. Above the onset of queuing loss, all transport 1621 protocols are expected to experience periodic losses determined by 1622 the interaction between the congestion control and AQM algorithms. 1623 For standard congestion control algorithms the periodic losses are 1624 likely to be relatively widely spaced and the details are typically 1625 dominated by the behavior of the transport protocol itself. For the 1626 stiffened transport protocols case (with non-standard, aggressive 1627 congestion control algorithms) the details of periodic losses will be 1628 dominated by how the window increase function responds to loss. 1630 8.2.1. Congestion Avoidance 1632 A subpath passes the congestion avoidance standing queue test if more 1633 than target_run_length packets are delivered between the onset of 1634 queueing (as determined by the window with the maximum network power 1635 as described above) and the first loss or ECN CE mark. If this test 1636 is implemented using a standard congestion control algorithm with a 1637 clamp, it can be performed in situ in the production internet as a 1638 capacity test. For an example of such a test see [Pathdiag]. 1640 For technologies that do not have conventional queues, use the 1641 test_window in place of the onset of queueing. i.e. A subpath passes 1642 the congestion avoidance standing queue test if more than 1643 target_run_length packets are delivered between start of the scan at 1644 test_window and the first loss or ECN CE mark. 1646 8.2.2. Bufferbloat 1648 This test confirms that there is some mechanism to limit buffer 1649 occupancy (e.g. that prevents bufferbloat). Note that this is not 1650 strictly a requirement for single stream bulk transport capacity, 1651 however if there is no mechanism to limit buffer queue occupancy then 1652 a single stream with sufficient data to deliver is likely to cause 1653 the problems described in [RFC7567], and [wikiBloat]. This may cause 1654 only minor symptoms for the dominant flow, but has the potential to 1655 make the subpath unusable for other flows and applications. 1657 Pass if the onset of loss occurs before a standing queue has 1658 introduced more delay than than twice target_RTT, or other well 1659 defined and specified limit. Note that there is not yet a model for 1660 how much standing queue is acceptable. The factor of two chosen here 1661 reflects a rule of thumb. In conjunction with the previous test, 1662 this test implies that the first loss should occur at a queueing 1663 delay which is between one and two times the target_RTT. 1665 Specified RTT limits that are larger than twice the target_RTT must 1666 be fully justified in the FSTIDS. 1668 8.2.3. Non excessive loss 1670 This test confirms that the onset of loss is not excessive. Pass if 1671 losses are equal or less than the increase in the cross traffic plus 1672 the test stream window increase since the previous RTT. This could 1673 be restated as non-decreasing total throughput of the subpath at the 1674 onset of loss. (Note that when there is a transient drop in subpath 1675 throughput and there is not already a standing queue, a subpath that 1676 passes other queue tests in this document will have sufficient queue 1677 space to hold one full RTT worth of data). 1679 Note that token bucket policers will not pass this test, which is as 1680 intended. TCP often stumbles badly if more than a small fraction of 1681 the packets are dropped in one RTT. Many TCP implementations will 1682 require a timeout and slowstart to recover their self clock. Even if 1683 they can recover from the massive losses the sudden change in 1684 available capacity at the bottleneck waists serving and front path 1685 capacity until TCP can adapt to the new rate [Policing]. 1687 8.2.4. Duplex Self Interference 1689 This engineering test confirms a bound on the interactions between 1690 the forward data path and the ACK return path. 1692 Some historical half duplex technologies had the property that each 1693 direction held the channel until it completely drained its queue. 1694 When a self clocked transport protocol, such as TCP, has data and 1695 ACKs passing in opposite directions through such a link, the behavior 1696 often reverts to stop-and-wait. Each additional packet added to the 1697 window raises the observed RTT by two packet times, once as it passes 1698 through the data path, and once for the additional delay incurred by 1699 the ACK waiting on the return path. 1701 The duplex self interference test fails if the RTT rises by more than 1702 a fixed bound above the expected queueing time computed from trom the 1703 excess window divided by the subpath IP Capacity. This bound must be 1704 smaller than target_RTT/2 to avoid reverting to stop and wait 1705 behavior. (e.g. Data packets and ACKs both have to be released at 1706 least twice per RTT.) 1708 8.3. Slowstart tests 1710 These tests mimic slowstart: data is sent at twice the effective 1711 bottleneck rate to exercise the queue at the dominant bottleneck. 1713 8.3.1. Full Window slowstart test 1715 This is a capacity test to confirm that slowstart is not likely to 1716 exit prematurely. Send slowstart bursts that are target_window_size 1717 total packets. 1719 Accumulate packet transfer statistics as described in Section 7.2 to 1720 score the outcome. Pass if it is statistically significant that the 1721 observed number of good packets delivered between losses or ECN CE 1722 marks is larger than the target_run_length. Fail if it is 1723 statistically significant that the observed interval between losses 1724 or ECN CE marks is smaller than the target_run_length. 1726 It is deemed inconclusive if the elapsed time to send the data burst 1727 is not less than half of the time to receive the ACKs. (i.e. It is 1728 acceptable to send data too fast, but sending it slower than twice 1729 the actual bottleneck rate as indicated by the ACKs is deemed 1730 inconclusive). The headway for the slowstart bursts should be the 1731 target_RTT. 1733 Note that these are the same parameters as the Sender Full Window 1734 burst test, except the burst rate is at slowestart rate, rather than 1735 sender interface rate. 1737 8.3.2. Slowstart AQM test 1739 Do a continuous slowstart (send data continuously at twice the 1740 implied IP bottleneck capacity), until the first loss, stop, allow 1741 the network to drain and repeat, gathering statistics on how many 1742 packets were delivered before the loss, the pattern of losses, 1743 maximum observed RTT and window size. Justify the results. There is 1744 not currently sufficient theory justifying requiring any particular 1745 result, however design decisions that affect the outcome of this 1746 tests also affect how the network balances between long and short 1747 flows (the "mice vs elephants" problem). The queue sojourn time for 1748 the first packet delivered after the first loss should be at least 1749 one half of the target_RTT. 1751 This is an engineering test: It should be performed on a quiescent 1752 network or testbed, since cross traffic has the potential to change 1753 the results in ill defined ways. 1755 8.4. Sender Rate Burst tests 1757 These tests determine how well the network can deliver bursts sent at 1758 sender's interface rate. Note that this test most heavily exercises 1759 the front path, and is likely to include infrastructure may be out of 1760 scope for an access ISP, even though the bursts might be caused by 1761 ACK compression, thinning or channel arbitration in the access ISP. 1762 See Appendix B. 1764 Also, there are a several details that are not precisely defined. 1765 For starters there is not a standard server interface rate. 1 Gb/s 1766 and 10 Gb/s are common today, but higher rates will become cost 1767 effective and can be expected to be dominant some time in the future. 1769 Current standards permit TCP to send a full window bursts following 1770 an application pause. (Congestion Window Validation [RFC2861] 1771 [RFC7661], is not required, but even if was, it does not take effect 1772 until an application pause is longer than an RTO.) Since full window 1773 bursts are consistent with standard behavior, it is desirable that 1774 the network be able to deliver such bursts, otherwise application 1775 pauses will cause unwarranted losses. Note that the AIMD sawtooth 1776 requires a peak window that is twice target_window_size, so the worst 1777 case burst may be 2*target_window_size. 1779 It is also understood in the application and serving community that 1780 interface rate bursts have a cost to the network that has to be 1781 balanced against other costs in the servers themselves. For example 1782 TCP Segmentation Offload (TSO) reduces server CPU in exchange for 1783 larger network bursts, which increase the stress on network buffer 1784 memory. Some newer TCP implementations can pace traffic at scale 1785 [TSO_pacing][TSO_fq_pacing]. It remains to be determined if and how 1786 quickly these changes will be deployed. 1788 There is not yet theory to unify these costs or to provide a 1789 framework for trying to optimize global efficiency. We do not yet 1790 have a model for how much the network should tolerate server rate 1791 bursts. Some bursts must be tolerated by the network, but it is 1792 probably unreasonable to expect the network to be able to efficiently 1793 deliver all data as a series of bursts. 1795 For this reason, this is the only test for which we encourage 1796 derating. A TIDS could include a table of pairs of derating 1797 parameters: burst sizes and how much each burst size is permitted to 1798 reduce the run length, relative to to the target_run_length. 1800 8.5. Combined and Implicit Tests 1802 Combined tests efficiently confirm multiple network properties in a 1803 single test, possibly as a side effect of normal content delivery. 1804 They require less measurement traffic than other testing strategies 1805 at the cost of conflating diagnostic signatures when they fail. 1806 These are by far the most efficient for monitoring networks that are 1807 nominally expected to pass all tests. 1809 8.5.1. Sustained Bursts Test 1811 The sustained burst test implements a combined worst case version of 1812 all of the capacity tests above. It is simply: 1814 Send target_window_size bursts of packets at server interface rate 1815 with target_RTT burst headway (burst start to next burst start). 1816 Verify that the observed packet transfer statistics meets the 1817 target_run_length. 1819 Key observations: 1821 o The subpath under test is expected to go idle for some fraction of 1822 the time, determined by the difference between the time to drain 1823 the queue at the subpath IP capacity, and the target_RTT. If the 1824 queue does not drain completely it may be an indication that the 1825 the subpath has insufficient IP capacity or that there is some 1826 other problem with the test (e.g. inconclusive). 1827 o The burst sensitivity can be derated by sending smaller bursts 1828 more frequently. E.g. send target_window_size*derate packet 1829 bursts every target_RTT*derate, where "derate" is less than one. 1830 o When not derated, this test is the most strenuous capacity test. 1831 o A subpath that passes this test is likely to be able to sustain 1832 higher rates (close to subpath_IP_capacity) for paths with RTTs 1833 significantly smaller than the target_RTT. 1834 o This test can be implemented with instrumented TCP [RFC4898], 1835 using a specialized measurement application at one end [MBMSource] 1836 and a minimal service at the other end [RFC0863] [RFC0864]. 1837 o This test is efficient to implement, since it does not require 1838 per-packet timers, and can make use of TSO in modern NIC hardware. 1839 o If a subpath is known to pass the Standing Queue engineering tests 1840 (particularly that it has a progressive onset of loss at an 1841 appropriate queue depth), then the Sustained Burst Test is 1842 sufficient to assure that the subpath under test will not impair 1843 Bulk Transport Capacity at the target performance under all 1844 conditions. See Section 8.2 for a discussion of the standing 1845 queue tests. 1847 Note that this test is clearly independent of the subpath RTT, or 1848 other details of the measurement infrastructure, as long as the 1849 measurement infrastructure can accurately and reliably deliver the 1850 required bursts to the subpath under test. 1852 8.5.2. Streaming Media 1854 Model Based Metrics can be implicitly implemented as a side effect 1855 any non-throughput maximizing application, such as streaming media, 1856 with some additional controls and instrumentation in the servers. 1857 The essential requirement is that the data rate be constrained such 1858 that even with arbitrary application pauses and bursts, the data rate 1859 and burst sizes stay within the envelope defined by the individual 1860 tests described above. 1862 If the application's serving data rate can be constrained to be less 1863 than or equal to the target_data_rate and the serving_RTT (the RTT 1864 between the sender and client) is less than the target_RTT, this 1865 constraint is most easily implemented by clamping the transport 1866 window size to serving_window_clamp, set to the test_window, computed 1867 for the actual serving path. 1869 Under the above constraints the serving_window_clamp will limit the 1870 both the serving data rate and burst sizes to be no larger than the 1871 procedures in Section 8.1.2 and Section 8.4 or Section 8.5.1. Since 1872 the serving RTT is smaller than the target_RTT, the worst case bursts 1873 that might be generated under these conditions will be smaller than 1874 called for by Section 8.4 and the sender rate burst sizes are 1875 implicitly derated by the serving_window_clamp divided by the 1876 target_window_size at the very least. (Depending on the application 1877 behavior, the data might be significantly smoother than specified by 1878 any of the burst tests.) 1880 In an alternative implementation the data rate and bursts might be 1881 explicitly controlled by a programmable traffic shaper or pacing at 1882 the sender. This would provide better control over transmissions but 1883 is more complicated to implement, although the required technology is 1884 available[TSO_pacing][TSO_fq_pacing]. 1886 Note that these techniques can be applied to any content delivery 1887 that can be constrained to a reduced data rate in order to inhibit 1888 TCP equilibrium behavior. 1890 9. An Example 1892 In this section a we illustrate a TIDS designed to confirm that an 1893 access ISP can reliably deliver HD video from multiple content 1894 providers to all of their customers. With modern codecs, minimal HD 1895 video (720p) generally fits in 2.5 Mb/s. Due to their geographical 1896 size, network topology and modem characteristics the ISP determines 1897 that most content is within a 50 mS RTT of their users (This example 1898 RTT is a sufficient to cover the propagation delay to continental 1899 Europe or either US coast with low delay modems or somewhat smaller 1900 geographical regions if the modems require additional delay to 1901 implement advanced compression and error recovery). 1903 2.5 Mb/s over a 50 ms path 1905 +----------------------+-------+---------+ 1906 | End-to-End Parameter | value | units | 1907 +----------------------+-------+---------+ 1908 | target_rate | 2.5 | Mb/s | 1909 | target_RTT | 50 | ms | 1910 | target_MTU | 1500 | bytes | 1911 | header_overhead | 64 | bytes | 1912 | | | | 1913 | target_window_size | 11 | packets | 1914 | target_run_length | 363 | packets | 1915 +----------------------+-------+---------+ 1917 Table 1 1919 Table 1 shows the default TCP model with no derating, and as such is 1920 quite conservative. The simplest TIDS would be to use the sustained 1921 burst test, described in Section 8.5.1. Such a test would send 11 1922 packet bursts every 50mS, and confirming that there was no more than 1923 1 packet loss per 33 bursts (363 total packets in 1.650 seconds). 1925 Since this number represents is the entire end-to-end loss budget, 1926 independent subpath tests could be implemented by apportioning the 1927 packet loss ratio across subpaths. For example 50% of the losses 1928 might be allocated to the access or last mile link to the user, 40% 1929 to the interconnects with other ISPs and 1% to each internal hop 1930 (assuming no more than 10 internal hops). Then all of the subpaths 1931 can be tested independently, and the spatial composition of passing 1932 subpaths would be expected to be within the end-to-end loss budget. 1934 9.1. Observations about applicability 1936 Guidance on deploying and using MBM belong in a future document. 1937 However this example illustrates some the issues that may need to be 1938 considered. 1940 Note that a another ISP, with different geographical coverage, 1941 topology or modem technology may need to assume a different 1942 target_RTT, and as a consequence different target_window_size and 1943 target_run_length, even for the same target_data rate. One of the 1944 implications of this is that infrastructure shared by multiple ISPs, 1945 such as inter-exchange points (IXPs) and other interconnects may need 1946 to be evaluated on the basis of the most stringent target_window_size 1947 and target_run_length of any participating ISP. One way to do this 1948 might be to choose target parameters for evaluating such shared 1949 infrastructure on the basis of a hypothetical reference path that 1950 does not necessarily match any actual paths. 1952 Testing interconnects has generally been problematic: conventional 1953 performance tests run between measurement points adjacent to either 1954 side of the interconnect are not generally useful. Unconstrained TCP 1955 tests, such as iPerf [iPerf] are usually overly aggressive due to the 1956 small RTT (often less than 1 mS). With a short RTT these tools are 1957 likely to report inflated data rates because on a short RTT these 1958 tools can tolerate very high packet loss ratios and can push other 1959 cross traffic off of the network. As a consequence these 1960 measurements are useless for predicting actual user performance over 1961 longer paths, and may themselves be quite disruptive. Model Based 1962 Metrics solves this problem. The interconnect can be evaluated with 1963 the same TIDS as other subpaths. Continuing our example, if the 1964 interconnect is apportioned 40% of the losses, 11 packet bursts sent 1965 every 50mS should have fewer than one loss per 82 bursts (902 1966 packets). 1968 10. Validation 1970 Since some aspects of the models are likely to be too conservative, 1971 Section 5.2 permits alternate protocol models and Section 5.3 permits 1972 test parameter derating. If either of these techniques are used, we 1973 require demonstrations that such a TIDS can robustly detect subpaths 1974 that will prevent authentic applications using state-of-the-art 1975 protocol implementations from meeting the specified Target Transport 1976 Performance. This correctness criteria is potentially difficult to 1977 prove, because it implicitly requires validating a TIDS against all 1978 possible paths and subpaths. The procedures described here are still 1979 experimental. 1981 We suggest two approaches, both of which should be applied: first, 1982 publish a fully open description of the TIDS, including what 1983 assumptions were used and and how it was derived, such that the 1984 research community can evaluate the design decisions, test them and 1985 comment on their applicability; and second, demonstrate that an 1986 applications running over an infinitesimally passing testbed do meet 1987 the performance targets. 1989 An infinitesimally passing testbed resembles a epsilon-delta proof in 1990 calculus. Construct a test network such that all of the individual 1991 tests of the TIDS pass by only small (infinitesimal) margins, and 1992 demonstrate that a variety of authentic applications running over 1993 real TCP implementations (or other protocol as appropriate) meets the 1994 Target Transport Performance over such a network. The workloads 1995 should include multiple types of streaming media and transaction 1996 oriented short flows (e.g. synthetic web traffic). 1998 For example, for the HD streaming video TIDS described in Section 9, 1999 the IP capacity should be exactly the header overhead above 2.5 Mb/s, 2000 the per packet random background loss ratio should be 1/363, for a 2001 run length of 363 packets, the bottleneck queue should be 11 packets 2002 and the front path should have just enough buffering to withstand 11 2003 packet interface rate bursts. We want every one of the TIDS tests to 2004 fail if we slightly increase the relevant test parameter, so for 2005 example sending a 12 packet bursts should cause excess (possibly 2006 deterministic) packet drops at the dominant queue at the bottleneck. 2007 On this infinitesimally passing network it should be possible for a 2008 real application using a stock TCP implementation in the vendor's 2009 default configuration to attain 2.5 Mb/s over an 50 mS path. 2011 The most difficult part of setting up such a testbed is arranging for 2012 it to infinitesimally pass the individual tests. Two approaches are 2013 suggested: constraining the network devices not to use all available 2014 resources (e.g. by limiting available buffer space or data rate); and 2015 pre-loading subpaths with cross traffic. Note that is it important 2016 that a single environment be constructed which infinitesimally passes 2017 all tests at the same time, otherwise there is a chance that TCP can 2018 exploit extra latitude in some parameters (such as data rate) to 2019 partially compensate for constraints in other parameters (queue 2020 space, or vice-versa). 2022 To the extent that a TIDS is used to inform public dialog it should 2023 be fully publicly documented, including the details of the tests, 2024 what assumptions were used and how it was derived. All of the 2025 details of the validation experiment should also be published with 2026 sufficient detail for the experiments to be replicated by other 2027 researchers. All components should either be open source of fully 2028 described proprietary implementations that are available to the 2029 research community. 2031 11. Security Considerations 2033 Measurement is often used to inform business and policy decisions, 2034 and as a consequence is potentially subject to manipulation. Model 2035 Based Metrics are expected to be a huge step forward because 2036 equivalent measurements can be performed from multiple vantage 2037 points, such that performance claims can be independently validated 2038 by multiple parties. 2040 Much of the acrimony in the Net Neutrality debate is due by the 2041 historical lack of any effective vantage independent tools to 2042 characterize network performance. Traditional methods for measuring 2043 Bulk Transport Capacity are sensitive to RTT and as a consequence 2044 often yield very different results when run local to an ISP or 2045 interconnect and when run over a customer's complete path. Neither 2046 the ISP nor customer can repeat the others measurements, leading to 2047 high levels of distrust and acrimony. Model Based Metrics are 2048 expected to greatly improve this situation. 2050 This document only describes a framework for designing Fully 2051 Specified Targeted IP Diagnostic Suite. Each FS-TIDS MUST include 2052 its own security section. 2054 12. Acknowledgements 2056 Ganga Maguluri suggested the statistical test for measuring loss 2057 probability in the target run length. Alex Gilgur for helping with 2058 the statistics. 2060 Meredith Whittaker for improving the clarity of the communications. 2062 Ruediger Geib provided feedback which greatly improved the document. 2064 This work was inspired by Measurement Lab: open tools running on an 2065 open platform, using open tools to collect open data. See 2066 http://www.measurementlab.net/ 2068 13. IANA Considerations 2070 This document has no actions for IANA. 2072 14. References 2074 14.1. Normative References 2076 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 2077 Requirement Levels", BCP 14, RFC 2119, March 1997. 2079 14.2. Informative References 2081 [RFC0863] Postel, J., "Discard Protocol", STD 21, RFC 863, May 1983. 2083 [RFC0864] Postel, J., "Character Generator Protocol", STD 22, 2084 RFC 864, May 1983. 2086 [RFC2330] Paxson, V., Almes, G., Mahdavi, J., and M. Mathis, 2087 "Framework for IP Performance Metrics", RFC 2330, May 2088 1998. 2090 [RFC2861] Handley, M., Padhye, J., and S. Floyd, "TCP Congestion 2091 Window Validation", RFC 2861, June 2000. 2093 [RFC3148] Mathis, M. and M. Allman, "A Framework for Defining 2094 Empirical Bulk Transfer Capacity Metrics", RFC 3148, July 2095 2001. 2097 [RFC3465] Allman, M., "TCP Congestion Control with Appropriate Byte 2098 Counting (ABC)", RFC 3465, February 2003. 2100 [RFC4015] Ludwig, R. and A. Gurtov, "The Eifel Response Algorithm 2101 for TCP", RFC 4015, February 2005. 2103 [RFC4737] Morton, A., Ciavattone, L., Ramachandran, G., Shalunov, 2104 S., and J. Perser, "Packet Reordering Metrics", RFC 4737, 2105 November 2006. 2107 [RFC4898] Mathis, M., Heffner, J., and R. Raghunarayan, "TCP 2108 Extended Statistics MIB", RFC 4898, May 2007. 2110 [RFC5136] Chimento, P. and J. Ishac, "Defining Network Capacity", 2111 RFC 5136, February 2008. 2113 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion 2114 Control", RFC 5681, September 2009. 2116 [RFC5835] Morton, A. and S. Van den Berghe, "Framework for Metric 2117 Composition", RFC 5835, April 2010. 2119 [RFC6049] Morton, A. and E. Stephan, "Spatial Composition of 2120 Metrics", RFC 6049, January 2011. 2122 [RFC6576] Geib, R., Ed., Morton, A., Fardid, R., and A. Steinmitz, 2123 "IP Performance Metrics (IPPM) Standard Advancement 2124 Testing", BCP 176, RFC 6576, DOI 10.17487/RFC6576, March 2125 2012, . 2127 [RFC6673] Morton, A., "Round-Trip Packet Loss Metrics", RFC 6673, 2128 August 2012. 2130 [RFC6928] Chu, J., Dukkipati, N., Cheng, Y., and M. Mathis, 2131 "Increasing TCP's Initial Window", RFC 6928, 2132 DOI 10.17487/RFC6928, April 2013, 2133 . 2135 [RFC7312] Fabini, J. and A. Morton, "Advanced Stream and Sampling 2136 Framework for IP Performance Metrics (IPPM)", RFC 7312, 2137 August 2014. 2139 [RFC7398] Bagnulo, M., Burbridge, T., Crawford, S., Eardley, P., and 2140 A. Morton, "A Reference Path and Measurement Points for 2141 Large-Scale Measurement of Broadband Performance", 2142 RFC 7398, February 2015. 2144 [RFC7567] Baker, F., Ed. and G. Fairhurst, Ed., "IETF 2145 Recommendations Regarding Active Queue Management", 2146 BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015, 2147 . 2149 [RFC7661] Fairhurst, G., Sathiaseelan, A., and R. Secchi, "Updating 2150 TCP to Support Rate-Limited Traffic", RFC 7661, 2151 DOI 10.17487/RFC7661, October 2015, 2152 . 2154 [RFC7680] Almes, G., Kalidindi, S., Zekauskas, M., and A. Morton, 2155 Ed., "A One-Way Loss Metric for IP Performance Metrics 2156 (IPPM)", STD 82, RFC 7680, DOI 10.17487/RFC7680, January 2157 2016, . 2159 [MSMO97] Mathis, M., Semke, J., Mahdavi, J., and T. Ott, "The 2160 Macroscopic Behavior of the TCP Congestion Avoidance 2161 Algorithm", Computer Communications Review volume 27, 2162 number3, July 1997. 2164 [WPING] Mathis, M., "Windowed Ping: An IP Level Performance 2165 Diagnostic", INET 94, June 1994. 2167 [mpingSource] 2168 Fan, X., Mathis, M., and D. Hamon, "Git Repository for 2169 mping: An IP Level Performance Diagnostic", Sept 2013, 2170 . 2172 [MBMSource] 2173 Hamon, D., Stuart, S., and H. Chen, "Git Repository for 2174 Model Based Metrics", Sept 2013, . 2177 [Pathdiag] 2178 Mathis, M., Heffner, J., O'Neil, P., and P. Siemsen, 2179 "Pathdiag: Automated TCP Diagnosis", Passive and Active 2180 Measurement , June 2008. 2182 [iPerf] Wikipedia Contributors, , "iPerf", Wikipedia, The Free 2183 Encyclopedia , cited March 2015, 2184 . 2187 [StatQC] Montgomery, D., "Introduction to Statistical Quality 2188 Control - 2nd ed.", ISBN 0-471-51988-X, 1990. 2190 [Rtool] R Development Core Team, , "R: A language and environment 2191 for statistical computing. R Foundation for Statistical 2192 Computing, Vienna, Austria. ISBN 3-900051-07-0, URL 2193 http://www.R-project.org/", , 2011. 2195 [CVST] Krueger, T. and M. Braun, "R package: Fast Cross- 2196 Validation via Sequential Testing", version 0.1, 11 2012. 2198 [AFD] Pan, R., Breslau, L., Prabhakar, B., and S. Shenker, 2199 "Approximate fairness through differential dropping", 2200 SIGCOMM Comput. Commun. Rev. 33, 2, April 2003. 2202 [wikiBloat] 2203 Wikipedia, , "Bufferbloat", http://en.wikipedia.org/ 2204 w/ index.php?title=Bufferbloat&oldid=608805474, March 2205 2015. 2207 [CCscaling] 2208 Fernando, F., Doyle, J., and S. Steven, "Scalable laws for 2209 stable network congestion control", Proceedings of 2210 Conference on Decision and 2211 Control, http://www.ee.ucla.edu/~paganini, December 2001. 2213 [TSO_pacing] 2214 Corbet, J., "TSO sizing and the FQ scheduler", 2215 LWN.net https://lwn.net/Articles/564978/, Aug 2013. 2217 [TSO_fq_pacing] 2218 Dumazet, E. and Y. Chen, "TSO, fair queuing, pacing: 2219 three's a charm", Proceedings of IETF 88, TCPM WG 2220 https://www.ietf.org/proceedings/88/slides/slides-88-tcpm- 2221 9.pdf, Nov 2013. 2223 [Policing] 2224 Flach, T., Papageorge, P., Terzis, A., Pedrosa, L., Cheng, 2225 Y., Karim, T., Katz-Bassett, E., and R. Govindan, "An 2226 Internet-Wide Analysis of Traffic Policing", ACM SIGCOMM , 2227 August 2016. 2229 Appendix A. Model Derivations 2231 The reference target_run_length described in Section 5.2 is based on 2232 very conservative assumptions: that all window above 2233 target_window_size contributes to a standing queue that raises the 2234 RTT, and that classic Reno congestion control with delayed ACKs are 2235 in effect. In this section we provide two alternative calculations 2236 using different assumptions. 2238 It may seem out of place to allow such latitude in a measurement 2239 standard, but this section provides offsetting requirements. 2241 The estimates provided by these models make the most sense if network 2242 performance is viewed logarithmically. In the operational Internet, 2243 data rates span more than 8 orders of magnitude, RTT spans more than 2244 3 orders of magnitude, and packet loss ratio spans at least 8 orders 2245 of magnitude if not more. When viewed logarithmically (as in 2246 decibels), these correspond to 80 dB of dynamic range. On an 80 dB 2247 scale, a 3 dB error is less than 4% of the scale, even though it 2248 represents a factor of 2 in untransformed parameter. 2250 This document gives a lot of latitude for calculating 2251 target_run_length, however people designing a TIDS should consider 2252 the effect of their choices on the ongoing tussle about the relevance 2253 of "TCP friendliness" as an appropriate model for Internet capacity 2254 allocation. Choosing a target_run_length that is substantially 2255 smaller than the reference target_run_length specified in Section 5.2 2256 strengthens the argument that it may be appropriate to abandon "TCP 2257 friendliness" as the Internet fairness model. This gives developers 2258 incentive and permission to develop even more aggressive applications 2259 and protocols, for example by increasing the number of connections 2260 that they open concurrently. 2262 A.1. Queueless Reno 2264 In Section 5.2 models were derived based on the assumption that the 2265 subpath IP rate matches the target rate plus overhead, such that the 2266 excess window needed for the AIMD sawtooth causes a fluctuating queue 2267 at the bottleneck. 2269 An alternate situation would be a bottleneck where there is no 2270 significant queue and losses are caused by some mechanism that does 2271 not involve extra delay, for example by the use of a virtual queue as 2272 done in Approximate Fair Dropping [AFD]. A flow controlled by such a 2273 bottleneck would have a constant RTT and a data rate that fluctuates 2274 in a sawtooth due to AIMD congestion control. Assume the losses are 2275 being controlled to make the average data rate meet some goal which 2276 is equal or greater than the target_rate. The necessary run length 2277 to meet the target_rate can be computed as follows: 2279 For some value of Wmin, the window will sweep from Wmin packets to 2280 2*Wmin packets in 2*Wmin RTT (due to delayed ACK). Unlike the 2281 queueing case where Wmin = target_window_size, we want the average of 2282 Wmin and 2*Wmin to be the target_window_size, so the average data 2283 rate is the target rate. Thus we want Wmin = 2284 (2/3)*target_window_size. 2286 Between losses each sawtooth delivers (1/2)(Wmin+2*Wmin)(2Wmin) 2287 packets in 2*Wmin round trip times. 2289 Substituting these together we get: 2291 target_run_length = (4/3)(target_window_size^2) 2293 Note that this is 44% of the reference_run_length computed earlier. 2294 This makes sense because under the assumptions in Section 5.2 the 2295 AMID sawtooth caused a queue at the bottleneck, which raised the 2296 effective RTT by 50%. 2298 Appendix B. The effects of ACK scheduling 2300 For many network technologies simple queuing models don't apply: the 2301 network schedules, thins or otherwise alters the timing of ACKs and 2302 data, generally to raise the efficiency of the channel allocation 2303 algorithms when confronted with relatively widely spaced small ACKs. 2304 These efficiency strategies are ubiquitous for half duplex, wireless 2305 and broadcast media. 2307 Altering the ACK stream by holding or thinning ACKs typically has two 2308 consequences: it raises the implied bottleneck IP capacity, making 2309 the fine grained slowstart bursts either faster or larger and it 2310 raises the effective RTT by the average time that the ACKs and data 2311 are delayed. The first effect can be partially mitigated by re- 2312 clocking ACKs once they are beyond the bottleneck on the return path 2313 to the sender, however this further raises the effective RTT. 2315 The most extreme example of this sort of behavior would be a half 2316 duplex channel that is not released as long as the endpoint currently 2317 holding the channel has more traffic (data or ACKs) to send. Such 2318 environments cause self clocked protocols under full load to revert 2319 to extremely inefficient stop and wait behavior. The channel 2320 constrains the protocol to send an entire window of data as a single 2321 contiguous burst on the forward path, followed by the entire window 2322 of ACKs on the return path. 2324 If a particular return path contains a subpath or device that alters 2325 the timing of the ACK stream, then the entire front path from the 2326 sender up to the bottleneck must be tested at the burst parameters 2327 implied by the ACK scheduling algorithm. The most important 2328 parameter is the Implied Bottleneck IP Capacity, which is the average 2329 rate at which the ACKs advance snd.una. Note that thinning the ACK 2330 stream (relying on the cumulative nature of seg.ack to permit 2331 discarding some ACKs) causes most TCP implementation to send 2332 interface rate bursts to offset the longer times between ACKs in 2333 order to maintain the average data rate. 2335 Note that due to ubiquitous self clocking in Internet protocols, ill 2336 conceived channel allocation mechanisms are likely to increases the 2337 queuing stress on the front path because they cause larger full 2338 sender rate data bursts. 2340 Holding data or ACKs for channel allocation or other reasons (such as 2341 forward error correction) always raises the effective RTT relative to 2342 the minimum delay for the path. Therefore it may be necessary to 2343 replace target_RTT in the calculation in Section 5.2 by an 2344 effective_RTT, which includes the target_RTT plus a term to account 2345 for the extra delays introduced by these mechanisms. 2347 Appendix C. Version Control 2349 This section to be removed prior to publication. 2351 Formatted: Thu Apr 7 18:12:37 PDT 2016 2353 Authors' Addresses 2355 Matt Mathis 2356 Google, Inc 2357 1600 Amphitheater Parkway 2358 Mountain View, California 94043 2359 USA 2361 Email: mattmathis@google.com 2363 Al Morton 2364 AT&T Labs 2365 200 Laurel Avenue South 2366 Middletown, NJ 07748 2367 USA 2369 Phone: +1 732 420 1571 2370 Email: acmorton@att.com