idnits 2.17.1 draft-mathis-ippm-model-based-metrics-01.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- ** The document seems to lack a Security Considerations section. ** The document seems to lack an IANA Considerations section. (See Section 2.2 of https://www.ietf.org/id-info/checklist for how to handle the case when there are no actions for IANA.) == There are 1 instance of lines with non-RFC2606-compliant FQDNs in the document. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year == Line 766 has weird spacing: '...lcement the m...' == The document seems to lack the recommended RFC 2119 boilerplate, even if it appears to use RFC 2119 keywords. (The document does seem to have the reference to RFC 2119 which the ID-Checklist requires). -- The document date (Feb 25, 2013) is 4076 days in the past. Is this intentional? Checking references for intended status: Experimental ---------------------------------------------------------------------------- == Missing Reference: 'SFQ' is mentioned on line 195, but not defined == Missing Reference: 'AFD' is mentioned on line 382, but not defined == Missing Reference: 'ConEx' is mentioned on line 426, but not defined == Missing Reference: 'QueueSize' is mentioned on line 506, but not defined == Missing Reference: 'ABC' is mentioned on line 552, but not defined == Missing Reference: 'WPING' is mentioned on line 586, but not defined == Missing Reference: 'PATHDIAG' is mentioned on line 586, but not defined == Missing Reference: 'BufferBloat' is mentioned on line 852, but not defined == Missing Reference: 'TBD' is mentioned on line 902, but not defined == Missing Reference: 'RFC 4898' is mentioned on line 1083, but not defined == Missing Reference: 'RFC 863' is mentioned on line 1084, but not defined == Missing Reference: 'RFC 864' is mentioned on line 1084, but not defined == Unused Reference: 'RFC2119' is defined on line 1131, but no explicit reference was found in the text == Unused Reference: 'RFC2026' is defined on line 1134, but no explicit reference was found in the text == Outdated reference: A later version (-01) exists of draft-morton-ippm-lmap-path-00 Summary: 2 errors (**), 0 flaws (~~), 19 warnings (==), 1 comment (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 IP Performance Working Group M. Mathis 3 Internet-Draft Google, Inc 4 Intended status: Experimental A. Morton 5 Expires: August 29, 2013 AT&T Labs 6 Feb 25, 2013 8 Model Based Internet Performance Metrics 9 draft-mathis-ippm-model-based-metrics-01.txt 11 Abstract 13 We introduce a new class of model based metrics designed to determine 14 if a long path can meet predefined end-to-end application performance 15 targets. This is done by section-by-section testing -- by applying a 16 suite of single property tests to successive sections of a long path. 17 In many cases these single property tests are based on existing IPPM 18 metrics, with the addition of success and validity criteria. The 19 sub-path at a time tests are designed to facilitate IP providers 20 eliminating all known conditions that might prevent the full end-to- 21 end path from meeting the users target performance. 23 This approach makes it possible to to determine the IP performance 24 requirements needed to support the desired end-to-end TCP 25 performance. The IP metrics are based on traffic patterns that mimic 26 TCP but are precomputed independently of the actual behavior of TCP 27 over the sub-path under test. This makes the measurements open loop, 28 eliminating nearly all of the difficulties encountered by traditional 29 bulk transport metrics, which rely on congestion control equilibrium 30 behavior. 32 A natural consequence of this methodology is verifiable network 33 measurement: measurements from any given vantage point are repeatable 34 from other vantage points. 36 Status of this Memo 38 This Internet-Draft is submitted in full conformance with the 39 provisions of BCP 78 and BCP 79. 41 Internet-Drafts are working documents of the Internet Engineering 42 Task Force (IETF). Note that other groups may also distribute 43 working documents as Internet-Drafts. The list of current Internet- 44 Drafts is at http://datatracker.ietf.org/drafts/current/. 46 Internet-Drafts are draft documents valid for a maximum of six months 47 and may be updated, replaced, or obsoleted by other documents at any 48 time. It is inappropriate to use Internet-Drafts as reference 49 material or to cite them other than as "work in progress." 51 This Internet-Draft will expire on August 29, 2013. 53 Copyright Notice 55 Copyright (c) 2013 IETF Trust and the persons identified as the 56 document authors. All rights reserved. 58 This document is subject to BCP 78 and the IETF Trust's Legal 59 Provisions Relating to IETF Documents 60 (http://trustee.ietf.org/license-info) in effect on the date of 61 publication of this document. Please review these documents 62 carefully, as they describe your rights and restrictions with respect 63 to this document. Code Components extracted from this document must 64 include Simplified BSD License text as described in Section 4.e of 65 the Trust Legal Provisions and are provided without warranty as 66 described in the Simplified BSD License. 68 Table of Contents 70 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 5 71 2. New requirements relative to RFC 2330 . . . . . . . . . . . . 7 72 3. Background . . . . . . . . . . . . . . . . . . . . . . . . . . 7 73 4. Common Models and Parameters . . . . . . . . . . . . . . . . . 9 74 4.1. Target End-to-end parameters . . . . . . . . . . . . . . . 9 75 4.2. End-to-end parameters from sub-paths . . . . . . . . . . . 11 76 4.3. Per sub-path parameters . . . . . . . . . . . . . . . . . 11 77 4.4. Common Calculations for Single Property Tests . . . . . . 11 78 4.5. Parameter Derating . . . . . . . . . . . . . . . . . . . . 13 79 5. Common testing procedures . . . . . . . . . . . . . . . . . . 13 80 5.1. Traffic generating techniques . . . . . . . . . . . . . . 13 81 5.1.1. Paced transmission . . . . . . . . . . . . . . . . . . 14 82 5.1.2. Constant window pseudo CBR . . . . . . . . . . . . . . 14 83 5.1.3. Scanned window pseudo CBR . . . . . . . . . . . . . . 14 84 5.1.4. Intermittent Testing . . . . . . . . . . . . . . . . . 15 85 5.1.5. Intermittent Scatter Testing . . . . . . . . . . . . . 15 86 5.2. Interpreting the Results . . . . . . . . . . . . . . . . . 15 87 5.2.1. Inconclusive test outcomes . . . . . . . . . . . . . . 15 88 5.2.2. Statistical criteria for measuring run_length . . . . 16 89 5.2.3. Classifications of tests . . . . . . . . . . . . . . . 17 90 5.3. Reordering Tolerance . . . . . . . . . . . . . . . . . . . 18 91 5.4. Verify the absence of cross traffic . . . . . . . . . . . 18 92 5.5. Additional test preconditions . . . . . . . . . . . . . . 19 93 6. Single Property Tests . . . . . . . . . . . . . . . . . . . . 20 94 6.1. CBR Tests . . . . . . . . . . . . . . . . . . . . . . . . 21 95 6.1.1. Loss Rate at Full Data Rate . . . . . . . . . . . . . 21 96 6.1.2. Background Loss Rate Tests . . . . . . . . . . . . . . 21 97 6.2. Standing Queue tests . . . . . . . . . . . . . . . . . . . 22 98 6.2.1. Congestion Avoidance . . . . . . . . . . . . . . . . . 22 99 6.2.2. Buffer Bloat . . . . . . . . . . . . . . . . . . . . . 23 100 6.2.3. Self Interference . . . . . . . . . . . . . . . . . . 23 101 6.3. Slow Start tests . . . . . . . . . . . . . . . . . . . . . 23 102 6.3.1. Full Window slow start test . . . . . . . . . . . . . 23 103 6.3.2. Slowstart AQM test . . . . . . . . . . . . . . . . . . 24 104 6.4. Server Rate tests . . . . . . . . . . . . . . . . . . . . 24 105 6.4.1. Server TCP Send Offload (TSO) tests . . . . . . . . . 24 106 6.4.2. Server Full Window test . . . . . . . . . . . . . . . 24 107 7. Combined Tests . . . . . . . . . . . . . . . . . . . . . . . . 24 108 7.1. Sustained burst test . . . . . . . . . . . . . . . . . . . 25 109 8. Calibration . . . . . . . . . . . . . . . . . . . . . . . . . 26 110 9. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . 26 111 10. References . . . . . . . . . . . . . . . . . . . . . . . . . . 26 112 10.1. Normative References . . . . . . . . . . . . . . . . . . . 26 113 10.2. Informative References . . . . . . . . . . . . . . . . . . 26 114 Appendix A. Model Derivations . . . . . . . . . . . . . . . . . . 27 115 Appendix B. Old text from an earlier document . . . . . . . . . . 27 116 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . . 29 118 1. Introduction 120 We introduce a new class of model based metrics designed to determine 121 if a long path can be expected to meet a predefined application end- 122 to-end performance target by applying a suite of single property 123 tests to successive sections of the long path. In many cases these 124 single property tests are based on existing IPPM metrics, with the 125 addition of specific success and validity criteria. The sub-path at 126 a time tests are designed to eliminate all known conditions that will 127 potentially prevent the full path from meeting the target 128 performance. 130 The end-to-end target performance must be specified in advance, and 131 models are used to compute the IP layer properties necessary to 132 support that performance. The IP metrics are based on traffic 133 patterns that mimic TCP but are precomputed independently of the 134 actual behavior of TCP over the sub-path under test. 136 This approach makes the measurements open loop, eliminating nearly 137 all of the difficulties encountered by traditional bulk transport 138 metrics, which depend on congestion control equilibrium behavior. 139 Otherwise these control systems inherently have a number of 140 properties that interfere with measurement: they have circular 141 dependencies such that every component affects every property. 143 Since a singleton (see [RFC2330]) is only a pass/fail measurement of 144 a sub-path, these metrics are most useful in composition over large 145 pools of samples, such as across a collection of paths or a time 146 interval [RFC5835] [RFC6049] . 148 For Bulk Transport Capacity (BTC) the target performance to be 149 confirmed is a data rate. TCP's ability to compensate for less than 150 ideal network conditions is fundamentally affected by the RTT and MTU 151 of the end-to-end Internet path that it traverses. Since the minimum 152 RTT and maximum MTU are both fixed properties of the path, they are 153 also taken as parameters to the modeling process. The target values 154 for these three parameters, Data Rate, RTT and MTU, are determined by 155 the application, its intended use and the physical infrastructure 156 over which it traverses. 158 For BTC the following tests are sufficient: 159 o raw data rate, 160 o background loss rate, 161 o queue burst capacity, 162 o reordering extent [RFC4737], 163 o onset of congestion/AQM 164 o and corresponding metrics on return path quality. 165 If every sub-path passes all of these tests, then an end-to-end 166 application using any reasonably modern TCP or similar protocol 167 should be able to attain the specified target data rate, over the 168 full end-to-end path at the specified RTT and MTU. 170 Traditional end-to-end BTC metrics have proven to be difficult or 171 unsatisfactory due to some overlooked requirements described in 172 Section 2 and some intrinsic difficulties with using protocols for 173 measurement described in Section 3. In Section 4 we describe the 174 models and common parameters used to derive single property test 175 parameters. In Section 5 we describe common testing procedures used 176 by all of the tests. Rather than testing the end-to-end path with 177 TCP or other some other BTC, each sub-path is evaluated using suite 178 of far simpler and more predictable single property tests described 179 in Section 6. Section 7 describes some combined tests that are more 180 efficient to implement and deploy. However, if they fail they may 181 not clearly indicate the nature of the problem. 183 There exists the potential that model based metric itself might yield 184 a false pass result, in the sense that every sub-path of an end-to- 185 end path passes every single property test and yet a real application 186 might still fall to attain its performance target over the path. If 187 so, then a traditional BTC needs to be used to validate the tests for 188 each sub-path, as described in Section 8. 190 Future text (or a more likely a future document) will describe model 191 based metrics for real time traffic. The salient point will be that 192 concurrently meeting the goals of both RT and throughput maximizing 193 traffic implicitly requires some form of traffic segregation, such 194 that the two traffic classes are not placed in the same queue. Some 195 technique as simple as Fair Queueing [SFQ] might be a sufficient 196 alternative to full QoS. 198 TODO: 199 o Better terminology for: single property test, test targets (as 200 opposed to end-to-end targets), packet layer properties(?), 201 testing suites, combined tests, etc. All to strengthen of the 202 linguistic differences between transport and network layer. 203 o Make it clear that this document is about traffic patterns and 204 delivery statistics. Other aspects of the test procedures are out 205 of scope. 206 o Add description of assumed TCP behaviors used to derived the 207 models. 208 o Eliminate sequentiality both as a modeling process and for section 209 by section testing. Treatment of link under test being different 210 from the bottleneck link (e.g. testing to an aggregation point). 212 o Add "effective bottleneck rate" as an end-to-end parameter. 213 Discussion of ACK compression and its intrinsic consequences. 214 o Tests for a given subpath can be designed w/o knowing the rest of 215 the path. Tests suites can be designed for link types in the 216 abstract and standardized independently. Add example "complete 217 test suites" following the combined tests. 218 o Add add concept of untargeted tests and algebra on loss rate. 219 o Make the background traffic test have an explicit procedure, and 220 clearly delineate between users background traffic and other 221 traffic. Connect preloading to intermittent and not intermittent, 222 and as a way to control radio power states. Note that nearly all 223 devices have some preloading effects (e.g. ARP on LANs) 224 o Better uniformity about: applies to all transport protocols, but 225 defined in terms of TCP parameters. 226 o Clean and uniform descriptions of all tests. 227 o Write model appendix. Deprecate "old doc" appendix. 229 2. New requirements relative to RFC 2330 231 The Model Based Metrics are designed to fulfil some additional 232 requirement that were not recognized at the time RFC 2330 was 233 written. These missing requirements may have significantly 234 contributed to policy difficulties in the IP measurement space. The 235 additional requirements are: 236 o Metrics must be actionable by the ISP - they have to be 237 interpreted in terms of behaviors or properties at the IP or lower 238 layers, that an ISP can test, repair and verify. 239 o Metrics must be vantage point invariant over a significant range 240 of measurement point choices (e.g., measurement points as 241 described in [I-D.morton-ippm-lmap-path]), including off path 242 measurement points. The only requirements on MP selection should 243 be that the portion of the path that is not under test is 244 effectively ideal (or is non ideal in calibratable ways) and the 245 end-to-end RTT between MPs is below some reasonable bound. 246 o Metrics must be repeatable by multiple parties. It must be 247 possible for different parties to make the same measurement and 248 observe the same results. In particular it is specifically 249 important that both a consumer (or their delegate) and ISP be able 250 to perform the same measurement and get the same result. 252 NB: all of the requirements for metrics in RFC 2330 should be 253 reviewed and potentially revised. 255 3. Background 257 The holy grail of IPPM has been BTC measurement, but it has proven to 258 be a very hard problem for a number of reasons: 259 o TCP is a control system with circular dependencies - everything 260 affects performance, including components that are explicitly not 261 part of the test. 262 o Congestion control is an equilibrium process, transport protocols 263 change the network (raise loss probability and/or RTT) to conform 264 to their behavior. 265 o TCP's ability to compensate for network flaws is directly 266 proportional to the number of round trips per second (e.g. 267 inversely proportional to the RTT). As a consequence a flawed 268 link that passes a local test is likely to completely fail when 269 the path is extended by a perfect network to some larger RTT. 270 o TCP has a meta Heisenberg problem - Measurement and cross traffic 271 interact in unknown and ill defined ways. The situation is 272 actually worse than the traditional physics problem where you can 273 at least estimate the relative masses of the measurement and 274 measured particles. For network measurement you can not in 275 general determine the relative "masses" of the measurement traffic 276 and cross traffic, so you can not even gage the relative magnitude 277 of their effects on each other. 279 The new approach is to "open loop" mandatory congestion control 280 algorithms, typically by throttling TCP (or other protocol) to a 281 lower rate, such that it does not react to changing network 282 conditions. In this approack the measurement software explicitly 283 controls the data rate, transmission pattern or cwnd (TCP's primary 284 congestion control state variables) to create repeatable traffic 285 patterns that mimic TCP behavior but are almost entirely independent 286 of the actual network behavior. These patterns are manipulated to 287 probe the network to verify that it can deliver all of the traffic 288 patterns that a transport protocol is likely to generate under normal 289 operation at the target rate and RTT. 291 Models are used to determine the actual test parameters (burst size, 292 loss rate, etc) from the target parameters. The basic method is to 293 use models to estimate specific network properties required to 294 sustain a given transport flow (or set of flows), and using a suite 295 of simpler metrics to confirm that the network meets the required 296 properties. For example a network can sustain a Bulk TCP flow of a 297 given data rate, MTU and RTT when 4 (and probably more) conditions 298 are met: 299 o The raw link rate is higher than the target data rate. 300 o The raw packet loss rate is lower than required by a suitable TCP 301 performance model 302 o There is sufficient buffering at any bottleneck smooth bursts. 303 o When the link is overfilled (congested), the onset of packet loss 304 is progressive. 306 These condition can all be verified with simple tests, using model 307 parameters and acceptance thresholds derived from the target data 308 rate, MTU and RTT. Note that this procedure is not invertible: a 309 singleton measurement is a pass/fail evaluation of a given path or 310 subpath at a given performance. Measurements to confirm that a link 311 passes at one particular performance may not be generally be useful 312 to predict if the link will pass at a different performance. 314 Although they are not invertible, they do have several other valuable 315 properties, such as natural ways to define several different 316 composition metrics. 318 [Add text on algebra on metrics (A-Frame) and tomography.] The 319 Spatial Composition of fundamental IPPM metrics has been studied and 320 standardized. For example, the algebra to combine empirical 321 assessments of loss ratio to estimate complete path performance is 322 described in section 5.1.5. of [RFC6049]. We intend to use this and 323 other composition metrics as necessary. 325 4. Common Models and Parameters 327 Transport performance models are used to derive the test parameters 328 for each single property test from the end-to-end target parameters 329 and additional ancillary parameters. 331 It is envisioned that the modeling phase (to compute the test 332 parameters) and testing phases will be decoupled. This section 333 covers common derived parameters, used by multiple single property 334 tests. For some tests, additional modeling is described with the 335 tests. MAKE THIS NON SEQUENTIAL 337 Since some aspects of the models may be excessively conservative, the 338 modeling framework permits some latitude in derating some test 339 parameters, as described in Section 4.5. 341 For certain sub-paths (e.g. common types of access links) it would be 342 appropriate for the single property test parameters to be documented 343 as a "measurement profile" together with the modeling assumptions and 344 derating factors described in Section 4.4 and Section 4.5. 346 4.1. Target End-to-end parameters 348 These parameters are determined by the needs of the application or 349 the ultimate end user and the end-to-end Internet path. They are in 350 units that make sense to the upper layer: payload bytes delivered, 351 excluding header overheads for IP, TCP and other protocol. 353 Target Data Rate: The application or ultimate user's performance 354 goal. 355 Target RTT (Round Trip Time): For fundamental reasons a long path 356 makes it more difficult for TCP or other transport protocol to 357 meet the target rate. The target RTT must be representative of 358 the actual expected application use of the network. It may be 359 subject to conventions about assumed application usage (e.g. 360 continental scale paths should be assumed to be some fixed RTT, 361 such as 100 ms) or alternatively be an property of an ISP's 362 topology (e.g. a ISP with richer or better placed peering may be 363 able to justify assuming lower RTTs than other ISPs.) 364 Target MTU (Maximum Transmission Unit): Assume 1500 Bytes per packet 365 unless otherwise specified. If some sub-path forces a smaller 366 MTU, then all sub-paths must be tested with the same smaller MTU. 367 Header overhead The IP and TCP header sizes, which are the portion 368 of each MTU not available for carrying application payload. This 369 is also assumed to be the size for returning acknowledgements 370 (ACKs). The payload Maximum Segment Size (MSS) is the Target MTU 371 minus header overhead. 372 Permitted Number of Connections: The target rate can be more easily 373 obtained by dividing the traffic across more than one connection. 374 (Ideally this would be 1). In general the number of concurrent 375 connections is determined by the application, however see the 376 comments below on multiple connections. 377 Effective Bottleneck Data Rate This is the bottleneck data rate that 378 might be inferred from the ACK stream, by looking at how much data 379 the ACK stream reports was delivered per unit time. For 380 traditional networks, the effective bottleneck rate would be the 381 same as the actual bottleneck rate. If the forward path is 382 subject AFD [AFD] style policing using a virtual queue, the 383 effective bottleneck rate would be the same as the actual physical 384 link rate, even though this rate is not available to the user. 385 For systems that batch ACKs, for example due to half duplex 386 channel allocation, the effective bottleneck data rate might be 387 much higher than any link in the system. 389 The use of multiple connections has been very controversial since the 390 beginning of the World-Wide-Web[first complaint]. Modern browsers 391 open many connections [BScope]. Experts associated with IETF 392 transport area have frequently spoken against this practice [long 393 list]. It is not inappropriate to assume some small number of 394 concurrent connections (e.g. 4 or 6), to compensate for limitation in 395 TCP. However, choosing too large a number is at risk of being 396 interpreted as a signal by the web browser community that this 397 practice has been embraced by the Internet service provider 398 community. It may not be desirable to send such a signal. 400 4.2. End-to-end parameters from sub-paths 402 [This entire section needs to be overhauled and should be skipped on 403 a first reading. The concepts defined here are not used elsewhere.] 405 The following optional parameters apply for testing generalized end- 406 to-end paths that include subpaths with known specific types of 407 behaviors that are not well represented by simple queueing models: 408 Bottleneck link clock rate: This applies to links that are using 409 virtual queues or other techniques to police or shape users 410 traffic at lower rates full link rate. The bottleneck link clock 411 rate should be representative of queue drain times for short 412 bursts of packets on an otherwise unloaded link. 413 Channel hold time: For channels that have relatively expensive 414 channel arbitration algorithms, this is the typical (maximum?) 415 time that data and or ACKs are held pending acquiring the channel. 416 While under heavy load, the RTT may be inflated by this parameter, 417 unless it is built into the target RTT 418 Preload traffic volume: If the user's traffic is shaped on the basis 419 of average traffic volume, this is volume necessary to invoke 420 "heavy hitter" policies. 421 Unloaded traffic volume: If the user's traffic is shaped on the 422 basis of average traffic volume, this is the maximum traffic 423 volume that a test can use and stay within a "light user" 424 policies. 426 Note on a ConEx enabled network [ConEx], the word "traffic" in the 427 last two items should be replaced by "congestion" i.e. "preload 428 congestion volume" and "unloaded congestion volume". 430 4.3. Per sub-path parameters 432 [This entire section needs to be overhauled and should be skipped on 433 a first reading. The concepts defined here are not used elsewhere.] 435 Some single parameter tests also need parameter of the sub-path. 437 sub-path RTT: RTT of the sub-path under test. 438 sub-path link clock rate: If different than the Bottleneck link 439 clock rate 441 4.4. Common Calculations for Single Property Tests 443 The most important derived parameter is target_pipe_size (in 444 packets), which is the number of packets needed exactly meet the 445 target rate, with no cross traffic for the specified target RTT and 446 MTU. It is given by: 448 [Need to add multiple connections] 450 target_pipe_size = target_rate * target_RTT / ( target_MTU - 451 header_overhead ) 453 If the transport protocol (e.g. TCP) average window size is smaller 454 than this, it will not meet the target rate. 456 The reference_target_run_length, which is the most conservative model 457 for the minimum spacing between losses, can derived as follows: 458 assume the target_data_rate is equal to bottleneck link_data_rate. 459 Then target_pipe_size also predicts the onset of queueing. If the 460 transport protocol (e.g. TCP) has an average window size that is 461 larger than the target_pipe_size, the excess packets will form a 462 standing queue at the bottleneck. 464 If the transport protocol is using traditional Reno style Additive 465 Increase, Multiplicative Decrease congestion control [RFC5681], then 466 there must be target_pipe_size roundtrips between losses. Otherwise 467 the multiplicative window reduction triggered by a loss would cause 468 the network to be underfilled. Following [MSMO97], we derive the 469 losses must be no more frequent than every 1 in 470 (3/2)(target_pipe_size^2) packets. This provides the reference value 471 for target_run_length which is typically the number of packets that 472 must be delivered between loss episodes in the tests below: 474 reference_target_run_length = (3/2)(target_pipe_size^2) 476 Note that this calculation is based on a number of assumptions that 477 may not apply. Appendix A discusses these assumptions and provides 478 some alternative models. The actual method for computing 479 target_run_length MUST be documented along with the rationale for the 480 underlying assumptions and the ratio of chosen target_run_length to 481 reference_target_run_length. 483 Although this document gives a lot of latitude for calculating 484 target_run_length, people specifying profiles for suites of single 485 property tests need to consider the effect of their choices on the 486 ongoing conversation and tussle about the relevance of "TCP 487 friendliness" as an appropriate model for capacity allocation. 488 Choosing a target_run_length that is substantially smaller than 489 reference_target_run_length is equivalent to saying that it is 490 appropriate for the transport research community to abandon "TCP 491 friendliness" as a fairness model and to develop more aggressive 492 Internet transport protocols, and for applications to continue (or 493 even increase) the number of connections that they open concurrently. 495 The calculations for individual parameters are presented with the 496 each single property test. In general these calculations are 497 permitted some derating as described in Section 4.5 499 4.5. Parameter Derating 501 Since some aspects of the models are very conservative, the modeling 502 framework permits some latitude in derating some specific test 503 parameters, as indicated in Section 6. For example classical 504 performance models suggest that in order to be sure that a single TCP 505 stream can fill a link, it needs to have a full bandwidth-delay- 506 product worth of buffering at the bottleneck[QueueSize]. In real 507 networks with real applications this is often overly conservative. 508 Rather than trying to formalize more complicated models we permit 509 some test parameters to be relaxed as long as they meet some 510 additional procedural constraints: 511 o The method used compute and justify the derated metrics is 512 published in such a way that it becomes a matter of public record. 513 o The calibration procedures described in Section 8 are used to 514 demonstrate the feasibility of meeting the performance targets 515 with the derated test parameters. 516 o The calibration process itself is documented is such a way that 517 other researchers can duplicate the experiments and validate the 518 results. 520 In the test specifications in Section 6 assume 0 < derate <= 1, is a 521 derating parameter. These will be individually named in the final 522 document. In all cases making derate smaller makes the test more 523 tolerant. Derate = 1 is "full strenght". 525 Note that some single property test parameters are not permitted to 526 be derated. 528 5. Common testing procedures 530 5.1. Traffic generating techniques 532 A key property of Model Based Metrics is that the traffic patterns 533 are determined by the end-to-end target parameters and not the 534 network. The only exception are the constant window tests, which 535 rely on a TCP style self clock. This makes the tests "open loop", 536 which is key to preventing circular dependencies between test 537 prameters. The transmission pattern does not depend on the details 538 of the network's reaction to traffic. 540 5.1.1. Paced transmission 542 Paced (burst) transmissions: send bursts of data on a timer to meet a 543 particular target rate and pattern. 544 Single: Send individual packets at the specified rate or headway. 545 Burst: Send server interface rate bursts on a timer. Specify any 3 546 of average rate, packet size, burst size (number of packets) and 547 burst headway (start to start). These bursts are typically sent 548 as back-to-back packets on high speed media. 549 Slow start: Send 4 packets bursts at an average rate equal to twice 550 the effective bottleneck link rate (but not faster than the server 551 interface rate). This corresponds to the average rate during a 552 TCP slowstart when Appropriate Byte Counting [ABC] is present or 553 delayed ack is disabled. Note that slow start pacing itself is 554 typically part of larger scale burst patterns, such as sending 555 target_pipe_size packets as slow start bursts every on a 556 target_RTT headway (burst start to burst start). Such a stream 557 has three different average rates, depending on the averaging time 558 scale. At the finest time scale the average rate is the same as 559 the server rate, at a medium scale the average rate is twice the 560 bottleneck link rate and at the longest time scales the average 561 rate is the target data rate, adjusted to include header overhead. 563 Note that if the effective bottleneck link rate is more than half of 564 the server interface rate, slowstart bursts become server interface 565 rate bursts. 567 5.1.2. Constant window pseudo CBR 569 Implement pseudo CBR by running a standard protocol such as TCP at a 570 fixed window size. This has the advantage that it can be implemented 571 under real content delivery. The rate is only maintained in average 572 over each RTT, and is subject to limitations of the transport 573 protocol. 575 For tests that have strongly prescribed data rates, if the transport 576 protocol fails to maintain the test rate for any reason (especially 577 due to network congestion) the test should be considered 578 inconclusive, otherwise there are some cases where tester failures 579 might cause incorrect link tests results. 581 5.1.3. Scanned window pseudo CBR 583 Same as the above, except the window is incremented once per 584 2*target_RTT, starting from below target_pipe and sweeping up to 585 first loss or some other event. This is analogous to the tests 586 implemented in Windowed Ping [WPING] and pathdiag [PATHDIAG]. 588 5.1.4. Intermittent Testing 590 Any test which does not depend on queueing (e.g. the CBR tests) or 591 normally experiences periodic zero outstanding data (e.g. between 592 bursts for burst tests), can be formulated as an intermittent test. 594 The Intermittent testing can be used for ongoing monitoring for 595 changes in sub-path quality with minimal disruption users. It should 596 be used in conjunction with the full rate test because this method 597 assesses an average_run_length over a long time interval w.r.t. user 598 sessions. It may false fail due to other legitimate congestion 599 causing traffic or may false pass changes in underlying link 600 properties (e.g. a modem retraining to an out of contract lower 601 rate). 603 5.1.5. Intermittent Scatter Testing 605 Intermittent scatter testing: when testing the network path to or 606 from an ISP subscriber aggregation point (Cable headend [better 607 word?] or DSLAM, etc), intermittent tests can be spread across a pool 608 of (idle) users such that no one users experiences the full impact of 609 the testing, even though the traffic to or from the ISP subscriber 610 aggregation point is sustained at full rate. EXPAND 612 5.2. Interpreting the Results 614 (This section needs major reorganization.) 616 MOVE ELSEWHERE: General comment about types of loss: masking BER type 617 loss with ARQ/FEC is not an issue. What we are most concerned with 618 is congestion or AQM related losses. These are explicitly considered 619 to be a signal back to the sender to slow down. Note that even with 620 ARQ or FEC at some point the link will accumulate enough backlog 621 where it will need to cause AQM (or drop tail overflow) losses. 623 5.2.1. Inconclusive test outcomes 625 A singleton is a pass fail measurement. 627 In addition we use "inconclusive" outcome to indicate that a test 628 failed to attain the required test conditions. This is important to 629 the extent that the tests themselves have built in control systems 630 which might interfere with some aspect of the test. 632 For example if a test is implemented with an controled and 633 instrumented TCP, failing to attain the specified data rate may 634 indicate a problem with either the TCP implementation or the test 635 vantage point. 637 One of the goal of the testing process should be to drive the number 638 of inconclusive tests to zero. 640 5.2.2. Statistical criteria for measuring run_length 642 When evaluating the traget_run_length, we need to determine 643 appropriate packet stream sizes and acceptable error levels to test 644 efficiently. In practice, can we compare the empirically estimated 645 loss probabilities with the targets as the sample size grows? How 646 large a sample is needed to say that the measurements of packet 647 transfer indicate a particular run-length is present? 649 The generalized measurement can be described as recursive testing: 650 send a flight of packets and observe the packet transfer performance 651 (loss ratio or other metric, any defect we define). 653 As each flight is sent and measured, we have an ongoing estimate of 654 the performance in terms of defect to total packet ratio (or an 655 empirical probability). Continue to send until conditions support a 656 conclusion or a maximum sending limit has been reached. 658 We have a target_defect_probability, 1 defect per target_run_length, 659 where a "defect" is defined as a lost packet, a packet with ECN mark, 660 or other impairment. This constitutes the null Hypothesis: 662 H0: no more than one defects in target_run_length = (3/2)*(flight)^2 663 packets 665 and we can stop sending flights of packets if measurements support 666 accepting H0 with the specified Type I error = alpha (= 0.05 for 667 example). 669 We also have an alternative Hypothesis to evaluate: if performance is 670 significantly lower than the target_defect_probability, say half the 671 target: 673 H1: one or more defects in target_run_length/2 packets 675 and we can stop sending flights of packets if measurements support 676 rejecting H0 with the specified Type II error = beta, thus preferring 677 the alternate H1. 679 H0 and H1 constitute the Success and Failure outcomes described 680 elsewhere in the memo, and while the ongoing measurements do not 681 support either hypothesis the current status of measurements is 682 indeterminate. 684 The problem above is formulated to match the Sequential Probability 685 Ratio Test (SPRT) [StatQC] [temp ref: 686 http://en.wikipedia.org/wiki/Sequential_probability_ratio_test ], 687 which also starts with a pair of hypothesis specified as above: 689 H0: p = p0 = one defect in target_run_length 690 H1: p = p1 = one defect in target_run_length/2 691 As flights are sent and measurements collected, the tester evaluates 692 the cumulative log-likelihood ratio: 694 S_i = S_i-1 + log(Lambda_i) 696 where Lambda_i is the ratio of the two likelihood functions 697 (calculated on the measurement at packet i, and index i increases 698 linearly over all flights of packets ) for p0 and p1 [temp ref: 699 http://en.wikipedia.org/wiki/Likelihood_function ]. 701 The SPRT specifies simple stopping rules: 703 o a < S_i < b: continue testing 704 o S_i <= a: Accept H0 705 o S_i >= b: Accept H1 706 where a and b are based on the Type I and II errors, alpha and beta: 708 a ~= Log((beta/(1-alpha)) and b ~= Log((1-beta)/alpha) 710 with the error probabilities decided beforehand, as above. 712 The calculations above are implemented in the R-tool for Statistical 713 Analysis, in the add-on package for Cross-Validation via Sequential 714 Testing (CVST) [http://www.r-project.org/] [Rtool] [CVST] . 716 5.2.3. Classifications of tests 718 These tests are annotated with "(load)", "(engineering)" or 719 "(monitoring)". WHY DO WE CARE? 721 A network would be expected to fail load tests in the presence excess 722 or uncontrolled cross traffic. As such, load tests identify 723 parameters that can transition from passing to failing as a 724 consequence of insufficient network capacity and the actions of other 725 network users. 727 Monitoring tests are design to capture the most important aspects of 728 a load test, but without causing unreasonable ongoing load 729 themselves. As such they may miss some details of the network 730 performance, but can serve as a useful reduced cost proxy for a load 731 test. 733 Engineering tests evaluate how network algorithms (such as AQM and 734 channel allocation) interact with transport protocols. Although 735 these tests may be to be sensitive to load, the interaction may be 736 quite complicated and might even have an inverted sensitivity. For 737 example a test to verify that an AQM algorithm causes ECN marks or 738 packet drops early enough to limit queue occupancy may experience a 739 false pass results in the presence of bursty cross traffic. It is 740 important that engineering tests be performed under a wide range of 741 conditions, including both in situ and bench testing, and under a 742 full range of load conditions. Ongoing monitoring is less likely to 743 be useful for these tests, although sparse testing might be 744 appropriate. 746 5.3. Reordering Tolerance 748 All tests must be instrumented for excessive reordering [RFC4737]. 750 NB: there is no global consensus for how much reordering tolerance is 751 appropriate or reasonable. ("None" is absolutely unreasonable.) 753 Section 5 of [RFC4737] proposed a metric that may be sufficient to 754 designate isolated reordered packets as effectively lost, because 755 TCP's retransmission response would be the same. 757 [As a strawman, we propose the following:] TCP should be able to 758 adapt to reordering as long as the reordering extent is no more than 759 the maximum of one half window or 1 mS, whichever is larger. Note 760 that there is a fundamental tradeoff between tolerance to reordering 761 and how quickly algorithms such as fast retransmit can repair losses. 762 Within this limit on extent, there should be no bound on the 763 frequency. 765 Parameters: 766 dispalcement the maximum of one half of target_pipe_size or 1 mS. 768 5.4. Verify the absence of cross traffic 770 Cross traffic should be monitored prior to and during testing. In 771 sub-paths where traffic of many users is aggregated, an excessive 772 level of cross traffic should be noted and prevent testing (the test 773 should be recorded as "inconclusive"). In sub-paths that include 774 individual subscriber service, the current approach is to suspend 775 testing when subscriber traffic is detected, because neither a flawed 776 test nor degraded user performance conditions are desired. 778 Note that canceling tests due to load on subscriber lines may 779 introduce sampling errors for other parts of the infrastructure. For 780 this reason tests that are scheduled but not run due to load should 781 be treated as a special case of "inconclusive". 783 The test is deemed to have passed only if the observed data rate 784 matches the target_data_rate and it is statistically significant that 785 the average_run_lenght is larger than target_run_lenght. It is 786 deemed inconclusive if: the statistical test is inconclusive; there 787 is too much background load; or the target_data_rate could not be 788 attained. 790 Use a passive packet or SNMP monitoring to verify that the traffic 791 volume on the sub-path agrees with the traffic generated by a test. 792 Ideally this should be performed before during and after each test. 794 The goal is provide quality assurance on the overall measurement 795 process, and specifically to detect the following measurement 796 failure: a user observes unexpectedly poor application performance, 797 the ISP observes that the access link is running at the rated 798 capacity. Both fail to observe that the user's computer has been 799 infected by a virus which is spewing traffic as fast as it can. 801 Parameters: 802 Maximum Cross Traffic Data Rate The amount of excess traffic 803 permitted. Note that this might be different for different tests. 804 Maximum Data Rate underage The permitted amount that the traffic can 805 be less than predicted for the current test. Normally this would 806 just be a statement of the maximum permitted measurement error, 807 however it might also detect cases where the passive and active 808 tests are misaligned: testing different subscriber lines. This is 809 important because the vantage points are so different: in-band 810 active measurement vs out-of-band passive measurement. 812 One possible method is an adaptation of: www-didc.lbl.gov/papers/ 813 SCNM-PAM03.pdf D Agarwal etal. "An Infrastructure for Passive 814 Network Monitoring of Application Data Streams". Use the same 815 technique as that paper to trigger the capture of SNMP statistics for 816 the link. 818 5.5. Additional test preconditions 820 Send pre-load traffic as needed to activate radios with a sleep mode, 821 or other "reactive network" elements (term defined in 822 [draft-morton-ippm-2330-update-01]). 824 Use the procedure above to confirm that the pre-test background 825 traffic is low enough. 827 6. Single Property Tests 829 The following tests are fully decomposed to verify individual network 830 properties required for TCP meet the target parameters. It is 831 believed that these properties apply to all self clocked throughput 832 maximizing protocols. Failing to meet any one of these tests will 833 cause poor TCP performance in some specific context. 835 These tests are pedantically separated: It would be more practical to 836 combine them. Failing such a combined test might imply ambiguous 837 consequences for TCP: it would be expected to fail under some 838 conditions, but a single test might not be able to indicate exactly 839 which conditions. The following section describes some combined 840 tests. 842 The single property tests confirm that each sub-path can sustain the 843 normal traffic patterns caused by TCP running at the specified target 844 performance. Specifically they confirm that each sub-path has: 845 sufficient raw capacity (e.g. sufficient data rate); low enough 846 background loss rate where mandatory congestion control stays out of 847 the way; large enough queue space to absorb TCP's normal bursts; does 848 not cause unreasonable packet reordering; progressive AQM to 849 appropriately invoke congestion control. Appropriately invoking 850 congestion control requires that packet losses or ECN marks start 851 progressively before TCP creates an excessive sustained 852 queues[BufferBloat] or and without causing excessively bursty losses. 853 The return path must also subject to a similar suite of tests, 854 although potentially with different test parameters (due to the 855 asymmetrical capabilities of many access link technologies). 856 Furthermore it is important that the forward and return path interact 857 appropriately, for example if they share they share a channel that 858 has to be allocated. 860 Note that many of the sub-path tests resemble metrics that have 861 already been defined in the IPPM context, with the specification of 862 additional of test parameters and success critera. The models used 863 to derive the test parameters make specific assumptions about network 864 conditions, a test is deemed "inconclusive" (as opposed to failing) 865 if tester does not meet the underlying assumption. For example a 866 loss rate test at a specified data rate is inconclusive if the tester 867 fails to send data at the specified rate for some reason. This 868 concept of an inconclusive test is necessary to build tests out of 869 protocols or technologies that they themselves have built in or 870 implicit control systems. 872 6.1. CBR Tests 874 We propose several versions of the CBR loss rate test. One, 875 performed at data full rate, is intrusive and recommend for 876 infrequent testing, such as when a service is first turned up or as 877 part of an auditing process. The second, background loss rate, is 878 designed for ongoing monitoring for change is sub-path quality. 880 6.1.1. Loss Rate at Full Data Rate 882 Send at target_rate, confirm that the observed run length is at least 883 the target_run_lenght. No additional derating is permitted (except 884 through the choice of model to calculate target_run_lenght in the 885 first place). 887 Note that this test also implicitly confirms that sub_path has 888 sufficient capacity to carry the target_data_rate. 890 Parameters: 891 Run Length Same as target_run_lenght 892 Data Rate Same as target_data_rate 894 Note that these parameters MUST NOT be derated. If the default 895 parameters are too stringent use an alternate model for 896 target_run_lenght as described in Appendix A. 898 Data is sent at the specified data_rate. The receiver accumulates 899 the total data delivered and packets lost [and ECN marks, which are 900 nominally treated as losses by conforming transport protocols]. The 901 observed average_run_lenght is computed from total_data_delivered 902 divided by the total_loss_rate. A [TBD] statistical test is applied 903 to determine when or if the average_run_length is larger than 904 target_run_length. 906 TODO: add language about monitoring cross traffic. acm attempted 907 below. 909 6.1.2. Background Loss Rate Tests 911 The background loss rate is a low rate version of the target rate 912 test above, designed for ongoing monitoring for changes in sub-path 913 quality without disrupting users. It should be used in conjunction 914 with the above full rate test because it may be subject to false 915 results under some conditions, in particular it may false pass 916 changes in underlying link properties (e.g. a modem retraining to an 917 out of contract lower rate). 919 Parameters: 921 Run Length Same as target_run_lenght 922 Data Rate Some small fraction of target_data_rate, such as 1%. 924 Once the preconditions described in Section 5.5 are met, the test 925 data is sent at the prescribed rate with a burst size of 1. The 926 receiver accumulates packet delivery statistics and the procedures 927 described in Section 5.2.1 and Section 5.4 are used to score the 928 outcome: 930 Pass: it is statistically significantly that the observed run length 931 is larger than the target_run_length. 933 Fail: it is statistically significantly that the observed run length 934 is smaller than the target_run_length. 936 Inconclusive: Neither test was statistically significant or there was 937 excess cross traffic during the test. 939 6.2. Standing Queue tests 941 These test confirm that the bottleneck is well behaved across the 942 onset of queueing. For conventional bottlenecks this will be from 943 the onset of queuing to the point where there is a full target_pipe 944 of standing data. Well behaved generally means lossless for 945 target_run_length, followed by a small number of losses to signal to 946 the transport protocol that it should slow down. Losses that are too 947 early can prevent the transport from averaging above the target_rate. 948 Losses that are too late indicate that the queue might be subject to 949 bufferbloat and subject other flows to excess queuing delay. Excess 950 losses (more than half of of target_pipe) make loss recovery 951 problematic for the transport protcol. 953 These tests can also observe some problems with channel acquisition 954 systems, especially at the onset of persistent queueing. Details 955 TBD. 957 6.2.1. Congestion Avoidance 959 Use the procedure in Section 5.1.3 to sweep the window (rate) from 960 below link_pipe up to beyond target_pipe+link_pipe. Depending on 961 events that happen during the scan, score the link. Identify the 962 power_point=MAX(rate/RTT) as the start of the test. 964 Fail if first loss is too early (loss rate too high) on repeated 965 tests or if the losses are more than half of the outstanding data. (a 966 load test) 968 6.2.2. Buffer Bloat 970 Use the procedure in Section 5.1.3 to sweep the window (rate) from 971 below link_pipe up to beyond target_pipe+link_pipe. Depending on 972 events that happen during the scan, score the link. Identify the 973 "power point:MAX(rate/RTT) as the start of the test (should be 974 window=target_pipe) 976 Fail if first loss is too late (insufficient AQM and subject to 977 bufferbloat - an engineering test). NO THEORY 979 6.2.3. Self Interference 981 Use the procedure in Section 5.1.3 to sweep the window (rate) from 982 below link_pipe up to beyond target_pipe+link_pipe. Depending on 983 events that happen during the scan, score the link. Identify the 984 "power point:MAX(rate/RTT) as the start of the test (should be 985 window=target_pipe) 987 Fail if RTT is non-monotonic by more than a small number of packet 988 times (channel allocation self interference - engineering) IS THIS 989 SUFFICIENT? 991 6.3. Slow Start tests 993 These tests mimic slow start: data is sent at twice subpath_rate. 994 They are deemed inconclusive if the elapsed time to send the data 995 burst is not less than half of the (extrapolated) time to receive the 996 ACKs. (i.e. sending data too fast is ok, but sending it slower than 997 twice the actual bottleneck rate is deemed inconclusive). Space the 998 bursts such that the average ACK rate matches the target_data_rate. 1000 These tests are not useful at burst sizes smaller than the server 1001 rate tests, since the server rate tests are more strenuous. If it is 1002 necessary to derate the server rate tests, then the full window 1003 slowstart test (un-derated) would be important. 1005 6.3.1. Full Window slow start test 1007 Send target_pipe_size*derate bursts must have fewer than one loss per 1008 target_run_length*derate. Note that these are the same parameters as 1009 the Server Full Window test, except the burst rate is at slowestart 1010 rate, rather than server interface rate. SHOULD derate=1. 1012 Otherwise TCP will exit from slowstart prematurely, and only reach a 1013 full target_pipe_size window by way of congestion avoidance. 1015 This is a load test: cross traffic may cause premature losses. 1017 6.3.2. Slowstart AQM test 1019 Do a continuous slowstart (date rate = 2*subpath_rate), until first 1020 loss, and repeat, gathering statistics on the last delivered packet's 1021 RTT and window size. Fail if too large (NO THEORY for value). 1023 This is an engineering test: It would be best performed on a 1024 quiescent network or testbed, since cross traffic might cause a false 1025 pass. 1027 6.4. Server Rate tests 1029 These tests us "server interface rate" bursts. Although this is not 1030 well defined it should be assumed to be current state of the art 1031 server grade hardware (probably 10Gb/s today). (load) 1033 6.4.1. Server TCP Send Offload (TSO) tests 1035 If MIN(target_pipe_size, 42) packet bursts meet target_run_lenght 1036 (Not derated!). 1038 Otherwise the link will interact badly with modern server NIC 1039 implementations, which as an optimization to reduce host side 1040 interactions (interrupts etc) accept up to 64kB super packets and 1041 send them as 42 seperate packets on the wire side.cc (load) 1043 6.4.2. Server Full Window test 1045 target_pipe_size*derate bursts have fewer than one loss per 1046 target_run_length*derate. 1048 Otherwise application pauses will cause unwarranted losses. Current 1049 standards permit TCP to send a full cwnd burst following an 1050 application pause. (Cwnd validation in not required, but even so 1051 does not take effect until the pause is longer than RTO). 1053 NB: there is no model here for what is good enough. derate=1 is 1054 safest, but may be unnecessarily conservative for some applications. 1055 Some application, such as streaming video need derate=1 to be 1056 efficient when the application pacing quanta is larger than cwnd. 1057 (load) 1059 7. Combined Tests 1061 These tests are more efficient from a deployment/operational 1062 perspective, but may not be possible to diagnose if they fail. 1064 7.1. Sustained burst test 1066 Send target_pipe_size server rate bursts every target_RTT, verify 1067 that the observed run length meets target_run_length. Key 1068 observations: 1069 o This test is RTT invariant, as long as the tester can generate the 1070 required pattern. 1071 o The subpath under test is expected to go idle for some fraction of 1072 the time: (link_rate-target_rate)/link_rate. Failing to do so 1073 suggests a problem with the procedure. 1074 o This test is more strenuous than the slow start tests: they are 1075 not needed if the link passes underated server rate burst tests. 1076 o This test could be derated by reducing both the burst size and 1077 headway (same average data rate). 1078 o A link that passes this test is likely to be able to sustain 1079 higher rates (close to link_rate) for paths with RTTs smaller than 1080 the target_RTT. Offsetting this performance underestimation is 1081 the rationale behind permitting derating in general. 1082 o This test should be implementable with standard instrumented TCP, 1083 [RFC 4898] using a specialized measurement application at one end 1084 and a minimal service at the other end [RFC 863, RFC 864]. It may 1085 require tweaks to the TCP implementation. 1086 o This test is efficient to implement, since it does not require 1087 per-packet timers, and can make maximal use of TSO in modern NIC 1088 hardware. 1089 o This test is not totally sufficient: the standing window 1090 engineering tests are also needed to be sure that the link is well 1091 behaved at and beyond the onset of congestion. 1092 o I believe that this test can be proven to be the one load test to 1093 supplant them all. 1095 Example 1097 To confirm that a 100 Mb/s link can reliably deliver single 10 1098 MByte/s stream at a distance of 50 mS, test the link by sending 346 1099 packet bursts every 50 mS (10 MByte/s payload rate, assuming a 1500 1100 Byte IP MTU and 52 Byte TCP/IP headers). These bursts are 4196288 1101 bits on the wire (assuming 16 bytes of link overhead and framing) for 1102 an aggregate test data rate of 8.4 Mb/s. 1104 To pass the test using the most conservative TCP model for a single 1105 stream the observed run length must be larger than 179574 packets. 1106 This is the same as less than one loss per 519 bursts (1.5*346) or 1107 every 26 seconds. 1109 Note that this test potentially cause transient 346 packet queues at 1110 the bottleneck. 1112 8. Calibration 1114 If using derated metrics, or when something goes wrong, the results 1115 must be calibrated against a traditional BTC. The preferred 1116 diagnostic follow-up to calibration issues is to run open end-to-end 1117 measurements on an open platform, such as Measurement Lab 1118 [http://www.measurementlab.net/] 1120 9. Acknowledgements 1122 Ganga Maguluri suggested the statistical test for measuring loss 1123 probability in the target run length. 1125 Meredith Whittaker for improving the clarity of the communications. 1127 10. References 1129 10.1. Normative References 1131 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 1132 Requirement Levels", BCP 14, RFC 2119, March 1997. 1134 [RFC2026] Bradner, S., "The Internet Standards Process -- Revision 1135 3", BCP 9, RFC 2026, October 1996. 1137 10.2. Informative References 1139 [RFC2330] Paxson, V., Almes, G., Mahdavi, J., and M. Mathis, 1140 "Framework for IP Performance Metrics", RFC 2330, 1141 May 1998. 1143 [RFC4737] Morton, A., Ciavattone, L., Ramachandran, G., Shalunov, 1144 S., and J. Perser, "Packet Reordering Metrics", RFC 4737, 1145 November 2006. 1147 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion 1148 Control", RFC 5681, September 2009. 1150 [RFC5835] Morton, A. and S. Van den Berghe, "Framework for Metric 1151 Composition", RFC 5835, April 2010. 1153 [RFC6049] Morton, A. and E. Stephan, "Spatial Composition of 1154 Metrics", RFC 6049, January 2011. 1156 [MSMO97] Mathis, M., Semke, J., Mahdavi, J., and T. Ott, "The 1157 Macroscopic Behavior of the TCP Congestion Avoidance 1158 Algorithm", Computer Communications Review volume 27, 1159 number3, July 1997. 1161 [BScope] Broswerscope, "Browserscope Network tests", Sept 2012, 1162 . 1164 See Max Connections column 1166 [I-D.morton-ippm-lmap-path] 1167 Bagnulo, M., Burbridge, T., Crawford, S., Eardley, P., and 1168 A. Morton, "A Reference Path and Measurement Points for 1169 LMAP", draft-morton-ippm-lmap-path-00 (work in progress), 1170 January 2013. 1172 [Rtool] R Development Core Team, "R: A language and environment 1173 for statistical computing. R Foundation for Statistical 1174 Computing, Vienna, Austria. ISBN 3-900051-07-0, URL 1175 http://www.R-project.org/", , 2011. 1177 [StatQC] Montgomery, D., "Introduction to Statistical Quality 1178 Control - 2nd ed.", ISBN 0-471-51988-X, 1990. 1180 [CVST] Krueger, T. and M. Braun, "R package: Fast Cross- 1181 Validation via Sequential Testing", version 0.1, 11 2012. 1183 Appendix A. Model Derivations 1185 This appendix describes several different ways to calculate 1186 target_run_length and the implication of the chosen calculation. 1188 Rederive MSMO97 under two different assumptions: target_rate = 1189 link_rate and target_rate < 2 * link_rate. 1191 Show equivalent derivation for CUBIC. 1193 Commentary on the consequence of the choice. 1195 Appendix B. Old text from an earlier document 1197 To be moved, removed or absorbed 1199 Step 0: select target end-to-end parameters: a target rate and target 1200 RTT. The primary test will be to confirm that the link quality is 1201 sufficient to meet the specified target rate for the link under test, 1202 when extended to the target RTT by an ideal network. The target rate 1203 must be below the actual link rate and nominally the target RTT would 1204 be longer than the link RTT. There should probably be a convention 1205 for the relationship between link and target rates (e.g. 85%). 1207 For example on a 10 Mb/s link, the target rate might be 1 MBytes/s, 1208 at an RTT of 100 mS (a typical continental scale path). 1210 Step 1: On the basis of the target rate and RTT and your favorite TCP 1211 performance model, compute the "required run length", which is the 1212 required number of consecutive non-losses between loss episodes. The 1213 run length resembles one over the loss probability, if clustered 1214 losses only count as a single event. Also select "test duration" and 1215 "test rate". The latter would nominally the same as the target rate, 1216 but might be different in some situations. There must be 1217 documentation connecting the test rate, duration and required run 1218 length, to the target rate and RTT selected in step 0. 1220 Continuing the above example: Assuming a 1500 Byte MTU. The 1221 calculated model loss rate for a single TCP stream is about 0.01% (1 1222 loss in 1E4 packets). 1224 Step 2, the actual measurement proceeds as follows: Start an 1225 unconstrained bulk data flow using any modern TCP (with large buffers 1226 and/or autotuning). During the first interval (no rate limits) 1227 observe the slowstart (e.g. tcpdump) and measure: Peak burst size; 1228 link clock rate (delivery rate for each round); peak data rate for 1229 the fastest single RTT interval; fraction of segments lost at the end 1230 of slow start. After the flow has fully recovered from the slowstart 1231 (details not important) throttle the flow down to the test rate (by 1232 clamping cwnd or application pacing at the sender or receiver). 1233 While clamped to the test rate, observe the losses (run length) for 1234 the chosen test duration. The link passes the test if the slowstart 1235 ends with less than approximately 50% losses and no timeouts, the 1236 peak rate is at least the target rate, and the measured run length is 1237 better than the required run length. There will also need to be some 1238 ancillary metrics, for example to discard tests where the receiver 1239 closes the window, invalidating the slowstart test. [This needs to 1240 be separated into multiple subtests] 1242 Optional step 3: In some cases it might make sense to compute an 1243 "extrapolated rate", which is the minimum of the observed peak rate, 1244 and the rate computed from the specified target RTT and the observed 1245 run length by using a suitable TCP performance model. The 1246 extrapolated rate should be annotated to indicate if it was run 1247 length or peak rate limited, since these have different predictive 1248 values. 1250 Other issues: 1252 If the link RTT is not substantially smaller than the target RTT and 1253 the actual run length is close to the target rate, a standards 1254 compliant TCP implementation might not be effective at accurately 1255 controlling the data rate. To be independent of the details of the 1256 TCP implementation, failing to control the rate has to be treated as 1257 a spoiled measurement, not a infrastructure failure. This can be 1258 overcome by "stiffening" TCP by using a non-standard congestion 1259 control algorithm. For example if the rate controlling by clamping 1260 cwnd then use "relentless TCP" style reductions on loss, and lock 1261 ssthresh to the cwnd clamp. Alternatively, implement an explicit 1262 rate controller for TCP. In either case the test must be abandoned 1263 (aborted) if the measured run length is substantially below the 1264 target run length. 1266 If the test is run "in situ" in a production environment, there also 1267 needs to be baseline tests using alternate paths to confirm that 1268 there are no bottlenecks or congested links between the test end 1269 points and the link under test. 1271 It might make sense to run multiple tests with different parameters, 1272 for example infrequent tests with test rate equal to the target rate, 1273 and more frequent, less disruptive tests with the same target rate 1274 but the test rate equal to 1% of the target rate. To observe the 1275 required run length, the low rate test would take 100 times longer to 1276 run. 1278 Returning to the example: a full rate test would entail sending 690 1279 pps (1 MByte/s) for several tens of seconds (e.g. 50k packets), and 1280 observing that the total loss rate is below 1:1e4. A less disruptive 1281 test might be to send at 6.9 pps for 100 times longer, and observing 1283 Formatted: Mon Feb 25 15:01:45 PST 2013 1285 Authors' Addresses 1287 Matt Mathis 1288 Google, Inc 1289 1600 Amphitheater Parkway 1290 Mountain View, California 93117 1291 USA 1293 Email: mattmathis@google.com 1294 Al Morton 1295 AT&T Labs 1296 200 Laurel Avenue South 1297 Middletown, NJ 07748 1298 USA 1300 Phone: +1 732 420 1571 1301 Email: acmorton@att.com 1302 URI: http://home.comcast.net/~acmacm/