idnits 2.17.1 draft-ietf-ippm-model-based-metrics-07.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 1371 has weird spacing: '... and n = h1...' -- The document date (Oct 19, 2015) is 3111 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: April 21, 2016 AT&T Labs 6 Oct 19, 2015 8 Model Based Metrics for Bulk Transport Capacity 9 draft-ietf-ippm-model-based-metrics-07.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 48 This Internet-Draft is submitted in full conformance with the 49 provisions of BCP 78 and BCP 79. 51 Internet-Drafts are working documents of the Internet Engineering 52 Task Force (IETF). Note that other groups may also distribute 53 working documents as Internet-Drafts. The list of current Internet- 54 Drafts is at http://datatracker.ietf.org/drafts/current/. 56 Internet-Drafts are draft documents valid for a maximum of six months 57 and may be updated, replaced, or obsoleted by other documents at any 58 time. It is inappropriate to use Internet-Drafts as reference 59 material or to cite them other than as "work in progress." 61 This Internet-Draft will expire on April 21, 2016. 63 Copyright Notice 65 Copyright (c) 2015 IETF Trust and the persons identified as the 66 document authors. All rights reserved. 68 This document is subject to BCP 78 and the IETF Trust's Legal 69 Provisions Relating to IETF Documents 70 (http://trustee.ietf.org/license-info) in effect on the date of 71 publication of this document. Please review these documents 72 carefully, as they describe your rights and restrictions with respect 73 to this document. Code Components extracted from this document must 74 include Simplified BSD License text as described in Section 4.e of 75 the Trust Legal Provisions and are provided without warranty as 76 described in the Simplified BSD License. 78 Table of Contents 80 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 5 81 1.1. Version Control . . . . . . . . . . . . . . . . . . . . . 6 82 2. Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 83 3. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 9 84 4. Background . . . . . . . . . . . . . . . . . . . . . . . . . . 16 85 4.1. TCP properties . . . . . . . . . . . . . . . . . . . . . . 17 86 4.2. Diagnostic Approach . . . . . . . . . . . . . . . . . . . 19 87 4.3. New requirements relative to RFC 2330 . . . . . . . . . . 19 88 5. Common Models and Parameters . . . . . . . . . . . . . . . . . 20 89 5.1. Target End-to-end parameters . . . . . . . . . . . . . . . 20 90 5.2. Common Model Calculations . . . . . . . . . . . . . . . . 21 91 5.3. Parameter Derating . . . . . . . . . . . . . . . . . . . . 22 92 5.4. Test Preconditions . . . . . . . . . . . . . . . . . . . . 22 93 6. Generating test streams . . . . . . . . . . . . . . . . . . . 23 94 6.1. Mimicking slowstart . . . . . . . . . . . . . . . . . . . 24 95 6.2. Constant window pseudo CBR . . . . . . . . . . . . . . . . 25 96 6.3. Scanned window pseudo CBR . . . . . . . . . . . . . . . . 25 97 6.4. Concurrent or channelized testing . . . . . . . . . . . . 26 98 7. Interpreting the Results . . . . . . . . . . . . . . . . . . . 27 99 7.1. Test outcomes . . . . . . . . . . . . . . . . . . . . . . 27 100 7.2. Statistical criteria for estimating run_length . . . . . . 29 101 7.3. Reordering Tolerance . . . . . . . . . . . . . . . . . . . 31 102 8. IP Diagnostic Tests . . . . . . . . . . . . . . . . . . . . . 31 103 8.1. Basic Data Rate and Packet Transfer Tests . . . . . . . . 32 104 8.1.1. Delivery Statistics at Paced Full Data Rate . . . . . 32 105 8.1.2. Delivery Statistics at Full Data Windowed Rate . . . . 33 106 8.1.3. Background Packet Transfer Statistics Tests . . . . . 33 107 8.2. Standing Queue Tests . . . . . . . . . . . . . . . . . . . 33 108 8.2.1. Congestion Avoidance . . . . . . . . . . . . . . . . . 35 109 8.2.2. Bufferbloat . . . . . . . . . . . . . . . . . . . . . 35 110 8.2.3. Non excessive loss . . . . . . . . . . . . . . . . . . 35 111 8.2.4. Duplex Self Interference . . . . . . . . . . . . . . . 36 112 8.3. Slowstart tests . . . . . . . . . . . . . . . . . . . . . 36 113 8.3.1. Full Window slowstart test . . . . . . . . . . . . . . 36 114 8.3.2. Slowstart AQM test . . . . . . . . . . . . . . . . . . 37 115 8.4. Sender Rate Burst tests . . . . . . . . . . . . . . . . . 37 116 8.5. Combined and Implicit Tests . . . . . . . . . . . . . . . 38 117 8.5.1. Sustained Bursts Test . . . . . . . . . . . . . . . . 38 118 8.5.2. Streaming Media . . . . . . . . . . . . . . . . . . . 39 119 9. An Example . . . . . . . . . . . . . . . . . . . . . . . . . . 40 120 10. Validation . . . . . . . . . . . . . . . . . . . . . . . . . . 41 121 11. Security Considerations . . . . . . . . . . . . . . . . . . . 42 122 12. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . 43 123 13. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 43 124 14. References . . . . . . . . . . . . . . . . . . . . . . . . . . 43 125 14.1. Normative References . . . . . . . . . . . . . . . . . . . 43 126 14.2. Informative References . . . . . . . . . . . . . . . . . . 44 127 Appendix A. Model Derivations . . . . . . . . . . . . . . . . . . 46 128 A.1. Queueless Reno . . . . . . . . . . . . . . . . . . . . . . 47 129 Appendix B. The effects of ACK scheduling . . . . . . . . . . . . 48 130 Appendix C. Version Control . . . . . . . . . . . . . . . . . . . 49 131 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . . 49 133 1. Introduction 135 Model Based Metrics (MBM) rely on mathematical models to specify a 136 Targeted Suite of IP Diagnostic tests, designed to assess whether 137 common transport protocols can be expected to meet a predetermined 138 Target Transport Performance over an Internet path. This note 139 describes the modeling framework to derive the test parameters for 140 accessing an Internet path's ability to support a predetermined Bulk 141 Transport Capacity. 143 Each test in the Targeted IP Diagnostic Suite (TIDS) measures some 144 aspect of IP packet transfer needed to meet the Target Transport 145 Performance. For Bulk Transport Capacity the TIDS includes IP 146 diagnostic tests to verify that there is: sufficient IP capacity 147 (data rate); sufficient queue space at bottlenecks to absorb and 148 deliver typical transport bursts; and that the background packet loss 149 ratio is low enough not to interfere with congestion control; and 150 other properties described below. Unlike typical IPPM metrics which 151 yield measures of network properties, Model Based Metrics nominally 152 yield pass/fail evaluations of the ability of standard transport 153 protocols to meet the specific performance objective over some 154 network path. 156 In most cases the IP diagnostic tests can be implemented by combining 157 existing IPPM metrics with additional controls for generating test 158 streams having a specified temporal structure (busts or standing 159 queues, etc) and statistical criteria for evaluating packet transfer. 160 The temporal structure of the test streams mimic transport protocol 161 behavior over the complete path, the statistical criteria models the 162 transport protocol's response to less than ideal IP packet transfer. 164 This note describes an alternative to the approach presented in "A 165 Framework for Defining Empirical Bulk Transfer Capacity Metrics" 166 [RFC3148]. In the future, other Model Based Metrics may cover other 167 applications and transports, such as VoIP over RTP. 169 The MBM approach, mapping Target Transport Performance to a Targeted 170 IP Diagnostic Suite (TIDS) of IP tests, solves some intrinsic 171 problems with using TCP or other throughput maximizing protocols for 172 measurement. In particular all throughput maximizing protocols (and 173 TCP congestion control in particular) cause some level of congestion 174 in order to detect when they have filled the network. This self 175 inflicted congestion obscures the network properties of interest and 176 introduces non-linear equilibrium behaviors that make any resulting 177 measurements useless as metrics because they have no predictive value 178 for conditions or paths different than that of the measurement 179 itself. These problems are discussed at length in Section 4. 181 A Targeted IP Diagnostic Suite does not have such difficulties. IP 182 diagnostics can be constructed such that they make strong statistical 183 statements about path properties that are independent of the 184 measurement details, such as vantage and choice of measurement 185 points. Model Based Metrics are designed to bridge the gap between 186 empirical IP measurements and expected TCP performance. 188 1.1. Version Control 190 RFC Editor: Please remove this entire subsection prior to 191 publication. 193 Please send comments about this draft to ippm@ietf.org. See 194 http://goo.gl/02tkD for more information including: interim drafts, 195 an up to date todo list and information on contributing. 197 Formatted: Mon Oct 19 15:59:51 PDT 2015 199 Changes since -06 draft: 200 o More language nits: 201 * "Targeted IP Diagnostic Suite (TIDS)" replaces "Targeted 202 Diagnostic Suite (TDS)". 203 * "implied bottleneck IP capacity" replaces "implied bottleneck 204 IP rate". 205 * Updated to ECN CE Marks. 206 * Added "specified temporal structure" 207 * "test stream" replaces "test traffic" 208 * "packet transfer" replaces "packet delivery" 209 * Reworked discussion of slowstart, bursts and pacing. 210 * RFC 7567 replaces RFC 2309. 212 Changes since -05 draft: 213 o Wordsmithing on sections overhauled in -05 draft. 214 o Reorganized the document: 215 * Relocated subsection "Preconditions". 216 * Relocated subsection "New Requirements relative to RFC 2330". 217 o Addressed nits and not so nits by Ruediger Geib. (Thanks!) 218 o Substantially tightened the entire definitions section. 219 o Many terminology changes, to better conform to other docs : 220 * IP rate and IP capacity (following RFC 5136) replaces various 221 forms of link data rate. 222 * subpath replaces link. 223 * target_window_size replaces target_pipe_size. 224 * implied bottleneck IP rate replaces effective bottleneck link 225 rate. 226 * Packet delivery statistics replaces delivery statistics. 228 Changes since -04 draft: 230 o The introduction was heavily overhauled: split into a separate 231 introduction and overview. 232 o The new shorter introduction: 233 * Is a problem statement; 234 * This document provides a framework; 235 * That it replaces TCP measurement by IP tests; 236 * That the results are pass/fail. 237 o Added a diagram of the framework to the overview 238 o and introduces all of the elements of the framework. 239 o Renumbered sections, reducing the depth of some section numbers. 240 o Updated definitions to better agree with other documents: 241 * Reordered section 2 242 * Bulk [data] performance -> Bulk Transport Capacity, everywhere 243 including the title. 244 * loss rate and loss probability -> packet loss ratio 245 * end-to-end path -> complete path 246 * [end-to-end][target] performance -> Target Transport 247 Performance 248 * load test -> capacity test 250 2. Overview 252 This document describes a modeling framework for deriving a Targeted 253 IP Diagnostic Suite from a predetermined Target Transport 254 Performance. It is not a complete specification, and relies on other 255 standards documents to define important details such as packet type-p 256 selection, sampling techniques, vantage selection, etc. We imagine 257 Fully Specified Targeted IP Diagnostic Suites (FSTIDS), that define 258 all of these details. We use Targeted IP Diagnostic Suite (TIDS) to 259 refer to the subset of such a specification that is in scope for this 260 document. This terminology is defined in Section 3. 262 Section 4 describes some key aspects of TCP behavior and what they 263 imply about the requirements for IP packet transfer. Most of the IP 264 diagnostic tests needed to confirm that the path meets these 265 properties can be built on existing IPPM metrics, with the addition 266 of statistical criteria for evaluating packet transfer and in a few 267 cases, new mechanisms to implement the required temporal structure. 268 (One group of tests, the standing queue tests described in 269 Section 8.2, don't correspond to existing IPPM metrics, but suitable 270 metrics can be patterned after existing tools.) 272 Figure 1 shows the MBM modeling and measurement framework. The 273 Target Transport Performance, at the top of the figure, is determined 274 by the needs of the user or application, outside the scope of this 275 document. For Bulk Transport Capacity, the main performance 276 parameter of interest is the Target Data Rate. However, since TCP's 277 ability to compensate for less than ideal network conditions is 278 fundamentally affected by the Round Trip Time (RTT) and the Maximum 279 Transmission Unit (MTU) of the complete path, these parameters must 280 also be specified in advance based on knowledge about the intended 281 application setting. They may reflect a specific application over 282 real path through the Internet or an idealized application and 283 hypothetical path representing a typical user community. Section 5 284 describes the common parameters and models derived from the Target 285 Transport Performance. 287 Target Transport Performance 288 (Target Data Rate, Target RTT and Target MTU) 289 | 290 ________V_________ 291 | mathematical | 292 | models | 293 | | 294 ------------------ 295 Traffic parameters | | Statistical criteria 296 | | 297 _______V____________V____Targeted_______ 298 | | * * * | Diagnostic Suite | 299 _____|_______V____________V________________ | 300 __|____________V____________V______________ | | 301 | IP diagnostic test | | | 302 | | | | | | 303 | _____________V__ __V____________ | | | 304 | | test | | Delivery | | | | 305 | | stream | | Evaluation | | | | 306 | | generation | | | | | | 307 | -------v-------- ------^-------- | | | 308 | | v test stream via ^ | | |-- 309 | | -->======================>-- | | | 310 | | subpath under test | |- 311 ----V----------------------------------V--- | 312 | | | | | | 313 V V V V V V 314 fail/inconclusive pass/fail/inconclusive 316 Overall Modeling Framework 318 Figure 1 320 The mathematical models are used to design traffic patterns that 321 mimic TCP or other transport protocol delivering bulk data and 322 operating at the Target Data Rate, MTU and RTT over a full range of 323 conditions, including flows that are bursty at multiple time scales. 324 The traffic patterns are generated based on the three target 325 parameters of complete path and independent of the properties of 326 individual subpaths using the techniques described in Section 6. As 327 much as possible the measurement traffic is generated 328 deterministically (precomputed) to minimize the extent to which test 329 methodology, measurement points, measurement vantage or path 330 partitioning affect the details of the measurement traffic. 332 Section 7 describes packet transfer statistics and methods test them 333 against the bounds provided by the mathematical models. Since these 334 statistics are typically the composition of subpaths of the complete 335 path [RFC6049] , in situ testing requires that the end-to-end 336 statistical bounds be apportioned as separate bounds for each 337 subpath. Subpaths that are expected to be bottlenecks may be 338 expected to contribute a larger fraction of the total packet loss. 339 In compensation, non-bottlenecked subpaths have to be constrained to 340 contribute less packet loss. The criteria for passing each test of a 341 TIDS is an apportioned share of the total bound determined by the 342 mathematical model from the Target Transport Performance. 344 Section 8 describes the suite of individual tests needed to verify 345 all of required IP delivery properties. A subpath passes if and only 346 if all of the individual IP diagnostic tests pass. Any subpath that 347 fails any test indicates that some users are likely fail to attain 348 their Target Transport Performance under some conditions. In 349 addition to passing or failing, a test can be deemed to be 350 inconclusive for a number of reasons including: the precomputed 351 traffic pattern was not accurately generated; the measurement results 352 were not statistically significant; and others such as failing to 353 meet some required test preconditions. If all tests pass, except 354 some are inconclusive then the entire suite is deemed to be 355 inconclusive. 357 In Section 9 we present an example TIDS that might be representative 358 of HD video, and illustrate how Model Based Metrics can be used to 359 address difficult measurement situations, such as confirming that 360 intercarrier exchanges have sufficient performance and capacity to 361 deliver HD video between ISPs. 363 Since there is some uncertainty in the modeling process, Section 10 364 describes a validation procedure to diagnose and minimize false 365 positive and false negative results. 367 3. Terminology 369 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 370 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 371 document are to be interpreted as described in [RFC2119]. 373 Note that terms containing underscores (rather than spaces) appear in 374 equations in the modeling sections. In some cases both forms are 375 used for aesthetic reasons, they do not have different meanings. 377 General Terminology: 379 Target: A general term for any parameter specified by or derived 380 from the user's application or transport performance requirements. 381 Target Transport Performance: Application or transport performance 382 goals for the complete path. For Bulk Transport Capacity defined 383 in this note the Target Transport Performance includes the Target 384 Data Rate, Target RTT and Target MTU as described below. 385 Target Data Rate: The specified application data rate required for 386 an application's proper operation. Conventional BTC metrics are 387 focused on the Target Data Rate, however these metrics had little 388 or no predictive value because they do not consider the effects of 389 the other two parameters of the Target Transport Performance, the 390 RTT and MTU of the complete paths. 391 Target RTT (Round Trip Time): The specified baseline (minimum) RTT 392 of the longest complete path over which the user expects to be 393 able meet the target performance. TCP and other transport 394 protocol's ability to compensate for path problems is generally 395 proportional to the number of round trips per second. The Target 396 RTT determines both key parameters of the traffic patterns (e.g. 397 burst sizes) and the thresholds on acceptable IP packet transfer 398 statistics. The Target RTT must be specified considering 399 appropriate packets sizes: MTU sized packets on the forward path, 400 ACK sized packets (typically header_overhead) on the return path. 401 Note that Target RTT is specified and not measured, it determines 402 the applicability of MBM measuremets for paths that are different 403 than the measured path. 404 Target MTU (Maximum Transmission Unit): The specified maximum MTU 405 supported by the complete path the over which the application 406 expects to meet the target performance. Assume 1500 Byte MTU 407 unless otherwise specified. If some subpath has a smaller MTU, 408 then it becomes the Target MTU for the complete path, and all 409 model calculations and subpath tests must use the same smaller 410 MTU. 411 Targeted IP Diagnostic Suite (TIDS): A set of IP diagnostic tests 412 designed to determine if an otherwise ideal complete path 413 containing the subpath under test can sustain flows at a specific 414 target_data_rate using target_MTU sized packets when the RTT of 415 the complete path is target_RTT. 416 Fully Specified Targeted IP Diagnostic Suite: A TIDS together with 417 additional specification such as "type-p", etc which are out of 418 scope for this document, but need to be drawn from other standards 419 documents. 421 Bulk Transport Capacity: Bulk Transport Capacity Metrics evaluate an 422 Internet path's ability to carry bulk data, such as large files, 423 streaming (non-real time) video, and under some conditions, web 424 images and other content. Prior efforts to define BTC metrics 425 have been based on [RFC3148], which predates our understanding of 426 TCP and the requirements described in Section 4 427 IP diagnostic tests: Measurements or diagnostics to determine if 428 packet transfer statistics meet some precomputed target. 429 traffic patterns: The temporal patterns or burstiness of traffic 430 generated by applications over transport protocols such as TCP. 431 There are several mechanisms that cause bursts at various time 432 scales as described in Section 4.1. Our goal here is to mimic the 433 range of common patterns (burst sizes and rates, etc), without 434 tying our applicability to specific applications, implementations 435 or technologies, which are sure to become stale. 436 packet transfer statistics: Raw, detailed or summary statistics 437 about packet transfer properties of the IP layer including packet 438 losses, ECN Congestion Experienced (CE) marks, reordering, or any 439 other properties that may be germane to transport performance. 440 packet loss ratio: As defined in [I-D.ietf-ippm-2680-bis]. 441 apportioned: To divide and allocate, for example budgeting packet 442 loss across multiple subpaths such that the losses will accumulate 443 to less than a specified end-to-end loss ratio. Apportioning 444 metrics is essentially the inverse of the process described in 445 [RFC5835]. 446 open loop: A control theory term used to describe a class of 447 techniques where systems that naturally exhibit circular 448 dependencies can be analyzed by suppressing some of the 449 dependencies, such that the resulting dependency graph is acyclic. 451 Terminology about paths, etc. See [RFC2330] and [RFC7398]. 453 data sender: Host sending data and receiving ACKs. 454 data receiver: Host receiving data and sending ACKs. 455 complete path: The end-to-end path from the data sender to the data 456 receiver. 457 subpath: A portion of the complete path. Note that there is no 458 requirement that subpaths be non-overlapping. A subpath can be a 459 small as a single device, link or interface. 460 measurement point: Measurement points as described in [RFC7398]. 461 test path: A path between two measurement points that includes a 462 subpath of the complete path under test. If the measurement 463 points are off path, the test path may include "test leads" 464 between the measurement points and the subpath. 466 dominant bottleneck: The bottleneck that generally determines most 467 of packet transfer statistics for the entire path. It typically 468 determines a flow's self clock timing, packet loss and ECN 469 Congestion Experienced (CE) marking rate, with other potential 470 bottlenecks having less effect on the packet transfer statistics. 471 See Section 4.1 on TCP properties. 472 front path: The subpath from the data sender to the dominant 473 bottleneck. 474 back path: The subpath from the dominant bottleneck to the receiver. 475 return path: The path taken by the ACKs from the data receiver to 476 the data sender. 477 cross traffic: Other, potentially interfering, traffic competing for 478 network resources (bandwidth and/or queue capacity). 480 Properties determined by the complete path and application. These 481 are described in more detail in Section 5.1. 483 Application Data Rate: General term for the data rate as seen by the 484 application above the transport layer in bytes per second. This 485 is the payload data rate, and explicitly excludes transport and 486 lower level headers (TCP/IP or other protocols), retransmissions 487 and other overhead that is not part to the total quantity of data 488 delivered to the application. 489 IP rate: The actual number of IP-layer bytes delivered through a 490 subpath, per unit time, including TCP and IP headers, retransmits 491 and other TCP/IP overhead. Follows from IP-type-P Link Usage 492 [RFC5136]. 493 IP capacity: The maximum number of IP-layer bytes that can be 494 transmitted through a subpath, per unit time, including TCP and IP 495 headers, retransmits and other TCP/IP overhead. Follows from IP- 496 type-P Link Capacity [RFC5136]. 497 bottleneck IP capacity: The IP capacity of the dominant bottleneck 498 in the forward path. All throughput maximizing protocols estimate 499 this capacity by observing the IP rate delivered through the 500 bottleneck. Most protocols derive their self clocks from the 501 timing of this data. See Section 4.1 and Appendix B for more 502 details. 503 implied bottleneck IP capacity: This is the bottleneck IP capacity 504 implied by the ACKs returning from the receiver. It is determined 505 by looking at how much application data the ACK stream at the 506 sender reports delivered to the data receiver per unit time at 507 various time scales. If the return path is thinning, batching or 508 otherwise altering the ACK timing the implied bottleneck IP 509 capacity over short time scales might be substantially larger than 510 the bottleneck IP capacity averaged over a full RTT. Since TCP 511 derives its clock from the data delivered through the bottleneck 512 the front path must have sufficient buffering to absorb any data 513 bursts at the dimensions (duration and IP rate) implied by the ACK 514 stream, potentially doubled during slowstart. If the return path 515 is not altering the ACK stream, then the implied bottleneck IP 516 capacity will be the same as the bottleneck IP capacity. See 517 Section 4.1 and Appendix B for more details. 518 sender interface rate: The IP rate which corresponds to the IP 519 capacity of the data sender's interface. Due to sender efficiency 520 algorithms including technologies such as TCP segmentation offload 521 (TSO), nearly all moderns servers deliver data in bursts at full 522 interface link rate. Today 1 or 10 Gb/s are typical. 523 Header_overhead: The IP and TCP header sizes, which are the portion 524 of each MTU not available for carrying application payload. 525 Without loss of generality this is assumed to be the size for 526 returning acknowledgements (ACKs). For TCP, the Maximum Segment 527 Size (MSS) is the Target MTU minus the header_overhead. 529 Basic parameters common to models and subpath tests are defined here 530 are described in more detail in Section 5.2. Note that these are 531 mixed between application transport performance (excludes headers) 532 and IP performance (which include TCP headers and retransmissions as 533 part of the IP payload). 535 Window [size]: The total quantity of data plus the data represented 536 by ACKs circulating in the network is referred to as the window. 537 See Section 4.1. Sometimes used with other qualifiers (congestion 538 window, cwnd or receiver window) to indicate which mechanism is 539 controlling the window. 540 pipe size: A general term for number of packets needed in flight 541 (the window size) to exactly fill some network path or subpath. 542 It corresponds to the window size which maximizes network power, 543 the observed data rate divided by the observed RTT. Often used 544 with additional qualifiers to specify which path, or under what 545 conditions, etc. 546 target_window_size: The average number of packets in flight (the 547 window size) needed to meet the Target Data Rate, for the 548 specified Target RTT, and MTU. It implies the scale of the bursts 549 that the network might experience. 550 run length: A general term for the observed, measured, or specified 551 number of packets that are (expected to be) delivered between 552 losses or ECN Congestion Experienced (CE) marks. Nominally one 553 over the sum of the loss and ECN CE marking probabilities, if 554 there are independently and identically distributed. 555 target_run_length: The target_run_length is an estimate of the 556 minimum number of non-congestion marked packets needed between 557 losses or ECN Congestion Experienced (CE) marks necessary to 558 attain the target_data_rate over a path with the specified 559 target_RTT and target_MTU, as computed by a mathematical model of 560 TCP congestion control. A reference calculation is shown in 561 Section 5.2 and alternatives in Appendix A 563 reference target_run_length: target_run_length computed precisely by 564 the method in Section 5.2. This is likely to be more slightly 565 conservative than required by modern TCP implementations. 567 Ancillary parameters used for some tests: 569 derating: Under some conditions the standard models are too 570 conservative. The modeling framework permits some latitude in 571 relaxing or "derating" some test parameters as described in 572 Section 5.3 in exchange for a more stringent TIDS validation 573 procedures, described in Section 10. 574 subpath_IP_capacity: The IP capacity of a specific subpath. 575 test path: A subpath of a complete path under test. 576 test_path_RTT: The RTT observed between two measurement points using 577 packet sizes that are consistent with the transport protocol. 578 Generally MTU sized packets of the forward path, header_overhead 579 sized packets on the return path. 580 test_path_pipe: The pipe size of a test path. Nominally the test 581 path RTT times the test path IP_capacity. 582 test_window: The window necessary to meet the target_rate over a 583 test path. Typically test_window=target_data_rate*test_path_RTT/ 584 (target_MTU - header_overhead). 586 The terminology below is used to define temporal patterns for test 587 stream. These patterns are designed to mimic TCP behavior, as 588 described in Section 4.1. 589 packet headway: Time interval between packets, specified from the 590 start of one to the start of the next. e.g. If packets are sent 591 with a 1 mS headway, there will be exactly 1000 packets per 592 second. 593 burst headway: Time interval between bursts, specified from the 594 start of the first packet one burst to the start of the first 595 packet of the next burst. e.g. If 4 packet bursts are sent with a 596 1 mS burst headway, there will be exactly 4000 packets per second. 597 paced single packets: Send individual packets at the specified rate 598 or packet headway. 599 paced bursts: Send bursts on a timer. Specify any 3 of: average 600 data rate, packet size, burst size (number of packets) and burst 601 headway (burst start to start). By default the bursts are assumed 602 full sender interface rate, such that the packet headway within 603 each burst is the minimum supported by the sender's interface. 604 Under some conditions it is useful to explicitly specify the 605 packet headway within each burst. 606 slowstart rate: Mimic TCP slowstart by sending 4 packet paced bursts 607 at an average data rate equal to twice the implied bottleneck IP 608 capacity (but not more than the sender interface rate). This is a 609 two level burst pattern described in more detail in Section 6.1. 610 If the implied bottleneck IP capacity is more than half of the 611 sender interface rate, slowstart rate becomes sender interface 612 rate. 613 slowstart burst: Mimic one round of TCP slowstart by sending a 614 specified number of packets packets in a two level burst pattern 615 that resembles slowstart. 616 repeated slowstart bursts: Repeat Slowstart bursts once per 617 target_RTT. For TCP each burst would be twice as large as the 618 prior burst, and the sequence would end at the first ECN CE mark 619 or lost packet. For measurement, all slowstart bursts would be 620 the same size (nominally target_window_size but other sizes might 621 be specified), and the ECN CE marks and lost packets are counted. 623 The tests described in this note can be grouped according to their 624 applicability. 626 Capacity tests: Capacity tests determine if a network subpath has 627 sufficient capacity to deliver the Target Transport Performance. 628 As long as the test stream is within the proper envelope for the 629 Target Transport Performance, the average packet losses or ECN 630 Congestion Experienced (CE) marks must be below the threshold 631 computed by the model. As such, capacity tests reflect parameters 632 that can transition from passing to failing as a consequence of 633 cross traffic, additional presented load or the actions of other 634 network users. By definition, capacity tests also consume 635 significant network resources (data capacity and/or queue buffer 636 space), and the test schedules must be balanced by their cost. 637 Monitoring tests: Monitoring tests are designed to capture the most 638 important aspects of a capacity test, but without presenting 639 excessive ongoing load themselves. As such they may miss some 640 details of the network's performance, but can serve as a useful 641 reduced-cost proxy for a capacity test, for example to support 642 ongoing monitoring. 643 Engineering tests: Engineering tests evaluate how network algorithms 644 (such as AQM and channel allocation) interact with TCP-style self 645 clocked protocols and adaptive congestion control based on packet 646 loss and ECN Congestion Experienced (CE) marks. These tests are 647 likely to have complicated interactions with cross traffic and 648 under some conditions can be inversely sensitive to load. For 649 example a test to verify that an AQM algorithm causes ECN CE marks 650 or packet drops early enough to limit queue occupancy may 651 experience a false pass result in the presence of cross traffic. 652 It is important that engineering tests be performed under a wide 653 range of conditions, including both in situ and bench testing, and 654 over a wide variety of load conditions. Ongoing monitoring is 655 less likely to be useful for engineering tests, although sparse in 656 situ testing might be appropriate. 658 4. Background 660 At the time the IPPM WG was chartered, sound Bulk Transport Capacity 661 measurement was known to be well beyond our capabilities. Even at 662 the time that Framework for IP Performance Metrics [RFC3148] was 663 written we knew that we didn't fully understand the problem. Now, by 664 hindsight we understand why BTC is such a hard problem: 665 o TCP is a control system with circular dependencies - everything 666 affects performance, including components that are explicitly not 667 part of the test. 668 o Congestion control is an equilibrium process, such that transport 669 protocols change the packet transfer statistics (raise the packet 670 loss ratio and/or RTT) to conform to their behavior. By design 671 TCP congestion control keeps raising the data rate until the 672 network gives some indication that it is full by dropping or ECN 673 CE marking packets. If TCP successfully fills the network (e.g. 674 the bottleneck is 100% utilized) the packet loss and ECN CE marks 675 are mostly determined by TCP and how hard TCP drives the network 676 at that rate and not by the network itself. 677 o TCP's ability to compensate for network flaws is directly 678 proportional to the number of roundtrips per second (i.e. 679 inversely proportional to the RTT). As a consequence a flawed 680 subpath may pass a short RTT local test even though it fails when 681 the subpath is extended by an effectively perfect network to some 682 larger RTT. 683 o TCP has an extreme form of the Heisenberg problem - Measurement 684 and cross traffic interact in unknown and ill defined ways. The 685 situation is actually worse than the traditional physics problem 686 where you can at least estimate bounds on the relative momentum of 687 the measurement and measured particles. For network measurement 688 you can not in general determine the relative "mass" of either the 689 test stream or the cross traffic, so you can not gauge the 690 relative magnitude of the uncertainty that might be introduced by 691 any interaction. 693 These properties are a consequence of the equilibrium behavior 694 intrinsic to how all throughput maximizing protocols interact with 695 the Internet. These protocols rely on control systems based on 696 estimated network parameters to regulate the quantity of data sent 697 into the network. The sent data in turn alters the network 698 properties observed by the estimators, such that there are circular 699 dependencies between every component and every property. Since some 700 of these properties are nonlinear, the entire system is nonlinear, 701 and any change anywhere causes difficult to predict changes in every 702 parameter. 704 Model Based Metrics overcome these problems by making the measurement 705 system open loop: the packet transfer statistics (akin to the network 706 estimators) do not affect the traffic or traffic patterns (bursts), 707 which computed on the basis of the Target Transport Performance. In 708 order for a network to pass, the resulting packet transfer statistics 709 and corresponding network estimators have to be such that they would 710 not cause the control systems slow the traffic below the Target Data 711 Rate. 713 4.1. TCP properties 715 TCP and SCTP are self clocked protocols that carry the vast majority 716 of all Internet data. Their dominant behavior is to have an 717 approximately fixed quantity of data and acknowledgements (ACKs) 718 circulating in the network. The data receiver reports arriving data 719 by returning ACKs to the data sender, the data sender typically 720 responds by sending exactly the same quantity of data back into the 721 network. The total quantity of data plus the data represented by 722 ACKs circulating in the network is referred to as the window. The 723 mandatory congestion control algorithms incrementally adjust the 724 window by sending slightly more or less data in response to each ACK. 725 The fundamentally important property of this system is that it is 726 self clocked: The data transmissions are a reflection of the ACKs 727 that were delivered by the network, the ACKs are a reflection of the 728 data arriving from the network. 730 A number of protocol features cause bursts of data, even in idealized 731 networks that can be modeled as simple queueing systems. 733 During slowstart the IP rate is doubled on each RTT by sending twice 734 as much data as was delivered to the receiver during the prior RTT. 735 Each returning ACK causes the sender to transmit twice the data the 736 ACK reported arriving at the receiver. For slowstart to be able to 737 fill a network the network must be able to tolerate slowstart bursts 738 up to the full pipe size inflated by the anticipated window reduction 739 on the first loss or ECN CE mark. For example, with classic Reno 740 congestion control, an optimal slowstart has to end with a burst that 741 is twice the bottleneck rate for one RTT in duration. This burst 742 causes a queue which is equal to the pipe size (i.e. the window is 743 twice the pipe size) so when the window is halved in response to the 744 first packet loss, the new window will be the pipe size. 746 Note that if the bottleneck IP rate is less that half of the capacity 747 of the front path (which is almost always the case), the slowstart 748 bursts will not by themselves cause significant queues anywhere else 749 along the front path; they primarily exercise the queue at the 750 dominant bottleneck. 752 Several common efficiency algorithms also cause bursts. The self 753 clock is typically applied to groups of packets: the receiver's 754 delayed ACK algorithm generally sends only one ACK per two data 755 segments. Furthermore the modern senders use TCP segmentation 756 offload (TSO) to reduce CPU overhead. The sender's software stack 757 builds supersized TCP segments that the TSO hardware splits into MTU 758 sized segments on the wire. The net effect of TSO, delayed ACK and 759 other efficiency algorithms is to send bursts of segments at full 760 sender interface rate. 762 Note that these efficiency algorithms are almost always in effect, 763 including during slowstart, such that slowstart typically has a two 764 level burst structure. Section 6.1 describes slowstart in more 765 detail. 767 Additional sources of bursts include TCP's initial window [RFC6928], 768 application pauses, channel allocation mechanisms and network devices 769 that schedule ACKs. Appendix B describes these last two items. If 770 the application pauses (stops reading or writing data) for some 771 fraction of an RTT, many TCP implementations catch up to their 772 earlier window size by sending a burst of data at the full sender 773 interface rate. To fill a network with a realistic application, the 774 network has to be able to tolerate sender interface rate bursts large 775 enough to restore the prior window following application pauses. 777 Although the sender interface rate bursts are typically smaller than 778 the last burst of a slowstart, they are at a higher IP rate so they 779 potentially exercise queues at arbitrary points along the front path 780 from the data sender up to and including the queue at the dominant 781 bottleneck. There is no model for how frequent or what sizes of 782 sender rate bursts the network should tolerate. 784 In conclusion, to verify that a path can meet a Target Transport 785 Performance, it is necessary to independently confirm that the path 786 can tolerate bursts at the scales that can be caused by these 787 mechanisms. Three cases are believed to be sufficient: 789 o Two level slowstart bursts sufficient to get connections started 790 properly. 791 o Ubiquitous sender interface rate bursts caused by efficiency 792 algorithms. We assume 4 packet bursts to be the most common case, 793 since it matches the effects of delayed ACK during slowstart. 794 These bursts should be assumed not to significantly affect packet 795 transfer statistics. 796 o Infrequent sender interface rate bursts that are full 797 target_window_size. Target_run_length may be derated for these 798 large fast bursts. 800 If a subpath can meet the required packet loss ratio for bursts at 801 all of these scales then it has sufficient buffering at all potential 802 bottlenecks to tolerate any of the bursts that are likely introduced 803 by TCP or other transport protocols. 805 4.2. Diagnostic Approach 807 A complete path of a given RTT and MTU, which are equal to or smaller 808 than the Target RTT and equal to or larger than the Target MTU 809 respectively, is expected to be able to attain a specified Bulk 810 Transport Capacity when all of the following conditions are met: 811 1. The IP capacity is above the Target Data Rate by sufficient 812 margin to cover all TCP/IP overheads. This can be confirmed by 813 the tests described in Section 8.1 or any number of IP capacity 814 tests adapted to implement MBM. 815 2. The observed packet transfer statistics are better than required 816 by a suitable TCP performance model (e.g. fewer packet losses or 817 ECN CE marks). See Section 8.1 or any number of low rate packet 818 loss tests outside of MBM. 819 3. There is sufficient buffering at the dominant bottleneck to 820 absorb a slowstart bursts large enough to get the flow out of 821 slowstart at a suitable window size. See Section 8.3. 822 4. There is sufficient buffering in the front path to absorb and 823 smooth sender interface rate bursts at all scales that are likely 824 to be generated by the application, any channel arbitration in 825 the ACK path or any other mechanisms. See Section 8.4. 826 5. When there is a slowly rising standing queue at the bottleneck 827 the onset of packet loss has to be at an appropriate point (time 828 or queue depth) and progressive [RFC7567]. See Section 8.2. 829 6. When there is a standing queue at a bottleneck for a shared media 830 subpath (e.g. half duplex), there must be a suitable bounds on 831 the interaction between ACKs and data, for example due to the 832 channel arbitration mechanism. See Section 8.2.4. 834 Note that conditions 1 through 4 require capacity tests for 835 validation, and thus may need to be monitored on an ongoing basis. 836 Conditions 5 and 6 require engineering tests, which are best 837 performed in controlled environments such as a bench test. They 838 won't generally fail due to load, but may fail in the field due to 839 configuration errors, etc. and should be spot checked. 841 We are developing a tool that can perform many of the tests described 842 here [MBMSource]. 844 4.3. New requirements relative to RFC 2330 846 Model Based Metrics are designed to fulfill some additional 847 requirements that were not recognized at the time RFC 2330 was 848 written [RFC2330]. These missing requirements may have significantly 849 contributed to policy difficulties in the IP measurement space. Some 850 additional requirements are: 851 o IP metrics must be actionable by the ISP - they have to be 852 interpreted in terms of behaviors or properties at the IP or lower 853 layers, that an ISP can test, repair and verify. 854 o Metrics should be spatially composable, such that measures of 855 concatenated paths should be predictable from subpaths. 856 o Metrics must be vantage point invariant over a significant range 857 of measurement point choices, including off path measurement 858 points. The only requirements on MP selection should be that the 859 RTT between the MPs is below some reasonable bound, and that the 860 effects of the "test leads" connecting MPs to the subpath under 861 test can be can be calibrated out of the measurements. The latter 862 might be be accomplished if the test leads are effectively ideal 863 or their properties can be deducted from the measurements between 864 the MPs. While many of tests require that the test leads have at 865 least as much IP capacity as the subpath under test, some do not, 866 for example Background Packet Transfer Tests described in 867 Section 8.1.3. 868 o Metric measurements must be repeatable by multiple parties with no 869 specialized access to MPs or diagnostic infrastructure. It must 870 be possible for different parties to make the same measurement and 871 observe the same results. In particular it is specifically 872 important that both a consumer (or their delegate) and ISP be able 873 to perform the same measurement and get the same result. Note 874 that vantage independence is key to meeting this requirement. 876 5. Common Models and Parameters 878 5.1. Target End-to-end parameters 880 The target end-to-end parameters are the Target Data Rate, Target RTT 881 and Target MTU as defined in Section 3. These parameters are 882 determined by the needs of the application or the ultimate end user 883 and the complete Internet path over which the application is expected 884 to operate. The target parameters are in units that make sense to 885 upper layers: payload bytes delivered to the application, above TCP. 886 They exclude overheads associated with TCP and IP headers, 887 retransmits and other protocols (e.g. DNS). 889 Other end-to-end parameters defined in Section 3 include the 890 effective bottleneck data rate, the sender interface data rate and 891 the TCP and IP header sizes. 893 The target_data_rate must be smaller than all subpath IP capacities 894 by enough headroom to carry the transport protocol overhead, 895 explicitly including retransmissions and an allowance for 896 fluctuations in TCP's actual data rate. Specifying a 897 target_data_rate with insufficient headroom is likely to result in 898 brittle measurements having little predictive value. 900 Note that the target parameters can be specified for a hypothetical 901 path, for example to construct TIDS designed for bench testing in the 902 absence of a real application; or for a live in situ test of 903 production infrastructure. 905 The number of concurrent connections is explicitly not a parameter to 906 this model. If a subpath requires multiple connections in order to 907 meet the specified performance, that must be stated explicitly and 908 the procedure described in Section 6.4 applies. 910 5.2. Common Model Calculations 912 The Target Transport Performance is used to derive the 913 target_window_size and the reference target_run_length. 915 The target_window_size, is the average window size in packets needed 916 to meet the target_rate, for the specified target_RTT and target_MTU. 917 It is given by: 919 target_window_size = ceiling( target_rate * target_RTT / ( target_MTU 920 - header_overhead ) ) 922 Target_run_length is an estimate of the minimum required number of 923 unmarked packets that must be delivered between losses or ECN 924 Congestion Experienced (CE) marks, as computed by a mathematical 925 model of TCP congestion control. The derivation here follows 926 [MSMO97], and by design is quite conservative. 928 Reference target_run_length is derived as follows: assume the 929 subpath_IP_capacity is infinitesimally larger than the 930 target_data_rate plus the required header_overhead. Then 931 target_window_size also predicts the onset of queueing. A larger 932 window will cause a standing queue at the bottleneck. 934 Assume the transport protocol is using standard Reno style Additive 935 Increase, Multiplicative Decrease (AIMD) congestion control [RFC5681] 936 (but not Appropriate Byte Counting [RFC3465]) and the receiver is 937 using standard delayed ACKs. Reno increases the window by one packet 938 every pipe_size worth of ACKs. With delayed ACKs this takes 2 Round 939 Trip Times per increase. To exactly fill the pipe, losses must be no 940 closer than when the peak of the AIMD sawtooth reached exactly twice 941 the target_window_size otherwise the multiplicative window reduction 942 triggered by the loss would cause the network to be underfilled. 943 Following [MSMO97] the number of packets between losses must be the 944 area under the AIMD sawtooth. They must be no more frequent than 945 every 1 in ((3/2)*target_window_size)*(2*target_window_size) packets, 946 which simplifies to: 948 target_run_length = 3*(target_window_size^2) 950 Note that this calculation is very conservative and is based on a 951 number of assumptions that may not apply. Appendix A discusses these 952 assumptions and provides some alternative models. If a different 953 model is used, a fully specified TIDS or FSTIDS MUST document the 954 actual method for computing target_run_length and ratio between 955 alternate target_run_length and the reference target_run_length 956 calculated above, along with a discussion of the rationale for the 957 underlying assumptions. 959 These two parameters, target_window_size and target_run_length, 960 directly imply most of the individual parameters for the tests in 961 Section 8. 963 5.3. Parameter Derating 965 Since some aspects of the models are very conservative, the MBM 966 framework permits some latitude in derating test parameters. Rather 967 than trying to formalize more complicated models we permit some test 968 parameters to be relaxed as long as they meet some additional 969 procedural constraints: 970 o The TIDS or FSTIDS MUST document and justify the actual method 971 used to compute the derated metric parameters. 972 o The validation procedures described in Section 10 must be used to 973 demonstrate the feasibility of meeting the Target Transport 974 Performance with infrastructure that infinitesimally passes the 975 derated tests. 976 o The validation process for a FSTIDS itself must be documented is 977 such a way that other researchers can duplicate the validation 978 experiments. 980 Except as noted, all tests below assume no derating. Tests where 981 there is not currently a well established model for the required 982 parameters explicitly include derating as a way to indicate 983 flexibility in the parameters. 985 5.4. Test Preconditions 987 Many tests have preconditions which are required to assure their 988 validity. Examples include: the presence or nonpresence of cross 989 traffic on specific subpaths; negotiating ECN; and appropriate 990 preloading to put reactive network elements into the proper states 991 [RFC7312]. If preconditions are not properly satisfied for some 992 reason, the tests should be considered to be inconclusive. In 993 general it is useful to preserve diagnostic information as to why the 994 preconditions were not met, and any test data that was collected even 995 if it is not useful for the intended test. Such diagnostic 996 information and partial test data may be useful for improving the 997 test in the future. 999 It is important to preserve the record that a test was scheduled, 1000 because otherwise precondition enforcement mechanisms can introduce 1001 sampling bias. For example, canceling tests due to cross traffic on 1002 subscriber access links might introduce sampling bias in tests of the 1003 rest of the network by reducing the number of tests during peak 1004 network load. 1006 Test preconditions and failure actions MUST be specified in a FSTIDS. 1008 6. Generating test streams 1010 Many important properties of Model Based Metrics, such as vantage 1011 independence, are a consequence of using test streams that have 1012 temporal structures that mimic TCP or other transport protocols 1013 running over a complete path. As described in Section 4.1, self 1014 clocked protocols naturally have burst structures related to the RTT 1015 and pipe size of the complete path. These bursts naturally get 1016 larger (contain more packets) as either the Target RTT or Target Data 1017 Rate get larger, or the Target MTU gets smaller. An implication of 1018 these relationships is that test streams generated by running self 1019 clocked protocols over short subpaths may not adequately exercise the 1020 queueing at any bottleneck to determine if the subpath can support 1021 the full Target Transport Performance over the complete path. 1023 Failing to authentically mimic TCP's temporal structure is part the 1024 reason why simple performance tools such as iperf, netperf, nc, etc 1025 have the reputation of yielding false pass results over short test 1026 paths, even when some subpath has a flaw. 1028 The definitions in Section 3 are sufficient for most test streams. 1029 We describe the slowstart and standing queue test streams in more 1030 detail. 1032 In conventional measurement practice stochastic processes are used to 1033 eliminate many unintended correlations and sample biases. However 1034 MBM tests are designed to explicitly mimic temporal correlations 1035 caused by network or protocol elements themselves and are intended to 1036 accurately reflect implementation behavior. Some portion of the 1037 system, such as traffic arrival (test scheduling) are naturally 1038 stochastic. Other details, such as protocol processing times, are 1039 technically nondeterministic and might be modeled stochastically, but 1040 are only a tiny part of the overall behavior which is dominated by 1041 implementation specific deterministic effects. Furthermore, it is 1042 known that sampling bias is a real problem for some protocol 1043 implementations. For example TCP's RTT estimator used to determine 1044 the Retransmit Time Out (RTO), can be fooled by periodic cross 1045 traffic or start-stop applications. 1047 At some point in the future it may make sense to introduce fine 1048 grained noise sources into the models used for generating test 1049 streams, but they are not warranted at this time. 1051 6.1. Mimicking slowstart 1053 TCP slowstart has a two level burst structure as shown in Figure 2. 1054 The fine structure is caused by the interaction between the ACK clock 1055 and TCP efficiency algorithms. Each ACK passing through the return 1056 path triggers a small data burst. These bursts are typically full 1057 sender interface rate, with the same headway as the returning ACKs, 1058 but having twice as much data as the ACK reported was delivered to 1059 the receiver. Due to variations in delayed ACK and algorithms such 1060 as Appropriate Byte Counting [RFC3465], different pairs of senders 1061 and receivers produce different burst patterns. Without loss of 1062 generality, we assume each ACK causes 4 packet bursts at an average 1063 headway equal to the ACK headway, and corresponding to sending at an 1064 average rate equal to twice the effective bottleneck IP rate. This 1065 fine structure defines one slowstart rate burst. 1067 For a transport protocol the slowstart bursts are repeated every 1068 target_RTT. Each slowstart burst is twice as large as the previous 1069 burst, and slowstart ends on the first lost packet or ECN mark. For 1070 diagnostic tests described below we preserve the fine structure but 1071 manipulate the burst size and headway to measure the ability of the 1072 dominant bottleneck to absorb and smooth slowstart bursts. 1074 Note that a stream of repeated slowstart bursts has three different 1075 average rates, depending on the averaging interval. At the finest 1076 time scale (a few packet times at the sender interface) the peak of 1077 the average IP rate is the same as the sender interface rate; at a 1078 medium timescale (a few packet times at the dominant bottleneck) the 1079 peak of the average IP rate is twice the implied bottleneck IP 1080 capacity; and at time scales longer than the target_RTT and when the 1081 burst size is equal to the target_window_size the average rate is 1082 equal to the target_data_rate. This pattern corresponds to repeating 1083 the last RTT of TCP slowstart when delayed ACK and sender side byte 1084 counting are present but without the limits specified in Appropriate 1085 Byte Counting [RFC3465]. 1087 time --> ( - = one packet) 1088 Packet stream: 1090 ---- ---- ---- ---- ---- ---- ---- ... 1092 |<>| sender interface rate bursts (typically 3 or 4 packets) 1093 |<--->| burst headway (determined by ACK headway) 1094 |<------------------------>| slowstart burst size(from the window) 1095 |<---------------------------------------------->| slowstart headway 1096 \____________ _____________/ \______ __ ... 1097 V V 1098 One slowstart burst Repeated slowstart bursts 1100 Multiple levels of Slowstart Bursts 1102 Figure 2 1104 6.2. Constant window pseudo CBR 1106 Implement pseudo constant bit rate by running a standard protocol 1107 such as TCP with a fixed window size, such that it is self clocked. 1108 Data packets arriving at the receiver trigger acknowledgements (ACKs) 1109 which travel back to the sender where they trigger additional 1110 transmissions. The window size is computed from the target_data_rate 1111 and the actual RTT of the test path. The rate is only maintained in 1112 average over each RTT, and is subject to limitations of the transport 1113 protocol. 1115 Since the window size is constrained to be an integer number of 1116 packets, for small RTTs or low data rates there may not be 1117 sufficiently precise control over the data rate. Rounding the window 1118 size up (the default) is likely to be result in data rates that are 1119 higher than the target rate, but reducing the window by one packet 1120 may result in data rates that are too small. Also cross traffic 1121 potentially raises the RTT, implicitly reducing the rate. Cross 1122 traffic that raises the RTT nearly always makes the test more 1123 strenuous. A FSTIDS specifying a constant window CBR tests MUST 1124 explicitly indicate under what conditions errors in the data rate 1125 causes tests to inconclusive. 1127 Since constant window pseudo CBR testing is sensitive to RTT 1128 fluctuations it is less accurate at controling the data rate in 1129 environments with fluctuating delays. 1131 6.3. Scanned window pseudo CBR 1133 Scanned window pseudo CBR is similar to the constant window CBR 1134 described above, except the window is scanned across a range of sizes 1135 designed to include two key events, the onset of queueing and the 1136 onset of packet loss or ECN CE marks. The window is scanned by 1137 incrementing it by one packet every 2*target_window_size delivered 1138 packets. This mimics the additive increase phase of standard Reno 1139 TCP congestion avoidance when delayed ACKs are in effect. Normally 1140 the window increases separated by intervals slightly longer than 1141 twice the target_RTT. 1143 There are two ways to implement this test: one built by applying a 1144 window clamp to standard congestion control in a standard protocol 1145 such as TCP and the other built by stiffening a non-standard 1146 transport protocol. When standard congestion control is in effect, 1147 any losses or ECN CE marks cause the transport to revert to a window 1148 smaller than the clamp such that the scanning clamp loses control the 1149 window size. The NPAD pathdiag tool is an example of this class of 1150 algorithms [Pathdiag]. 1152 Alternatively a non-standard congestion control algorithm can respond 1153 to losses by transmitting extra data, such that it maintains the 1154 specified window size independent of losses or ECN CE marks. Such a 1155 stiffened transport explicitly violates mandatory Internet congestion 1156 control [RFC5681] and is not suitable for in situ testing. It is 1157 only appropriate for engineering testing under laboratory conditions. 1158 The Windowed Ping tool implements such a test [WPING]. The tool 1159 described in the paper has been updated.[mpingSource] 1161 The test procedures in Section 8.2 describe how to the partition the 1162 scans into regions and how to interpret the results. 1164 6.4. Concurrent or channelized testing 1166 The procedures described in this document are only directly 1167 applicable to single stream measurement, e.g. one TCP connection or 1168 measurement stream. In an ideal world, we would disallow all 1169 performance claims based multiple concurrent streams, but this is not 1170 practical due to at least two different issues. First, many very 1171 high rate link technologies are channelized and at last partially pin 1172 the flow to channel mapping to minimize packet reordering within 1173 flows. Second, TCP itself has scaling limits. Although the former 1174 problem might be overcome through different design decisions, the 1175 later problem is more deeply rooted. 1177 All congestion control algorithms that are philosophically aligned 1178 with the standard [RFC5681] (e.g. claim some level of TCP 1179 compatibility, friendliness or fairness) have scaling limits, in the 1180 sense that as a long fast network (LFN) with a fixed RTT and MTU gets 1181 faster, these congestion control algorithms get less accurate and as 1182 a consequence have difficulty filling the network[CCscaling]. These 1183 properties are a consequence of the original Reno AIMD congestion 1184 control design and the requirement in [RFC5681] that all transport 1185 protocols have similar responses to congestion. 1187 There are a number of reasons to want to specify performance in term 1188 of multiple concurrent flows, however this approach is not 1189 recommended for data rates below several megabits per second, which 1190 can be attained with run lengths under 10000 packets on many paths. 1191 Since the required run length goes as the square of the data rate, at 1192 higher rates the run lengths can be unreasonably large, and multiple 1193 flows might be the only feasible approach. 1195 If multiple flows are deemed necessary to meet aggregate performance 1196 targets then this MUST be stated both the design of the TIDS and in 1197 any claims about network performance. The IP diagnostic tests MUST 1198 be performed concurrently with the specified number of connections. 1199 For the the tests that use bursty test streams, the bursts should be 1200 synchronized across streams. 1202 7. Interpreting the Results 1204 7.1. Test outcomes 1206 To perform an exhaustive test of a complete network path, each test 1207 of the TIDS is applied to each subpath of the complete path. If any 1208 subpath fails any test then a standard transport protocol running 1209 over the complete path can also be expected to fail to attain the 1210 Target Transport Performance under some conditions. 1212 In addition to passing or failing, a test can be deemed to be 1213 inconclusive for a number of reasons. Proper instrumentation and 1214 treatment of inconclusive outcomes is critical to the accuracy and 1215 robustness of Model Based Metrics. Tests can be inconclusive if the 1216 precomputed traffic pattern or data rates were not accurately 1217 generated; the measurement results were not statistically 1218 significant; and others causes such as failing to meet some required 1219 preconditions for the test. See Section 5.4 1221 For example consider a test that implements Constant Window Pseudo 1222 CBR (Section 6.2) by adding rate controls and detailed IP packet 1223 transfer instrumentation to TCP (e.g. [RFC4898]). TCP includes 1224 built in control systems which might interfere with the sending data 1225 rate. If such a test meets the required packet transfer statistics 1226 (e.g. run length) while failing to attain the specified data rate it 1227 must be treated as an inconclusive result, because we can not a 1228 priori determine if the reduced data rate was caused by a TCP problem 1229 or a network problem, or if the reduced data rate had a material 1230 effect on the observed packet transfer statistics. 1232 Note that for capacity tests, if the observed packet transfer 1233 statistics meet the statistical criteria for failing (accepting 1234 hypnosis H1 in Section 7.2), the test can can be considered to have 1235 failed because it doesn't really matter that the test didn't attain 1236 the required data rate. 1238 The really important new properties of MBM, such as vantage 1239 independence, are a direct consequence of opening the control loops 1240 in the protocols, such that the test stream does not depend on 1241 network conditions or IP packets received. Any mechanism that 1242 introduces feedback between the path's measurements and the test 1243 stream generation is at risk of introducing nonlinearities that spoil 1244 these properties. Any exceptional event that indicates that such 1245 feedback has happened should cause the test to be considered 1246 inconclusive. 1248 One way to view inconclusive tests is that they reflect situations 1249 where a test outcome is ambiguous between limitations of the network 1250 and some unknown limitation of the IP diagnostic test itself, which 1251 may have been caused by some uncontrolled feedback from the network. 1253 Note that procedures that attempt to sweep the target parameter space 1254 to find the limits on some parameter such as target_data_rate are at 1255 risk of breaking the location independent properties of Model Based 1256 Metrics, if any part of the boundary between passing and inconclusive 1257 results is sensitive to RTT (which is normally the case). 1259 One of the goals for evolving TIDS designs will be to keep sharpening 1260 distinction between inconclusive, passing and failing tests. The 1261 criteria for for passing, failing and inconclusive tests MUST be 1262 explicitly stated for every test in the TIDS or FSTIDS. 1264 One of the goals of evolving the testing process, procedures, tools 1265 and measurement point selection should be to minimize the number of 1266 inconclusive tests. 1268 It may be useful to keep raw packet transfer statistics and ancillary 1269 metrics [RFC3148] for deeper study of the behavior of the network 1270 path and to measure the tools themselves. Raw packet transfer 1271 statistics can help to drive tool evolution. Under some conditions 1272 it might be possible to reevaluate the raw data for satisfying 1273 alternate Target Transport Performance. However it is important to 1274 guard against sampling bias and other implicit feedback which can 1275 cause false results and exhibit measurement point vantage 1276 sensitivity. Simply applying different delivery criteria based on a 1277 different Target Transport Performance is insufficient if the test 1278 traffic patterns (bursts, etc) does not match the alternate Target 1279 Transport Performance. 1281 7.2. Statistical criteria for estimating run_length 1283 When evaluating the observed run_length, we need to determine 1284 appropriate packet stream sizes and acceptable error levels for 1285 efficient measurement. In practice, can we compare the empirically 1286 estimated packet loss and ECN Congestion Experienced (CE) marking 1287 ratios with the targets as the sample size grows? How large a sample 1288 is needed to say that the measurements of packet transfer indicate a 1289 particular run length is present? 1291 The generalized measurement can be described as recursive testing: 1292 send packets (individually or in patterns) and observe the packet 1293 delivery performance (packet loss ratio or other metric, any marking 1294 we define). 1296 As each packet is sent and measured, we have an ongoing estimate of 1297 the performance in terms of the ratio of packet loss or ECN CE mark 1298 to total packets (i.e. an empirical probability). We continue to 1299 send until conditions support a conclusion or a maximum sending limit 1300 has been reached. 1302 We have a target_mark_probability, 1 mark per target_run_length, 1303 where a "mark" is defined as a lost packet, a packet with ECN CE 1304 mark, or other signal. This constitutes the null Hypothesis: 1306 H0: no more than one mark in target_run_length = 1307 3*(target_window_size)^2 packets 1309 and we can stop sending packets if on-going measurements support 1310 accepting H0 with the specified Type I error = alpha (= 0.05 for 1311 example). 1313 We also have an alternative Hypothesis to evaluate: if performance is 1314 significantly lower than the target_mark_probability. Based on 1315 analysis of typical values and practical limits on measurement 1316 duration, we choose four times the H0 probability: 1318 H1: one or more marks in (target_run_length/4) packets 1320 and we can stop sending packets if measurements support rejecting H0 1321 with the specified Type II error = beta (= 0.05 for example), thus 1322 preferring the alternate hypothesis H1. 1324 H0 and H1 constitute the Success and Failure outcomes described 1325 elsewhere in the memo, and while the ongoing measurements do not 1326 support either hypothesis the current status of measurements is 1327 inconclusive. 1329 The problem above is formulated to match the Sequential Probability 1330 Ratio Test (SPRT) [StatQC]. Note that as originally framed the 1331 events under consideration were all manufacturing defects. In 1332 networking, ECN CE marks and lost packets are not defects but 1333 signals, indicating that the transport protocol should slow down. 1335 The Sequential Probability Ratio Test also starts with a pair of 1336 hypothesis specified as above: 1338 H0: p0 = one defect in target_run_length 1339 H1: p1 = one defect in target_run_length/4 1340 As packets are sent and measurements collected, the tester evaluates 1341 the cumulative defect count against two boundaries representing H0 1342 Acceptance or Rejection (and acceptance of H1): 1344 Acceptance line: Xa = -h1 + s*n 1345 Rejection line: Xr = h2 + s*n 1346 where n increases linearly for each packet sent and 1348 h1 = { log((1-alpha)/beta) }/k 1349 h2 = { log((1-beta)/alpha) }/k 1350 k = log{ (p1(1-p0)) / (p0(1-p1)) } 1351 s = [ log{ (1-p0)/(1-p1) } ]/k 1352 for p0 and p1 as defined in the null and alternative Hypotheses 1353 statements above, and alpha and beta as the Type I and Type II 1354 errors. 1356 The SPRT specifies simple stopping rules: 1358 o Xa < defect_count(n) < Xb: continue testing 1359 o defect_count(n) <= Xa: Accept H0 1360 o defect_count(n) >= Xb: Accept H1 1362 The calculations above are implemented in the R-tool for Statistical 1363 Analysis [Rtool] , in the add-on package for Cross-Validation via 1364 Sequential Testing (CVST) [CVST] . 1366 Using the equations above, we can calculate the minimum number of 1367 packets (n) needed to accept H0 when x defects are observed. For 1368 example, when x = 0: 1370 Xa = 0 = -h1 + s*n 1371 and n = h1 / s 1373 7.3. Reordering Tolerance 1375 All tests must be instrumented for packet level reordering [RFC4737]. 1376 However, there is no consensus for how much reordering should be 1377 acceptable. Over the last two decades the general trend has been to 1378 make protocols and applications more tolerant to reordering (see for 1379 example [RFC4015]), in response to the gradual increase in reordering 1380 in the network. This increase has been due to the deployment of 1381 technologies such as multi threaded routing lookups and Equal Cost 1382 MultiPath (ECMP) routing. These techniques increase parallelism in 1383 network and are critical to enabling overall Internet growth to 1384 exceed Moore's Law. 1386 Note that transport retransmission strategies can trade off 1387 reordering tolerance vs how quickly they can repair losses vs 1388 overhead from spurious retransmissions. In advance of new 1389 retransmission strategies we propose the following strawman: 1390 Transport protocols should be able to adapt to reordering as long as 1391 the reordering extent is not more than the maximum of one quarter 1392 window or 1 mS, whichever is larger. Within this limit on reorder 1393 extent, there should be no bound on reordering density. 1395 By implication, recording which is less than these bounds should not 1396 be treated as a network impairment. However [RFC4737] still applies: 1397 reordering should be instrumented and the maximum reordering that can 1398 be properly characterized by the test (e.g. bound on history buffers) 1399 should be recorded with the measurement results. 1401 Reordering tolerance and diagnostic limitations, such as the size of 1402 the history buffer used to diagnose packets that are way out-of- 1403 order, MUST be specified in a FSTIDS. 1405 8. IP Diagnostic Tests 1407 The IP diagnostic tests below are organized by traffic pattern: basic 1408 data rate and packet transfer statistics, standing queues, slowstart 1409 bursts, and sender rate bursts. We also introduce some combined 1410 tests which are more efficient when networks are expected to pass, 1411 but conflate diagnostic signatures when they fail. 1413 There are a number of test details which are not fully defined here. 1414 They must be fully specified in a FSTIDS. From a standardization 1415 perspective, this lack of specificity will weaken this version of 1416 Model Based Metrics, however it is anticipated that this it be more 1417 than offset by the extent to which MBM suppresses the problems caused 1418 by using transport protocols for measurement. e.g. non-specific MBM 1419 metrics are likely to have better repeatability than many existing 1420 BTC like metrics. Once we have good field experience, the missing 1421 details can be fully specified. 1423 8.1. Basic Data Rate and Packet Transfer Tests 1425 We propose several versions of the basic data rate and packet 1426 transfer statistics test. All measure the number of packets 1427 delivered between losses or ECN Congestion Experienced (CE) marks, 1428 using a data stream that is rate controlled at or below the 1429 target_data_rate. 1431 The tests below differ in how the data rate is controlled. The data 1432 can be paced on a timer, or window controlled at full Target Data 1433 Rate. The first two tests implicitly confirm that sub_path has 1434 sufficient raw capacity to carry the target_data_rate. They are 1435 recommend for relatively infrequent testing, such as an installation 1436 or periodic auditing process. The third, background packet transfer 1437 statistics, is a low rate test designed for ongoing monitoring for 1438 changes in subpath quality. 1440 All rely on the data receiver accumulating packet transfer statistics 1441 as described in Section 7.2 to score the outcome: 1443 Pass: it is statistically significant that the observed interval 1444 between losses or ECN CE marks is larger than the target_run_length. 1446 Fail: it is statistically significant that the observed interval 1447 between losses or ECN CE marks is smaller than the target_run_length. 1449 A test is considered to be inconclusive if it failed to meet the data 1450 rate as specified below, meet the qualifications defined in 1451 Section 5.4 or neither run length statistical hypothesis was 1452 confirmed in the allotted test duration. 1454 8.1.1. Delivery Statistics at Paced Full Data Rate 1456 Confirm that the observed run length is at least the 1457 target_run_length while relying on timer to send data at the 1458 target_rate using the procedure described in in Section 6.1 with a 1459 burst size of 1 (single packets) or 2 (packet pairs). 1461 The test is considered to be inconclusive if the packet transmission 1462 can not be accurately controlled for any reason. 1464 RFC 6673 [RFC6673] is appropriate for measuring packet transfer 1465 statistics at full data rate. 1467 8.1.2. Delivery Statistics at Full Data Windowed Rate 1469 Confirm that the observed run length is at least the 1470 target_run_length while sending at an average rate approximately 1471 equal to the target_data_rate, by controlling (or clamping) the 1472 window size of a conventional transport protocol to a fixed value 1473 computed from the properties of the test path, typically 1474 test_window=target_data_rate*test_path_RTT/target_MTU. Note that if 1475 there is any interaction between the forward and return path, 1476 test_window may need to be adjusted slightly to compensate for the 1477 resulting inflated RTT. However see the discussion in Section 8.2.4. 1479 Since losses and ECN CE marks cause transport protocols to reduce 1480 their data rates, this test is expected to be less precise about 1481 controlling its data rate. It should not be considered inconclusive 1482 as long as at least some of the round trips reached the full 1483 target_data_rate without incurring losses or ECN CE marks. To pass 1484 this test the network MUST deliver target_window_size packets in 1485 target_RTT time without any losses or ECN CE marks at least once per 1486 two target_window_size round trips, in addition to meeting the run 1487 length statistical test. 1489 8.1.3. Background Packet Transfer Statistics Tests 1491 The background run length is a low rate version of the target target 1492 rate test above, designed for ongoing lightweight monitoring for 1493 changes in the observed subpath run length without disrupting users. 1494 It should be used in conjunction with one of the above full rate 1495 tests because it does not confirm that the subpath can support raw 1496 data rate. 1498 RFC 6673 [RFC6673] is appropriate for measuring background packet 1499 transfer statistics. 1501 8.2. Standing Queue Tests 1503 These engineering tests confirm that the bottleneck is well behaved 1504 across the onset of packet loss, which typically follows after the 1505 onset of queueing. Well behaved generally means lossless for 1506 transient queues, but once the queue has been sustained for a 1507 sufficient period of time (or reaches a sufficient queue depth) there 1508 should be a small number of losses to signal to the transport 1509 protocol that it should reduce its window. Losses that are too early 1510 can prevent the transport from averaging at the target_data_rate. 1511 Losses that are too late indicate that the queue might be subject to 1512 bufferbloat [wikiBloat] and inflict excess queuing delays on all 1513 flows sharing the bottleneck queue. Excess losses (more than half of 1514 the window) at the onset of congestion make loss recovery problematic 1515 for the transport protocol. Non-linear, erratic or excessive RTT 1516 increases suggest poor interactions between the channel acquisition 1517 algorithms and the transport self clock. All of the tests in this 1518 section use the same basic scanning algorithm, described here, but 1519 score the link or subpath on the basis of how well it avoids each of 1520 these problems. 1522 For some technologies the data might not be subject to increasing 1523 delays, in which case the data rate will vary with the window size 1524 all the way up to the onset of load induced packet loss or ECN CE 1525 marks. For theses technologies, the discussion of queueing does not 1526 apply, but it is still required that the onset of losses or ECN CE 1527 marks be at an appropriate point and progressive. 1529 Use the procedure in Section 6.3 to sweep the window across the onset 1530 of queueing and the onset of loss. The tests below all assume that 1531 the scan emulates standard additive increase and delayed ACK by 1532 incrementing the window by one packet for every 2*target_window_size 1533 packets delivered. A scan can typically be divided into three 1534 regions: below the onset of queueing, a standing queue, and at or 1535 beyond the onset of loss. 1537 Below the onset of queueing the RTT is typically fairly constant, and 1538 the data rate varies in proportion to the window size. Once the data 1539 rate reaches the subpath IP rate, the data rate becomes fairly 1540 constant, and the RTT increases in proportion to the increase in 1541 window size. The precise transition across the start of queueing can 1542 be identified by the maximum network power, defined to be the ratio 1543 data rate over the RTT. The network power can be computed at each 1544 window size, and the window with the maximum are taken as the start 1545 of the queueing region. 1547 For technologies that do not have conventional queues, start the scan 1548 at a window equal to the test_window=target_data_rate*test_path_RTT/ 1549 target_MTU, i.e. starting at the target rate, instead of the power 1550 point. 1552 If there is random background loss (e.g. bit errors, etc), precise 1553 determination of the onset of queue induced packet loss may require 1554 multiple scans. Above the onset of queuing loss, all transport 1555 protocols are expected to experience periodic losses determined by 1556 the interaction between the congestion control and AQM algorithms. 1557 For standard congestion control algorithms the periodic losses are 1558 likely to be relatively widely spaced and the details are typically 1559 dominated by the behavior of the transport protocol itself. For the 1560 stiffened transport protocols case (with non-standard, aggressive 1561 congestion control algorithms) the details of periodic losses will be 1562 dominated by how the the window increase function responds to loss. 1564 8.2.1. Congestion Avoidance 1566 A subpath passes the congestion avoidance standing queue test if more 1567 than target_run_length packets are delivered between the onset of 1568 queueing (as determined by the window with the maximum network power) 1569 and the first loss or ECN CE mark. If this test is implemented using 1570 a standards congestion control algorithm with a clamp, it can be 1571 performed in situ in the production internet as a capacity test. For 1572 an example of such a test see [Pathdiag]. 1574 For technologies that do not have conventional queues, use the 1575 test_window in place of the onset of queueing. i.e. A subpath passes 1576 the congestion avoidance standing queue test if more than 1577 target_run_length packets are delivered between start of the scan at 1578 test_window and the first loss or ECN CE mark. 1580 8.2.2. Bufferbloat 1582 This test confirms that there is some mechanism to limit buffer 1583 occupancy (e.g. that prevents bufferbloat). Note that this is not 1584 strictly a requirement for single stream bulk transport capacity, 1585 however if there is no mechanism to limit buffer queue occupancy then 1586 a single stream with sufficient data to deliver is likely to cause 1587 the problems described in [RFC7567], and [wikiBloat]. This may cause 1588 only minor symptoms for the dominant flow, but has the potential to 1589 make the subpath unusable for other flows and applications. 1591 Pass if the onset of loss occurs before a standing queue has 1592 introduced more delay than than twice target_RTT, or other well 1593 defined and specified limit. Note that there is not yet a model for 1594 how much standing queue is acceptable. The factor of two chosen here 1595 reflects a rule of thumb. In conjunction with the previous test, 1596 this test implies that the first loss should occur at a queueing 1597 delay which is between one and two times the target_RTT. 1599 Specified RTT limits that are larger than twice the target_RTT must 1600 be fully justified in the FSTIDS. 1602 8.2.3. Non excessive loss 1604 This test confirm that the onset of loss is not excessive. Pass if 1605 losses are equal or less than the increase in the cross traffic plus 1606 the test stream window increase on the previous RTT. This could be 1607 restated as non-decreasing subpath throughput at the onset of loss, 1608 which is easy to meet as long as discarding packets is not more 1609 expensive than delivering them. (Note when there is a transient drop 1610 in subpath throughput, outside of a standing queue test, a subpath 1611 that passes other queue tests in this document will have sufficient 1612 queue space to hold one RTT worth of data). 1614 Note that conventional Internet policers will not pass this test, 1615 which is correct. TCP often fails to come into equilibrium at more 1616 than a small fraction of the available capacity, if the capacity is 1617 enforced by a policer. [Citation Pending]. 1619 8.2.4. Duplex Self Interference 1621 This engineering test confirms a bound on the interactions between 1622 the forward data path and the ACK return path. 1624 Some historical half duplex technologies had the property that each 1625 direction held the channel until it completely drained its queue. 1626 When a self clocked transport protocol, such as TCP, has data and 1627 ACKs passing in opposite directions through such a link, the behavior 1628 often reverts to stop-and-wait. Each additional packet added to the 1629 window raises the observed RTT by two packet times, once as it passes 1630 through the data path, and once for the additional delay incurred by 1631 the ACK waiting on the return path. 1633 The duplex self interference test fails if the RTT rises by more than 1634 a fixed bound above the expected queueing time computed from trom the 1635 excess window divided by the subpath IP Capacity. This bound must be 1636 smaller than target_RTT/2 to avoid reverting to stop and wait 1637 behavior. (e.g. Data packets and ACKs both have to be released at 1638 least twice per RTT.) 1640 8.3. Slowstart tests 1642 These tests mimic slowstart: data is sent at twice the effective 1643 bottleneck rate to exercise the queue at the dominant bottleneck. 1645 8.3.1. Full Window slowstart test 1647 This is a capacity test to confirm that slowstart is not likely to 1648 exit prematurely. Send slowstart bursts that are target_window_size 1649 total packets. 1651 Accumulate packet transfer statistics as described in Section 7.2 to 1652 score the outcome. Pass if it is statistically significant that the 1653 observed number of good packets delivered between losses or ECN CE 1654 marks is larger than the target_run_length. Fail if it is 1655 statistically significant that the observed interval between losses 1656 or ECN CE marks is smaller than the target_run_length. 1658 It is deemed inconclusive if the elapsed time to send the data burst 1659 is not less than half of the time to receive the ACKs. (i.e. sending 1660 data too fast is ok, but sending it slower than twice the actual 1661 bottleneck rate as indicated by the ACKs is deemed inconclusive). 1662 The headway for the slowstart bursts should be the target_RTT. 1664 Note that these are the same parameters as the Sender Full Window 1665 burst test, except the burst rate is at slowestart rate, rather than 1666 sender interface rate. 1668 8.3.2. Slowstart AQM test 1670 Do a continuous slowstart (send data continuously at twice the 1671 implied IP bottleneck capacity), until the first loss, stop, allow 1672 the network to drain and repeat, gathering statistics on how many 1673 packets were delivered before the loss, the pattern of losses, 1674 maximum observed RTT and window size. Justify the results. There is 1675 not currently sufficient theory justifying requiring any particular 1676 result, however design decisions that affect the outcome of this 1677 tests also affect how the network balances between long and short 1678 flows (the "mice vs elephants" problem). The queue at the time of 1679 the first loss should be at least one half of the target_RTT. 1681 This is an engineering test: It must be performed on a quiescent 1682 network or testbed, since cross traffic has the potential to change 1683 the results. 1685 8.4. Sender Rate Burst tests 1687 These tests determine how well the network can deliver bursts sent at 1688 sender's interface rate. Note that this test most heavily exercises 1689 the front path, and is likely to include infrastructure may be out of 1690 scope for an access ISP, even though the bursts might be caused by 1691 ACK compression, thinning or channel arbitration in the access ISP. 1692 See Appendix B. 1694 Also, there are a several details that are not precisely defined. 1695 For starters there is not a standard server interface rate. 1 Gb/s 1696 and 10 Gb/s are common today, but higher rates will become cost 1697 effective and can be expected to be dominant some time in the future. 1699 Current standards permit TCP to send a full window bursts following 1700 an application pause. (Congestion Window Validation [RFC2861], is 1701 not required, but even if was, it does not take effect until an 1702 application pause is longer than an RTO.) Since full window bursts 1703 are consistent with standard behavior, it is desirable that the 1704 network be able to deliver such bursts, otherwise application pauses 1705 will cause unwarranted losses. Note that the AIMD sawtooth requires 1706 a peak window that is twice target_window_size, so the worst case 1707 burst may be 2*target_window_size. 1709 It is also understood in the application and serving community that 1710 interface rate bursts have a cost to the network that has to be 1711 balanced against other costs in the servers themselves. For example 1712 TCP Segmentation Offload (TSO) reduces server CPU in exchange for 1713 larger network bursts, which increase the stress on network buffer 1714 memory. Some newer TCP implementations can pace traffic at scale 1715 [TSO_pacing][TSO_fq_pacing]. It remains to be determined if and how 1716 quickly these changes will be deployed. 1718 There is not yet theory to unify these costs or to provide a 1719 framework for trying to optimize global efficiency. We do not yet 1720 have a model for how much the network should tolerate server rate 1721 bursts. Some bursts must be tolerated by the network, but it is 1722 probably unreasonable to expect the network to be able to efficiently 1723 deliver all data as a series of bursts. 1725 For this reason, this is the only test for which we encourage 1726 derating. A TIDS could include a table of pairs of derating 1727 parameters: burst sizes and how much each burst size is permitted to 1728 reduce the run length, relative to to the target_run_length. 1730 8.5. Combined and Implicit Tests 1732 Combined tests efficiently confirm multiple network properties in a 1733 single test, possibly as a side effect of normal content delivery. 1734 They require less measurement traffic than other testing strategies 1735 at the cost of conflating diagnostic signatures when they fail. 1736 These are by far the most efficient for monitoring networks that are 1737 nominally expected to pass all tests. 1739 8.5.1. Sustained Bursts Test 1741 The sustained burst test implements a combined worst case version of 1742 all of the capacity tests above. It is simply: 1744 Send target_window_size bursts of packets at server interface rate 1745 with target_RTT burst headway (burst start to burst start). Verify 1746 that the observed packet transfer statistics meets the 1747 target_run_length. 1749 Key observations: 1750 o The subpath under test is expected to go idle for some fraction of 1751 the time: (subpath_IP_capacity-target_rate/ 1752 (target_MTU-header_overhead)*target_MTU)/subpath_IP_capacity. 1753 Failing to do so indicates a problem with the procedure and an 1754 inconclusive test result. 1755 o The burst sensitivity can be derated by sending smaller bursts 1756 more frequently. E.g. send target_window_size*derate packet 1757 bursts every target_RTT*derate. 1758 o When not derated, this test is the most strenuous capacity test. 1759 o A subpath that passes this test is likely to be able to sustain 1760 higher rates (close to subpath_IP_capacity) for paths with RTTs 1761 significantly smaller than the target_RTT. 1762 o This test can be implemented with instrumented TCP [RFC4898], 1763 using a specialized measurement application at one end [MBMSource] 1764 and a minimal service at the other end [RFC0863] [RFC0864]. 1765 o This test is efficient to implement, since it does not require 1766 per-packet timers, and can make use of TSO in modern NIC hardware. 1767 o If a subpath is known to pass the Standing Queue engineering tests 1768 (particularly that it has a progressive onset of loss at an 1769 appropriate queue depth), then the Sustained Burst Test is 1770 sufficient to assure that the subpath under test will not impair 1771 Bulk Transport Capacity at the target performance under all 1772 conditions. See Section 8.2 for a discussion of the standing 1773 queue tests. 1775 Note that this test is clearly independent of the subpath RTT, or 1776 other details of the measurement infrastructure, as long as the 1777 measurement infrastructure can accurately and reliably deliver the 1778 required bursts to the subpath under test. 1780 8.5.2. Streaming Media 1782 Model Based Metrics can be implicitly implemented as a side effect 1783 any non-throughput maximizing application, such as streaming media, 1784 with some additional controls and instrumentation in the servers. 1785 The essential requirement is that the data rate be constrained such 1786 that even with arbitrary application pauses and bursts the data rate 1787 and burst sizes stay within the envelope defined by the individual 1788 tests described above. 1790 If the application's serving_data_rate is less than or equal to the 1791 target_data_rate and the serving_RTT (the RTT between the sender and 1792 client) is less than the target_RTT, this constraint is most easily 1793 implemented by clamping the transport window size to be no larger 1794 than: 1796 serving_window_clamp=target_data_rate*serving_RTT/ 1797 (target_MTU-header_overhead) 1799 Under the above constraints the serving_window_clamp will limit the 1800 both the serving data rate and burst sizes to be no larger than the 1801 procedures in Section 8.1.2 and Section 8.4 or Section 8.5.1. Since 1802 the serving RTT is smaller than the target_RTT, the worst case bursts 1803 that might be generated under these conditions will be smaller than 1804 called for by Section 8.4 and the sender rate burst sizes are 1805 implicitly derated by the serving_window_clamp divided by the 1806 target_window_size at the very least. (Depending on the application 1807 behavior, the data might be significantly smoother than specified by 1808 any of the burst tests.) 1810 In an alternative implementation the data rate and bursts might be 1811 explicitly controlled by a programmable traffic shaper or pacing at 1812 the sender. This would provide better control over transmissions but 1813 it is substantially more complicated to implement and would be likely 1814 to have a higher CPU overhead. 1816 Note that these techniques can be applied to any content delivery 1817 that can be subjected to a reduced data rate in order to inhibit TCP 1818 equilibrium behavior. 1820 9. An Example 1822 In this section a we illustrate a TIDS designed to confirm that an 1823 access ISP can reliably deliver HD video from multiple content 1824 providers to all of their customers. With modern codecs, minimal HD 1825 video (720p) generally fits in 2.5 Mb/s. Due to their geographical 1826 size, network topology and modem designs the ISP determines that most 1827 content is within a 50 mS RTT from their users (This is a sufficient 1828 to cover continental Europe or either US coast from a single serving 1829 site.) 1831 2.5 Mb/s over a 50 ms path 1833 +----------------------+-------+---------+ 1834 | End-to-End Parameter | value | units | 1835 +----------------------+-------+---------+ 1836 | target_rate | 2.5 | Mb/s | 1837 | target_RTT | 50 | ms | 1838 | target_MTU | 1500 | bytes | 1839 | header_overhead | 64 | bytes | 1840 | target_window_size | 11 | packets | 1841 | target_run_length | 363 | packets | 1842 +----------------------+-------+---------+ 1844 Table 1 1846 Table 1 shows the default TCP model with no derating, and as such is 1847 quite conservative. The simplest TIDS would be to use the sustained 1848 burst test, described in Section 8.5.1. Such a test would send 11 1849 packet bursts every 50mS, and confirming that there was no more than 1850 1 packet loss per 33 bursts (363 total packets in 1.650 seconds). 1852 Since this number represents is the entire end-to-end loss budget, 1853 independent subpath tests could be implemented by apportioning the 1854 packet loss ratio across subpaths. For example 50% of the losses 1855 might be allocated to the access or last mile link to the user, 40% 1856 to the interconnects with other ISPs and 1% to each internal hop 1857 (assuming no more than 10 internal hops). Then all of the subpaths 1858 can be tested independently, and the spatial composition of passing 1859 subpaths would be expected to be within the end-to-end loss budget. 1861 Testing interconnects has generally been problematic: conventional 1862 performance tests run between measurement points adjacent to either 1863 side of the interconnect, are not generally useful. Unconstrained 1864 TCP tests, such as iperf [iperf] are usually overly aggressive 1865 because the RTT is so small (often less than 1 mS). With a short RTT 1866 these tools are likely to report inflated numbers because for short 1867 RTTs these tools can tolerate very high packet loss ratios and can 1868 push other cross traffic off of the network. As a consequence they 1869 are useless for predicting actual user performance, and may 1870 themselves be quite disruptive. Model Based Metrics solves this 1871 problem. The same test pattern as used on other subpaths can be 1872 applied to the interconnect. For our example, when apportioned 40% 1873 of the losses, 11 packet bursts sent every 50mS should have fewer 1874 than one loss per 82 bursts (902 packets). 1876 10. Validation 1878 Since some aspects of the models are likely to be too conservative, 1879 Section 5.2 permits alternate protocol models and Section 5.3 permits 1880 test parameter derating. If either of these techniques are used, we 1881 require demonstrations that such a TIDS can robustly detect subpaths 1882 that will prevent authentic applications using state-of-the-art 1883 protocol implementations from meeting the specified Target Transport 1884 Performance. This correctness criteria is potentially difficult to 1885 prove, because it implicitly requires validating a TIDS against all 1886 possible subpaths and subpaths. The procedures described here are 1887 still experimental. 1889 We suggest two approaches, both of which should be applied: first, 1890 publish a fully open description of the TIDS, including what 1891 assumptions were used and and how it was derived, such that the 1892 research community can evaluate the design decisions, test them and 1893 comment on their applicability; and second, demonstrate that an 1894 applications running over an infinitesimally passing testbed do meet 1895 the performance targets. 1897 An infinitesimally passing testbed resembles a epsilon-delta proof in 1898 calculus. Construct a test network such that all of the individual 1899 tests of the TIDS pass by only small (infinitesimal) margins, and 1900 demonstrate that a variety of authentic applications running over 1901 real TCP implementations (or other protocol as appropriate) meets the 1902 Target Transport Performance over such a network. The workloads 1903 should include multiple types of streaming media and transaction 1904 oriented short flows (e.g. synthetic web traffic). 1906 For example, for the HD streaming video TIDS described in Section 9, 1907 the IP capacity should be exactly the header overhead above 2.5 Mb/s, 1908 the per packet random background loss ratio should be 1/363, for a 1909 run length of 363 packets, the bottleneck queue should be 11 packets 1910 and the front path should have just enough buffering to withstand 11 1911 packet interface rate bursts. We want every one of the TIDS tests to 1912 fail if we slightly increase the relevant test parameter, so for 1913 example sending a 12 packet bursts should cause excess (possibly 1914 deterministic) packet drops at the dominant queue at the bottleneck. 1915 On this infinitesimally passing network it should be possible for a 1916 real application using a stock TCP implementation in the vendor's 1917 default configuration to attain 2.5 Mb/s over an 50 mS path. 1919 The most difficult part of setting up such a testbed is arranging for 1920 it to infinitesimally pass the individual tests. Two approaches: 1921 constraining the network devices not to use all available resources 1922 (e.g. by limiting available buffer space or data rate); and 1923 preloading subpaths with cross traffic. Note that is it important 1924 that a single environment be constructed which infinitesimally passes 1925 all tests at the same time, otherwise there is a chance that TCP can 1926 exploit extra latitude in some parameters (such as data rate) to 1927 partially compensate for constraints in other parameters (queue 1928 space, or viceversa). 1930 To the extent that a TIDS is used to inform public dialog it should 1931 be fully publicly documented, including the details of the tests, 1932 what assumptions were used and how it was derived. All of the 1933 details of the validation experiment should also be published with 1934 sufficient detail for the experiments to be replicated by other 1935 researchers. All components should either be open source of fully 1936 described proprietary implementations that are available to the 1937 research community. 1939 11. Security Considerations 1941 Measurement is often used to inform business and policy decisions, 1942 and as a consequence is potentially subject to manipulation. Model 1943 Based Metrics are expected to be a huge step forward because 1944 equivalent measurements can be performed from multiple vantage 1945 points, such that performance claims can be independently validated 1946 by multiple parties. 1948 Much of the acrimony in the Net Neutrality debate is due by the 1949 historical lack of any effective vantage independent tools to 1950 characterize network performance. Traditional methods for measuring 1951 Bulk Transport Capacity are sensitive to RTT and as a consequence 1952 often yield very different results when run local to an ISP or 1953 internconnect and when run over a customer's complete path. Neither 1954 the ISP nor customer can repeat the other's measurements, leading to 1955 high levels of distrust and acrimony. Model Based Metrics are 1956 expected to greatly improve this situation. 1958 This document only describes a framework for designing Fully 1959 Specified Targeted IP Diagnostic Suite. Each FSTIDS MUST include its 1960 own security section. 1962 12. Acknowledgements 1964 Ganga Maguluri suggested the statistical test for measuring loss 1965 probability in the target run length. Alex Gilgur for helping with 1966 the statistics. 1968 Meredith Whittaker for improving the clarity of the communications. 1970 Ruediger Geib provided feedback which greatly improved the document. 1972 This work was inspired by Measurement Lab: open tools running on an 1973 open platform, using open tools to collect open data. See 1974 http://www.measurementlab.net/ 1976 13. IANA Considerations 1978 This document has no actions for IANA. 1980 14. References 1982 14.1. Normative References 1984 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 1985 Requirement Levels", BCP 14, RFC 2119, March 1997. 1987 14.2. Informative References 1989 [RFC0863] Postel, J., "Discard Protocol", STD 21, RFC 863, May 1983. 1991 [RFC0864] Postel, J., "Character Generator Protocol", STD 22, 1992 RFC 864, May 1983. 1994 [RFC2330] Paxson, V., Almes, G., Mahdavi, J., and M. Mathis, 1995 "Framework for IP Performance Metrics", RFC 2330, 1996 May 1998. 1998 [RFC2861] Handley, M., Padhye, J., and S. Floyd, "TCP Congestion 1999 Window Validation", RFC 2861, June 2000. 2001 [RFC3148] Mathis, M. and M. Allman, "A Framework for Defining 2002 Empirical Bulk Transfer Capacity Metrics", RFC 3148, 2003 July 2001. 2005 [RFC3465] Allman, M., "TCP Congestion Control with Appropriate Byte 2006 Counting (ABC)", RFC 3465, February 2003. 2008 [RFC4015] Ludwig, R. and A. Gurtov, "The Eifel Response Algorithm 2009 for TCP", RFC 4015, February 2005. 2011 [RFC4737] Morton, A., Ciavattone, L., Ramachandran, G., Shalunov, 2012 S., and J. Perser, "Packet Reordering Metrics", RFC 4737, 2013 November 2006. 2015 [RFC4898] Mathis, M., Heffner, J., and R. Raghunarayan, "TCP 2016 Extended Statistics MIB", RFC 4898, May 2007. 2018 [RFC5136] Chimento, P. and J. Ishac, "Defining Network Capacity", 2019 RFC 5136, February 2008. 2021 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion 2022 Control", RFC 5681, September 2009. 2024 [RFC5835] Morton, A. and S. Van den Berghe, "Framework for Metric 2025 Composition", RFC 5835, April 2010. 2027 [RFC6049] Morton, A. and E. Stephan, "Spatial Composition of 2028 Metrics", RFC 6049, January 2011. 2030 [RFC6673] Morton, A., "Round-Trip Packet Loss Metrics", RFC 6673, 2031 August 2012. 2033 [RFC6928] Chu, J., Dukkipati, N., Cheng, Y., and M. Mathis, 2034 "Increasing TCP's Initial Window", RFC 6928, DOI 10.17487/ 2035 RFC6928, April 2013, 2036 . 2038 [RFC7312] Fabini, J. and A. Morton, "Advanced Stream and Sampling 2039 Framework for IP Performance Metrics (IPPM)", RFC 7312, 2040 August 2014. 2042 [RFC7398] Bagnulo, M., Burbridge, T., Crawford, S., Eardley, P., and 2043 A. Morton, "A Reference Path and Measurement Points for 2044 Large-Scale Measurement of Broadband Performance", 2045 RFC 7398, February 2015. 2047 [RFC7567] Baker, F., Ed. and G. Fairhurst, Ed., "IETF 2048 Recommendations Regarding Active Queue Management", 2049 BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015, 2050 . 2052 [I-D.ietf-ippm-2680-bis] 2053 Almes, G., Kalidindi, S., Zekauskas, M., and A. Morton, "A 2054 One-Way Loss Metric for IPPM", draft-ietf-ippm-2680-bis-05 2055 (work in progress), August 2015. 2057 [MSMO97] Mathis, M., Semke, J., Mahdavi, J., and T. Ott, "The 2058 Macroscopic Behavior of the TCP Congestion Avoidance 2059 Algorithm", Computer Communications Review volume 27, 2060 number3, July 1997. 2062 [WPING] Mathis, M., "Windowed Ping: An IP Level Performance 2063 Diagnostic", INET 94, June 1994. 2065 [mpingSource] 2066 Fan, X., Mathis, M., and D. Hamon, "Git Repository for 2067 mping: An IP Level Performance Diagnostic", Sept 2013, 2068 . 2070 [MBMSource] 2071 Hamon, D., Stuart, S., and H. Chen, "Git Repository for 2072 Model Based Metrics", Sept 2013, 2073 . 2075 [Pathdiag] 2076 Mathis, M., Heffner, J., O'Neil, P., and P. Siemsen, 2077 "Pathdiag: Automated TCP Diagnosis", Passive and Active 2078 Measurement , June 2008. 2080 [iperf] Wikipedia Contributors, "iPerf", Wikipedia, The Free 2081 Encyclopedia , cited March 2015, . 2084 [StatQC] Montgomery, D., "Introduction to Statistical Quality 2085 Control - 2nd ed.", ISBN 0-471-51988-X, 1990. 2087 [Rtool] R Development Core Team, "R: A language and environment 2088 for statistical computing. R Foundation for Statistical 2089 Computing, Vienna, Austria. ISBN 3-900051-07-0, URL 2090 http://www.R-project.org/", , 2011. 2092 [CVST] Krueger, T. and M. Braun, "R package: Fast Cross- 2093 Validation via Sequential Testing", version 0.1, 11 2012. 2095 [AFD] Pan, R., Breslau, L., Prabhakar, B., and S. Shenker, 2096 "Approximate fairness through differential dropping", 2097 SIGCOMM Comput. Commun. Rev. 33, 2, April 2003. 2099 [wikiBloat] 2100 Wikipedia, "Bufferbloat", http://en.wikipedia.org/w/ 2101 index.php?title=Bufferbloat&oldid=608805474, March 2015. 2103 [CCscaling] 2104 Fernando, F., Doyle, J., and S. Steven, "Scalable laws for 2105 stable network congestion control", Proceedings of 2106 Conference on Decision and 2107 Control, http://www.ee.ucla.edu/~paganini, December 2001. 2109 [TSO_pacing] 2110 Corbet, J., "TSO sizing and the FQ scheduler", 2111 LWN.net https://lwn.net/Articles/564978/, Aug 2013. 2113 [TSO_fq_pacing] 2114 Dumazet, E. and Y. Chen, "TSO, fair queuing, pacing: 2115 three's a charm", Proceedings of IETF 88, TCPM WG https:// 2116 www.ietf.org/proceedings/88/slides/slides-88-tcpm-9.pdf, 2117 Nov 2013. 2119 Appendix A. Model Derivations 2121 The reference target_run_length described in Section 5.2 is based on 2122 very conservative assumptions: that all window above 2123 target_window_size contributes to a standing queue that raises the 2124 RTT, and that classic Reno congestion control with delayed ACKs are 2125 in effect. In this section we provide two alternative calculations 2126 using different assumptions. 2128 It may seem out of place to allow such latitude in a measurement 2129 standard, but this section provides offsetting requirements. 2131 The estimates provided by these models make the most sense if network 2132 performance is viewed logarithmically. In the operational Internet, 2133 data rates span more than 8 orders of magnitude, RTT spans more than 2134 3 orders of magnitude, and packet loss ratio spans at least 8 orders 2135 of magnitude if not more. When viewed logarithmically (as in 2136 decibels), these correspond to 80 dB of dynamic range. On an 80 dB 2137 scale, a 3 dB error is less than 4% of the scale, even though it 2138 represents a factor of 2 in untransformed parameter. 2140 This document gives a lot of latitude for calculating 2141 target_run_length, however people designing a TIDS should consider 2142 the effect of their choices on the ongoing tussle about the relevance 2143 of "TCP friendliness" as an appropriate model for Internet capacity 2144 allocation. Choosing a target_run_length that is substantially 2145 smaller than the reference target_run_length specified in Section 5.2 2146 strengthens the argument that it may be appropriate to abandon "TCP 2147 friendliness" as the Internet fairness model. This gives developers 2148 incentive and permission to develop even more aggressive applications 2149 and protocols, for example by increasing the number of connections 2150 that they open concurrently. 2152 A.1. Queueless Reno 2154 In Section 5.2 models were derived based on the assumption that the 2155 subpath IP rate matches the target rate plus overhead, such that the 2156 excess window needed for the AIMD sawtooth causes a fluctuating queue 2157 at the bottleneck. 2159 An alternate situation would be a bottleneck where there is no 2160 significant queue and losses are caused by some mechanism that does 2161 not involve extra delay, for example by the use of a virtual queue as 2162 done in Approximate Fair Dropping [AFD]. A flow controlled by such a 2163 bottleneck would have a constant RTT and a data rate that fluctuates 2164 in a sawtooth due to AIMD congestion control. Assume the losses are 2165 being controlled to make the average data rate meet some goal which 2166 is equal or greater than the target_rate. The necessary run length 2167 to meet the target_rate can be computed as follows: 2169 For some value of Wmin, the window will sweep from Wmin packets to 2170 2*Wmin packets in 2*Wmin RTT (due to delayed ACK). Unlike the 2171 queueing case where Wmin = target_window_size, we want the average of 2172 Wmin and 2*Wmin to be the target_window_size, so the average data 2173 rate is the target rate. Thus we want Wmin = 2174 (2/3)*target_window_size. 2176 Between losses each sawtooth delivers (1/2)(Wmin+2*Wmin)(2Wmin) 2177 packets in 2*Wmin round trip times. 2179 Substituting these together we get: 2181 target_run_length = (4/3)(target_window_size^2) 2183 Note that this is 44% of the reference_run_length computed earlier. 2184 This makes sense because under the assumptions in Section 5.2 the 2185 AMID sawtooth caused a queue at the bottleneck, which raised the 2186 effective RTT by 50%. 2188 Appendix B. The effects of ACK scheduling 2190 For many network technologies simple queueing models don't apply: the 2191 network schedules, thins or otherwise alters the timing of ACKs and 2192 data, generally to raise the efficiency of the channel allocation 2193 algorithms when confronted with relatively widely spaced small ACKs. 2194 These efficiency strategies are ubiquitous for half duplex, wireless 2195 and broadcast media. 2197 Altering the ACK stream by holding or thinning ACKs typically has two 2198 consequences: it raises the implied bottleneck IP capacity, making 2199 the fine grained slowstart bursts either faster or larger and it 2200 raises the effective RTT by the average time that the ACKs and data 2201 are delayed. The first effect can be partially mitigated by 2202 reclocking ACKs once they are beyond the bottleneck on the return 2203 path to the sender, however this further raises the effective RTT. 2205 The most extreme example of this sort of behavior would be a half 2206 duplex channel that is not released as long as the endpoint currently 2207 holding the channel has more traffic (data or ACKs) to send. Such 2208 environments cause self clocked protocols under full load to revert 2209 to extremely inefficient stop and wait behavior. The channel 2210 constrains the protocol to send an entire window of data as a single 2211 contiguous burst on the forward path, followed by the entire window 2212 of ACKs on the return path. 2214 If a particular return path contains a subpath or device that alters 2215 the timing of the ACK stream, then the entire front path from the 2216 sender up to the bottleneck must be tested at the burst parameters 2217 implied by the ACK scheduling algorithm. The most important 2218 parameter is the Implied Bottleneck IP Capacity, which is the average 2219 rate at which the ACKs advance snd.una. Note that thinning the ACK 2220 stream (relying on the cumulative nature of seg.ack to permit 2221 discarding some ACKs) requires larger sender interface bursts to 2222 offset the longer times between ACK in order to maintain the average 2223 data rate. 2225 It is important to note that due to ubiquitous self clocking in 2226 Internet protocols, ill conceived channel allocation mechanisms 2227 increases the queueing stress on the front path because they cause 2228 larger full sender rate data bursts. 2230 Holding data or ACKs for channel allocation or other reasons (such as 2231 forward error correction) always raises the effective RTT relative to 2232 the minimum delay for the path. Therefore it may be necessary to 2233 replace target_RTT in the calculation in Section 5.2 by an 2234 effective_RTT, which includes the target_RTT plus a term to account 2235 for the extra delays introduced by these mechanisms. 2237 Appendix C. Version Control 2239 This section to be removed prior to publication. 2241 Formatted: Mon Oct 19 15:59:51 PDT 2015 2243 Authors' Addresses 2245 Matt Mathis 2246 Google, Inc 2247 1600 Amphitheater Parkway 2248 Mountain View, California 94043 2249 USA 2251 Email: mattmathis@google.com 2253 Al Morton 2254 AT&T Labs 2255 200 Laurel Avenue South 2256 Middletown, NJ 07748 2257 USA 2259 Phone: +1 732 420 1571 2260 Email: acmorton@att.com 2261 URI: http://home.comcast.net/~acmacm/