idnits 2.17.1 draft-ietf-ippm-model-based-metrics-00.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. ** The document seems to lack a both a reference to RFC 2119 and the recommended RFC 2119 boilerplate, even if it appears to use RFC 2119 keywords. RFC 2119 keyword, line 560: '...arget_run_length MUST be documented al...' RFC 2119 keyword, line 958: '...d target_data_rate parameters MUST NOT...' RFC 2119 keyword, line 992: '...d target_data_rate parameters MUST NOT...' RFC 2119 keyword, line 1114: '... SHOULD derate=1....' Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year -- The document date (June 21, 2013) is 3961 days in the past. Is this intentional? Checking references for intended status: Experimental ---------------------------------------------------------------------------- == Missing Reference: 'Dominant' is mentioned on line 251, but not defined == Missing Reference: 'QueueSize' is mentioned on line 592, but not defined == Missing Reference: 'ABC' is mentioned on line 629, but not defined == Missing Reference: 'RFC 4898' is mentioned on line 1188, but not defined == Missing Reference: 'RFC 863' is mentioned on line 1189, but not defined == Missing Reference: 'RFC 864' is mentioned on line 1189, but not defined == Missing Reference: 'ConEx' is mentioned on line 1415, but not defined == Outdated reference: A later version (-01) exists of draft-morton-ippm-lmap-path-00 Summary: 3 errors (**), 0 flaws (~~), 10 warnings (==), 1 comment (--).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: December 23, 2013 AT&T Labs 6 June 21, 2013 8 Model Based Bulk Performance Metrics 9 draft-ietf-ippm-model-based-metrics-00.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 subpath at a time testing -- by applying a 16 suite of single property tests to successive subpaths 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 subpath 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 subpath 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 Formatted: Fri Jun 21 18:23:29 PDT 2013 38 Status of this Memo 40 This Internet-Draft is submitted in full conformance with the 41 provisions of BCP 78 and BCP 79. 43 Internet-Drafts are working documents of the Internet Engineering 44 Task Force (IETF). Note that other groups may also distribute 45 working documents as Internet-Drafts. The list of current Internet- 46 Drafts is at http://datatracker.ietf.org/drafts/current/. 48 Internet-Drafts are draft documents valid for a maximum of six months 49 and may be updated, replaced, or obsoleted by other documents at any 50 time. It is inappropriate to use Internet-Drafts as reference 51 material or to cite them other than as "work in progress." 53 This Internet-Draft will expire on December 23, 2013. 55 Copyright Notice 57 Copyright (c) 2013 IETF Trust and the persons identified as the 58 document authors. All rights reserved. 60 This document is subject to BCP 78 and the IETF Trust's Legal 61 Provisions Relating to IETF Documents 62 (http://trustee.ietf.org/license-info) in effect on the date of 63 publication of this document. Please review these documents 64 carefully, as they describe your rights and restrictions with respect 65 to this document. Code Components extracted from this document must 66 include Simplified BSD License text as described in Section 4.e of 67 the Trust Legal Provisions and are provided without warranty as 68 described in the Simplified BSD License. 70 Table of Contents 72 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 5 73 1.1. TODO . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 74 2. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 6 75 3. New requirements relative to RFC 2330 . . . . . . . . . . . . 8 76 4. Background . . . . . . . . . . . . . . . . . . . . . . . . . . 9 77 4.1. TCP properties . . . . . . . . . . . . . . . . . . . . . . 11 78 5. Common Models and Parameters . . . . . . . . . . . . . . . . . 12 79 5.1. Target End-to-end parameters . . . . . . . . . . . . . . . 13 80 5.2. Common Model Calculations . . . . . . . . . . . . . . . . 13 81 5.3. Parameter Derating . . . . . . . . . . . . . . . . . . . . 14 82 6. Common testing procedures . . . . . . . . . . . . . . . . . . 15 83 6.1. Traffic generating techniques . . . . . . . . . . . . . . 15 84 6.1.1. Paced transmission . . . . . . . . . . . . . . . . . . 15 85 6.1.2. Constant window pseudo CBR . . . . . . . . . . . . . . 16 86 6.1.2.1. Scanned window pseudo CBR . . . . . . . . . . . . 16 87 6.1.3. Intermittent Testing . . . . . . . . . . . . . . . . . 16 88 6.1.4. Intermittent Scatter Testing . . . . . . . . . . . . . 17 89 6.2. Interpreting the Results . . . . . . . . . . . . . . . . . 17 90 6.2.1. Test outcomes . . . . . . . . . . . . . . . . . . . . 17 91 6.2.2. Statistical criteria for measuring run_length . . . . 17 92 6.2.3. Classifications of tests . . . . . . . . . . . . . . . 19 93 6.2.4. Reordering Tolerance . . . . . . . . . . . . . . . . . 20 94 6.3. Test Qualifications . . . . . . . . . . . . . . . . . . . 20 95 6.3.1. Verify the Traffic Generation Accuracy . . . . . . . . 20 96 6.3.2. Verify the absence of cross traffic . . . . . . . . . 21 97 6.3.3. Additional test preconditions . . . . . . . . . . . . 22 98 7. Single Property Tests . . . . . . . . . . . . . . . . . . . . 22 99 7.1. Basic Data and Loss Rate Tests . . . . . . . . . . . . . . 22 100 7.1.1. Loss Rate at Paced Full Data Rate . . . . . . . . . . 22 101 7.1.2. Loss Rate at Full Data Windowed Rate . . . . . . . . . 23 102 7.1.3. Background Loss Rate Tests . . . . . . . . . . . . . . 23 103 7.2. Standing Queue tests . . . . . . . . . . . . . . . . . . . 24 104 7.2.1. Congestion Avoidance . . . . . . . . . . . . . . . . . 24 105 7.2.2. Buffer Bloat . . . . . . . . . . . . . . . . . . . . . 25 106 7.2.3. Duplex Self Interference . . . . . . . . . . . . . . . 25 107 7.3. Slowstart tests . . . . . . . . . . . . . . . . . . . . . 25 108 7.3.1. Full Window slowstart test . . . . . . . . . . . . . . 25 109 7.3.2. Slowstart AQM test . . . . . . . . . . . . . . . . . . 26 110 7.4. Sender Rate Burst tests . . . . . . . . . . . . . . . . . 26 111 7.4.1. Sender TCP Send Offload (TSO) tests . . . . . . . . . 26 112 7.4.2. Sender Full Window burst test . . . . . . . . . . . . 26 113 8. Combined Tests . . . . . . . . . . . . . . . . . . . . . . . . 27 114 8.1. Sustained burst test . . . . . . . . . . . . . . . . . . . 27 115 9. Calibration . . . . . . . . . . . . . . . . . . . . . . . . . 28 116 10. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . 28 117 11. Informative References . . . . . . . . . . . . . . . . . . . . 28 118 Appendix A. Model Derivations . . . . . . . . . . . . . . . . . . 29 119 Appendix B. old text . . . . . . . . . . . . . . . . . . . . . . 29 120 B.1. An earlier document . . . . . . . . . . . . . . . . . . . 30 121 B.2. End-to-end parameters from subpaths . . . . . . . . . . . 31 122 B.3. Per subpath parameters . . . . . . . . . . . . . . . . . . 32 123 B.4. Version Control . . . . . . . . . . . . . . . . . . . . . 32 124 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . . 32 126 1. Introduction 128 Model based bulk performance metrics evaluate an Internet paths 129 ability to carry bulk data. TCP models are used to design a targeted 130 diagnostic suite of IP performance tests which can be applied 131 independently to each subpath of the full end-to-end path. The 132 targeted diagnostic suites are constructed such that independent 133 tests of the subpaths will accurately predict if the full end-to-end 134 path can deliver bulk data at the specified performance target, 135 independent of the measurement vantage points or other details of the 136 test procedures used to measure each subpath. 138 Each test in the targeted diagnostic suite consists of a precomputed 139 traffic pattern and statistical criteria for evaluating packet 140 delivery. 142 TCP models are used to design traffic patterns that mimic TCP or 143 other bulk transport protocol operating at the target performance and 144 RTT over a full range of conditions, including flows that are bursty 145 at multiple time scales. The traffic patterns are computed in 146 advance based on the properties of the full end-to-end path and 147 independent of the properties of individual subpaths. As much as 148 possible the traffic is generated deterministically in ways that 149 minimizes the extent to which test methodology, measurement points, 150 measurement vantage or path partitioning effect the details of the 151 traffic. 153 Models are also used to compute the statistical criteria for 154 evaluating the IP diagnostics tests. The criteria for passing each 155 test must be determined from the end-to-end target performance and 156 independent of the RTT or other properties of the subpath under test. 157 In addition to passing or failing, a test can be inconclusive if the 158 precomputed traffic pattern was not authentically generated, test 159 preconditions were not met or the measurement results were not 160 statistically significantly. 162 TCP's ability to compensate for less than ideal network conditions is 163 fundamentally affected by the RTT and MTU of the end-to-end Internet 164 path that it traverses which are both fixed properties of the end-to- 165 end path. The target values for these three parameters, Data Rate, 166 RTT and MTU, are determined by the application, its intended use and 167 the physical infrastructure over which it traverses. They are used 168 to inform the models used to design the targeted diagnostic suite. 170 Section 2 defines terminology used throughout this document. It has 171 been difficult to develop BTC metrics due to some overlooked 172 requirements described in Section 3 and some intrinsic problems with 173 using protocols for measurement, described in Section 4. In 174 Section 5 we describe the models and common parameters used to derive 175 the targeted diagnostic suite. In Section 6 we describe common 176 testing procedures used by all of the tests. Each subpath is 177 evaluated using suite of far simpler and more predictable single 178 property tests described in Section 7. Section 8 describes some 179 combined tests that are more efficient to implement and deploy. 180 However, if they fail they may not clearly indicate the nature of the 181 problem. 183 There exists a small risk that model based metric itself might yield 184 a false pass result, in the sense that every subpath of an end-to-end 185 path passes every IP diagnostic test and yet a real application falls 186 to attain the performance target over the end-to-end path. If this 187 happens, then the calibration procedure described in Section 9 needs 188 to be used to validate and potentially revise the models. 190 Future document will define model based metrics for other traffic 191 classes and application types, such as real time. 193 1.1. TODO 195 Please send comments on this draft to ippm@ietf.org. See 196 http://goo.gl/02tkD for more information including: interim drafts, 197 an up to date todo list and information on contributing. 199 Formatted: Fri Jun 21 18:23:29 PDT 2013 201 2. Terminology 203 Properties determined by the end-to-end path and application. They 204 are described in more detail in Section 5.1. 206 end-to-end target parameters: Application or transport performance 207 goals for the end-to-end path. They include the target data rate, 208 RTT and MTU described below. 209 Target Data Rate: The application or ultimate user's performance 210 goal. This must be slightly smaller than the actual link rate, 211 otherwise there is no margin for compensating for RTT or other 212 path protperties. 213 Target RTT (Round Trip Time): The RTT over which the application 214 must meet the target performance. 215 Target MTU (Maximum Transmission Unit): Assume 1500 Bytes per packet 216 unless otherwise specified. If some subpath forces a smaller MTU, 217 then it becomes the target MTU, and all subpaths must be tested 218 with the same smaller MTU. 220 Effective Bottleneck Data Rate: This is the bottleneck data rate 221 that might be inferred from the ACK stream, by looking at how much 222 data the ACK stream reports was delivered per unit time. See 223 Section 4.1 for more details. 224 Permitted Number of Connections: The target rate can be more easily 225 obtained by dividing the traffic across more than one connection. 226 In general the number of concurrent connections is determined by 227 the application, however see the comments below on multiple 228 connections. 229 [sender] [interface] rate: The burst data rate, constrained by the 230 data sender's interfaces. Today 1 or 10 Gb/s are typical. 231 Header overhead: The IP and TCP header sizes, which are the portion 232 of each MTU not available for carrying application payload. 233 Without loss of generality this is assumed to be the size for 234 returning acknowledgements (ACKs). For TCP, the Maximum Segment 235 Size (MSS) is the Target MTU minus the header overhead. 237 Terminology about paths, etc. See [RFC2330] and 238 [I-D.morton-ippm-lmap-path]. 240 [data] sender Host sending data and receiving ACKs, typically via 241 TCP. 242 [data] receiver Host receiving data and sending ACKs, typically via 243 TCP. 244 subpath Subpath as defined in [RFC2330]. 245 Measurement Point Measurement points as described in 246 [I-D.morton-ippm-lmap-path]. 247 test path A path between two measurement points that includes a 248 subpath of the end-to-end path under test, plus possibly 249 additional infrastructure between the measurement points and the 250 subpath. 251 [Dominant] Bottleneck The Bottleneck that determines a flow's self 252 clock. It generally determines the traffic statistics for the 253 entire path. See Section 4.1. 254 front path The subpath from the data sender to the dominant 255 bottleneck. 256 back path The subpath from the dominant bottleneck to the receiver. 257 return path The path taken by the ACKs from the data receiver to the 258 data sender. 259 cross traffic Other, potentially interfering, traffic competing for 260 resources (network and/or queue capacity). 262 Basic parameters common to all models and subpath tests. They are 263 described in more detail in Section 5.2. 265 @ @@@ 266 pipe size The number of packets needed in flight (the window size) 267 to exactly fill some network path or sub path. The is the window 268 size which in normally the onset of queueing. 269 target_pipe_size: The number of packets in flight (the window size) 270 needed to exactly meet the target rate, with a single stream and 271 no cross traffic for the specified target data rate, RTT and MTU. 272 subpath pipe size 273 run length Observed, measured or specified number of packets that 274 are (to be) delivered between losses or ECN marks. Nominally one 275 over the loss probability. 276 target_run_length Required run length computed from the target data 277 rate, RTT and MTU. 278 reference_target_run_length: One specific conservative estimate of 279 the number of packets that must be delivered between loss episodes 280 in most diagnostic tests. 281 derating: The modeling framework permits some latitude in derating 282 some specific test parameters as described in Section 5.3. 284 Test types [These need work] 286 capacity tests: For "capacity tests" is required that as long as the 287 test traffic is within the proper envelope for the target end-to- 288 end performance, the average packet losses must be below the 289 threshold computed by the model. 290 Engineering tests: Engineering tests verify that the subpath under 291 test interacts well with TCP style self clocked protocols using 292 adaptive congestion control based on packet loss and ECN marks. 293 For example "AQM Tests" verify that when the presented load 294 exceeds the capacity of the subpath, the subpath signals for the 295 transport protocol to slow down, by appropriately ECN marking or 296 dropping some of the packets. Note while that cross traffic is 297 can cause capacity tests to fail, it has the potential to cause 298 AQM tests to false pass, which is why AQM tests require separate 299 test procedures. 301 3. New requirements relative to RFC 2330 303 Model Based Metrics are designed to fulfil some additional 304 requirement that were not recognized at the time RFC 2330 was 305 written. These missing requirements may have significantly 306 contributed to policy difficulties in the IP measurement space. Some 307 additional requirements are: 309 o Metrics must be actionable by the ISP - they have to be 310 interpreted in terms of behaviors or properties at the IP or lower 311 layers, that an ISP can test, repair and verify. 312 o Metrics must be vantage point invariant over a significant range 313 of measurement point choices (e.g., measurement points as 314 described in [I-D.morton-ippm-lmap-path]), including off path 315 measurement points. The only requirements on MP selection should 316 be that the portion of the path that is not under test is 317 effectively ideal (or is non ideal in calibratable ways) and the 318 end-to-end RTT between MPs is below some reasonable bound. 319 o Metrics must be repeatable by multiple parties. It must be 320 possible for different parties to make the same measurement and 321 observe the same results. In particular it is specifically 322 important that both a consumer (or their delegate) and ISP be able 323 to perform the same measurement and get the same result. 325 NB: All of the metric requirements in RFC 2330 should be reviewed and 326 potentially revised. If such a document is opened soon enough, this 327 entire section should be dropped. 329 4. Background 331 At the time the IPPM WG was chartered, sound Bulk Transport Capacity 332 measurement was known to be beyond our capabilities. By hindsight it 333 is now clear why it is such a hard problem: 334 o TCP is a control system with circular dependencies - everything 335 affects performance, including components that are explicitly not 336 part of the test. 337 o Congestion control is an equilibrium process, transport protocols 338 change the network (raise loss probability and/or RTT) to conform 339 to their behavior. 340 o TCP's ability to compensate for network flaws is directly 341 proportional to the number of roundtrips per second (i.e. 342 inversely proportional to the RTT). As a consequence a flawed 343 link may pass a short RTT local test even though it fails when the 344 path is extended by a perfect network to some larger RTT. 345 o TCP has a meta Heisenberg problem - Measurement and cross traffic 346 interact in unknown and ill defined ways. The situation is 347 actually worse than the traditional physics problem where you can 348 at least estimate the relative momentum of the measurement and 349 measured particles. For network measurement you can not in 350 general determine the relative "elasticity" of the measurement 351 traffic and cross traffic, so you can not even gage the relative 352 magnitude of their effects on each other. 354 The MBM approach is to "open loop" TCP by precomputing traffic 355 patterns that are typically generated by TCP operating at the given 356 target parameters, and evaluating delivery statistics (losses and 357 delay). In this approach the measurement software explicitly 358 controls the data rate, transmission pattern or cwnd (TCP's primary 359 congestion control state variables) to create repeatable traffic 360 patterns that mimic TCP behavior but are independent of the actual 361 network behavior of the subpath under test. These patterns are 362 manipulated to probe the network to verify that it can deliver all of 363 the traffic patterns that a transport protocol is likely to generate 364 under normal operation at the target rate and RTT. 366 Models are used to determine the actual test parameters (burst size, 367 loss rate, etc) from the target parameters. The basic method is to 368 use models to estimate specific network properties required to 369 sustain a given transport flow (or set of flows), and using a suite 370 of metrics to confirm that the network meets the required properties. 372 A network is expected to be able to sustain a Bulk TCP flow of a 373 given data rate, MTU and RTT when the following conditions are met: 374 o The raw link rate is higher than the target data rate. 375 o The raw packet loss rate is lower than required by a suitable TCP 376 performance model 377 o There is sufficient buffering at the dominant bottleneck to absorb 378 a slowstart rate burst large enough to get the flow out of 379 slowstart at a suitable window size. 380 o There is sufficient buffering in the front path to absorb and 381 smooth sender interface rate bursts at all scales that are likely 382 to be generated by the application, any channel arbitration in the 383 ACK path or other mechanisms. 384 o When there is a standing queue at a bottleneck for a shared media 385 subpath, there are suitable bounds on how the data and ACKs 386 interact, for example due to the channel arbitration mechanism. 387 o When there is a slowly rising standing queue at the bottleneck the 388 onset of packet loss has to be at an appropriate point (time or 389 queue depth) and progressive. 391 The tests to verify these condition are described in Section 7. 393 Note that this procedure is not invertible: a singleton measurement 394 is a pass/fail evaluation of a given path or subpath at a given 395 performance. Measurements to confirm that a link passes at one 396 particular performance may not be generally be useful to predict if 397 the link will pass at a different performance. 399 Although they are not invertible, they do have several other valuable 400 properties, such as natural ways to define several different 401 composition metrics [RFC5835]. 403 [Add text on algebra on metrics (A-Frame from [RFC2330]) and 404 tomography.] The Spatial Composition of fundamental IPPM metrics has 405 been studied and standardized. For example, the algebra to combine 406 empirical assessments of loss ratio to estimate complete path 407 performance is described in section 5.1.5. of [RFC6049]. We intend 408 to use this and other composition metrics as necessary. 410 4.1. TCP properties 412 TCP and SCTP are self clocked protocols. The dominant steady state 413 behavior is to have an approximately fixed quantity of data and 414 acknowledgements (ACKs) circulating in the network. The receiver 415 reports arriving data by returning ACKs to the data sender, the data 416 sender most frequently responds by sending exactly the same quantity 417 of data back into the network. The quantity of data plus the data 418 represented by ACKs circulating in the network is referred to as the 419 window. The mandatory congestion control algorithms incrementally 420 adjust the widow by sending slightly more or less data in response to 421 each ACK. The fundamentally important property of this systems is 422 that it is entirely self clocked: The data transmissions are a 423 reflection of the ACKs that were delivered by the network, the ACKs 424 are a reflection of the data arriving from the network. 426 A number of phenomena can cause bursts of data, even in idealized 427 networks that are modeled as simple queueing systems. 429 During slowstart the data rate is doubled by sending twice as much 430 data as was delivered to the receiver. For slowstart to be able to 431 fill such a network the network must be able to tolerate slowstart 432 bursts up to the full pipe size inflated by the anticipated window 433 reduction on the first loss. For example, with classic Reno 434 congestion control, an optimal slowstart has to end with a burst that 435 is twice the bottleneck rate for exactly one RTT in duration. This 436 burst causes a queue which is exactly equal to the pipe size (the 437 window is exactly twice the pipe size) so when the window is halved, 438 the new window will be exactly the pipe size. 440 Another source of bursts are application pauses. If the application 441 pauses (stops reading or writing data) for some fraction of one RTT, 442 state-of-the-art TCP to "catches up" to the earlier window size by 443 sending a burst of data at the full sender interface rate. To fill 444 such a network with a realistic application, the network has to be 445 able to tolerate interface rate bursts from the data sender large 446 enough to cover the worst case application pause. 448 Note that if the bottleneck data rate is significantly slower than 449 the rest of the path, the slowstart bursts will not cause significant 450 queues anywhere else along the path; they primarily exercise the 451 queue at the dominant bottleneck. Furthermore although the interface 452 rate bursts caused by the application are likely to be smaller than 453 burst at the last RTT of slowstart, they are at a higher rate so they 454 can exercise queues at arbitrary points along the "front path" from 455 the data sender up to and including the queue at the bottleneck. 457 For many network technologies a simple queueing model does not apply: 458 the network schedules, thins or otherwise alters the ACKs and data 459 stream, generally to raise the efficiency of the channel allocation 460 process when confronted with relatively widely spaced ACKs. These 461 efficiency strategies are ubiquitous for wireless and other half 462 duplex or broadcast media. 464 Altering the ACK stream generally has two consequences: raising the 465 effective bottleneck rate making slowstart burst at higher rates 466 (possibly as high as the sender's interface rate) and effectively 467 raising the RTT by the time that the ACKs were postponed. The first 468 effect can be partially mitigated by reclocking ACKs once they are 469 through the bottleneck on the return to the sender, however this 470 further raises the effective RTT. The most extreme example of this 471 class of behaviors is a half duplex channel that is never released 472 until the current sender has no pending traffic. Such environments 473 intrinsically cause self clocked protocols revert to extremely 474 inefficient stop and wait behavior, where they send an entire window 475 of data as a single burst, followed by the entire window of ACKs on 476 the return path. 478 If a particular end-to-end path contains a link or device that alters 479 the ACK stream, then the entire path from the sender up to the 480 bottleneck must be tested at the burst parameters implied by the ACK 481 scheduling algorithms. The most important parameter is the Effective 482 Bottleneck Data Rate, which is the average rate at which the ACKs 483 advance snd.una. Note that thinning the ACKs (relying on the 484 cumulative nature of seg.ack to permit discarding some ACKs) is 485 implies an effectively infinite bottleneck data rate. 487 To verify that a path can meet the performance target, Model Based 488 Metrics need to independently confirm that the entire path can 489 tolerate bursts of the dimensions that are likely to be induced by 490 the application and any data or ACK scheduling. Two common cases are 491 the most important: slowstart bursts of with more than the 492 target_pipe_size data at twice the effective bottleneck data rate; 493 and somewhat smaller sender interface rate bursts. 495 5. Common Models and Parameters 497 Transport performance models are used to derive the test parameters 498 for test suites of simple diagnostics from the end-to-end target 499 parameters and additional ancillary parameters. 501 5.1. Target End-to-end parameters 503 The target end to end parameters are the target data rate, target RTT 504 and target MTU as defined in Section 2 These parameters are 505 determined by the needs of the application or the ultimate end user 506 and the end-to-end Internet path. They are in units that make sense 507 to the upper layer: payload bytes delivered, excluding header 508 overheads for IP, TCP and other protocol. 510 Ancillary parameters include the effective bottleneck rate and the 511 permitted number of connections (numb_cons). 513 The use of multiple connections has been very controversial since the 514 beginning of the World-Wide-Web[first complaint]. Modern browsers 515 open many connections [BScope]. Experts associated with IETF 516 transport area have frequently spoken against this practice [long 517 list]. It is not inappropriate to assume some small number of 518 concurrent connections (e.g. 4 or 6), to compensate for limitation in 519 TCP. However, choosing too large a number is at risk of being 520 interpreted as a signal by the web browser community that this 521 practice has been embraced by the Internet service provider 522 community. It may not be desirable to send such a signal. 524 5.2. Common Model Calculations 526 The most important derived parameter is target_pipe_size (in 527 packets), which is the number of packets needed exactly meet the 528 target rate, with numb_cons connections and no cross traffic for the 529 specified target RTT and MTU. It is given by: 531 target_pipe_size = (target_rate / numb_cons) * target_RTT / ( 532 target_MTU - header_overhead ) 534 If the transport protocol (e.g. TCP) average window size is smaller 535 than this, it will not meet the target rate. 537 The reference_target_run_length, which is the most conservative model 538 for the minimum spacing between losses, can derived as follows: 539 assume the link_data_rate is infinitesimally larger than the 540 target_data_rate. Then target_pipe_size also predicts the onset of 541 queueing. If the transport protocol (e.g. TCP) has an average 542 window size that is larger than the target_pipe_size, the excess 543 packets will form a standing queue at the bottleneck. 545 If the transport protocol is using standard Reno style Additive 546 Increase, Multiplicative Decrease congestion control [RFC5681], then 547 there must be target_pipe_size roundtrips between losses. Otherwise 548 the multiplicative window reduction triggered by a loss would cause 549 the network to be underfilled. Following [MSMO97], we derive the 550 losses must be no more frequent than every 1 in 551 (3/2)(target_pipe_size^2) packets. This provides the reference value 552 for target_run_length which is typically the number of packets that 553 must be delivered between loss episodes in the tests below: 555 reference_target_run_length = (3/2)(target_pipe_size^2) 557 Note that this calculation is based on a number of assumptions that 558 may not apply. Appendix A discusses these assumptions and provides 559 some alternative models. The actual method for computing 560 target_run_length MUST be documented along with the rationale for the 561 underlying assumptions and the ratio of chosen target_run_length to 562 reference_target_run_length. @@@ MOVE 564 Although this document gives a lot of latitude for calculating 565 target_run_length, people designing suites of tests need to consider 566 the effect of their choices on the ongoing conversation and tussle 567 about the relevance of "TCP friendliness" as an appropriate model for 568 capacity allocation. Choosing a target_run_length that is 569 substantially smaller than reference_target_run_length is equivalent 570 to saying that it is appropriate for the transport research community 571 to abandon "TCP friendliness" as a fairness model and to develop more 572 aggressive Internet transport protocols, and for applications to 573 continue (or even increase) the number of connections that they open 574 concurrently. 576 The calculations for individual parameters are presented with the 577 each single property test. In general these calculations permit some 578 derating as described in Section 5.3. For test parameters that can 579 be derated and are proportional to target_pipe_size, it is 580 recommended that the derating be specified relative to 581 target_pipe_size calculations using numb_cons=1, although the 582 derating may additionally be specified relative to the 583 target_pipe_size common to other tests. 585 5.3. Parameter Derating 587 Since some aspects of the models are very conservative, the modeling 588 framework permits some latitude in derating some specific test 589 parameters. For example classical performance models suggest that in 590 order to be sure that a single TCP stream can fill a link, it needs 591 to have a full bandwidth-delay-product worth of buffering at the 592 bottleneck[QueueSize]. In real networks with real applications this 593 is often overly conservative. Rather than trying to formalize more 594 complicated models we permit some test parameters to be relaxed as 595 long as they meet some additional procedural constraints: 596 o The method used compute and justify the derated metrics is 597 published in such a way that it becomes a matter of public record. 598 @@@ introduce earlier 599 o The calibration procedures described in Section 9 are used to 600 demonstrate the feasibility of meeting the performance targets 601 with the derated test parameters. 602 o The calibration process itself is documented is such a way that 603 other researchers can duplicate the experiments and validate the 604 results. 606 In the test specifications in Section 7 assume 0 < derate <= 1, is a 607 derating parameter. These will be individually named in the final 608 document. In all cases making derate smaller makes the test more 609 tolerant. Derate = 1 is "full strenght". 611 Note that some test parameters are not permitted to be derated. 613 6. Common testing procedures 615 6.1. Traffic generating techniques 617 6.1.1. Paced transmission 619 Paced (burst) transmissions: send bursts of data on a timer to meet a 620 particular target rate and pattern. 621 Single: Send individual packets at the specified rate or headway. 622 Burst: Send sender interface rate bursts on a timer. Specify any 3 623 of average rate, packet size, burst size (number of packets) and 624 burst headway (burst start to start). These bursts are typically 625 sent as back-to-back packets at the testers interface rate. 626 Slowstart: Send 4 packet sender interface rate bursts at an average 627 rate equal to the minimum of twice effective bottleneck link rate 628 or the sender interface rate. This corresponds to the average 629 rate during a TCP slowstart when Appropriate Byte Counting [ABC] 630 is present or delayed ack is disabled. 631 Repeated Slowstart: Slowstart pacing itself is typically part of 632 larger scale pattern of repeated bursts, such as sending 633 target_pipe_size packets as slowstart bursts on a target_RTT 634 headway (burst start to burst start). Such a stream has three 635 different average rates, depending on the averaging time scale. 636 At the finest time scale the average rate is the same as the 637 sender interface rate, at a medium scale the average rate is twice 638 the bottleneck link rate and at the longest time scales the 639 average rate is the target data rate, adjusted to include header 640 overhead. 642 Note that if the effective bottleneck link rate is more than half of 643 the sender interface rate, slowstart bursts become sender interface 644 rate bursts. 646 6.1.2. Constant window pseudo CBR 648 Implement pseudo CBR by running a standard protocol such as TCP with 649 a fixed window size. This has the advantage that it can be 650 implemented as part of real content delivery. The rate is only 651 maintained in average over each RTT, and is subject to limitations of 652 the transport protocol. 654 For tests that have strongly prescribed data rates, if the transport 655 protocol fails to maintain the test rate for any reason related to 656 the network itself, such as packet losses or congestion, the test 657 should be considered inconclusive. Otherwise there are some cases 658 where tester failures might cause false negative link test results. 660 6.1.2.1. Scanned window pseudo CBR 662 Same as the above, except the window is incremented once per 663 2*target_pipe_size, starting from below target_pipe[@@@ test pipe] 664 and sweeping up to first loss or some other event. This is analogous 665 to the tests implemented in Windowed Ping [WPING] and pathdiag 666 [Pathdiag] 668 6.1.3. Intermittent Testing 670 Any test which does not depend on queueing (e.g. the CBR tests) or 671 experiences periodic zero outstanding data during normal operation 672 (e.g. between bursts for burst tests), can be formulated as an 673 intermittent test. 675 The Intermittent testing can be used for ongoing monitoring for 676 changes in subpath quality with minimal disruption users. It should 677 be used in conjunction with the full rate test because this method 678 assesses an average_run_length over a long time interval w.r.t. user 679 sessions. It may false fail due to other legitimate congestion 680 causing traffic or may false pass changes in underlying link 681 properties (e.g. a modem retraining to an out of contract lower 682 rate). 684 [Need text about bias (false pass) in the shadow of loss caused by 685 excessive bursts] 687 6.1.4. Intermittent Scatter Testing 689 Intermittent scatter testing: when testing the network path to or 690 from an ISP subscriber aggregation point (CMTS, DSLAM, etc), 691 intermittent tests can be spread across a pool of users such that no 692 one users experiences the full impact of the testing, even though the 693 traffic to or from the ISP subscriber aggregation point is sustained 694 at full rate. 696 6.2. Interpreting the Results 698 6.2.1. Test outcomes 700 A singleton is a pass fail measurement. If any subpath fails any 701 test it can be assumed that the end-to-end path will also fail to 702 attain the target performance under some conditions. 704 In addition we use "inconclusive" outcome to indicate that a test 705 failed to attain the required test conditions. This is important to 706 the extent that the tests themselves use protocols that have built in 707 control systems which might interfere with some aspect of the test. 708 For example consider a test is implemented by adding rate controls 709 and instrumentation to TCP: failing to attain the specified data rate 710 has to be treated an inconclusive, unless the test clearly fails 711 (target_run_lenght is too small). This is because failing to reach 712 the target rate is an ambiguous signature for problems with either 713 the test procedure (a problem with the TCP implementation or the test 714 path RTT is too long) or the subpath itself. 716 The vantage independence properties of Model Based Metrics depends on 717 the accuracy of the distinction between failing and inconclusive 718 tests. One of the goals of evolving test designs will be to keep 719 sharpening the distinction between failing and inconclusive tests. 721 One of the goals of evolving the testing process, procedures and 722 measurement point selection should be to minimize the number of 723 inconclusive tests. 725 6.2.2. Statistical criteria for measuring run_length 727 When evaluating the observed run_length, we need to determine 728 appropriate packet stream sizes and acceptable error levels to test 729 efficiently. In practice, can we compare the empirically estimated 730 loss probabilities with the targets as the sample size grows? How 731 large a sample is needed to say that the measurements of packet 732 transfer indicate a particular run-length is present? 734 The generalized measurement can be described as recursive testing: 736 send a flight of packets and observe the packet transfer performance 737 (loss ratio or other metric, any defect we define). 739 As each flight is sent and measured, we have an ongoing estimate of 740 the performance in terms of defect to total packet ratio (or an 741 empirical probability). Continue to send until conditions support a 742 conclusion or a maximum sending limit has been reached. 744 We have a target_defect_probability, 1 defect per target_run_length, 745 where a "defect" is defined as a lost packet, a packet with ECN mark, 746 or other impairment. This constitutes the null Hypothesis: 748 H0: no more than one defects in target_run_length = (3/2)*(flight)^2 749 packets 751 and we can stop sending flights of packets if measurements support 752 accepting H0 with the specified Type I error = alpha (= 0.05 for 753 example). 755 We also have an alternative Hypothesis to evaluate: if performance is 756 significantly lower than the target_defect_probability, say half the 757 target: 759 H1: one or more defects in target_run_length/2 packets 761 and we can stop sending flights of packets if measurements support 762 rejecting H0 with the specified Type II error = beta, thus preferring 763 the alternate H1. 765 H0 and H1 constitute the Success and Failure outcomes described 766 elsewhere in the memo, and while the ongoing measurements do not 767 support either hypothesis the current status of measurements is 768 inconclusive. 770 The problem above is formulated to match the Sequential Probability 771 Ratio Test (SPRT) [StatQC] [temp ref: 772 http://en.wikipedia.org/wiki/Sequential_probability_ratio_test ], 773 which also starts with a pair of hypothesis specified as above: 775 H0: p = p0 = one defect in target_run_length 776 H1: p = p1 = one defect in target_run_length/2 777 As flights are sent and measurements collected, the tester evaluates 778 the cumulative log-likelihood ratio: 780 S_i = S_i-1 + log(Lambda_i) 782 where Lambda_i is the ratio of the two likelihood functions 783 (calculated on the measurement at packet i, and index i increases 784 linearly over all flights of packets ) for p0 and p1 [temp ref: 785 http://en.wikipedia.org/wiki/Likelihood_function ]. 787 The SPRT specifies simple stopping rules: 789 o a < S_i < b: continue testing 790 o S_i <= a: Accept H0 791 o S_i >= b: Accept H1 792 where a and b are based on the Type I and II errors, alpha and beta: 794 a ~= Log((beta/(1-alpha)) and b ~= Log((1-beta)/alpha) 796 with the error probabilities decided beforehand, as above. 798 The calculations above are implemented in the R-tool for Statistical 799 Analysis, in the add-on package for Cross-Validation via Sequential 800 Testing (CVST) [http://www.r-project.org/] [Rtool] [CVST] . 802 6.2.3. Classifications of tests 804 Tests are annotated with "(capacity)", "(engineering)" or 805 "(monitoring)". @@@@MOVE to definitions? 807 Capacity tests determine if a network subpath has sufficient capacity 808 to deliver the target performance. As such, they reflect parameters 809 that can transition from passing to failing as a consequence of 810 additional presented load or the actions of other network users. By 811 definition, capacity tests also consume network resources (capacity 812 and/or buffer space), and their test schedules must be balanced by 813 their cost. 815 Monitoring tests are design to capture the most important aspects of 816 a capacity test, but without causing unreasonable ongoing load 817 themselves. As such they may miss some details of the network 818 performance, but can serve as a useful reduced cost proxy for a 819 capacity test. 821 Engineering tests evaluate how network algorithms (such as AQM and 822 channel allocation) interact with transport protocols. These tests 823 are likely to have complicated interactions with other network 824 traffic and can be inversely sensitive to load. For example a test 825 to verify that an AQM algorithm causes ECN marks or packet drops 826 early enough to limit queue occupancy may experience a false pass 827 results in the presence of bursty cross traffic. It is important 828 that engineering tests be performed under a wide range of conditions, 829 including both in situ and bench testing, and under a variety of load 830 conditions. Ongoing monitoring is less likely to be useful for these 831 tests, although sparse in situ testing might be appropriate. 833 @@@ Add single property vs combined tests here? 835 6.2.4. Reordering Tolerance 837 All tests must be instrumented for reordering [RFC4737]. 839 NB: there is no global consensus for how much reordering tolerance is 840 appropriate or reasonable. ("None" is absolutely unreasonable.) 842 Section 5 of [RFC4737] proposed a metric that may be sufficient to 843 designate isolated reordered packets as effectively lost, because 844 TCP's retransmission response would be the same. 846 [As a strawman, we propose the following:] TCP should be able to 847 adapt to reordering as long as the reordering extent is no more than 848 the maximum of one half window or 1 mS, whichever is larger. Note 849 that there is a fundamental tradeoff between tolerance to reordering 850 and how quickly algorithms such as fast retransmit can repair losses. 851 Within this limit on reorder extent, there should be no bound on 852 reordering frequency. 854 NB: Current TCP implementations are not compatible with this metric. 855 We view this as bugs in current TCP implementations. 857 Parameters: 858 Reordering displacement: the maximum of one half of target_pipe_size 859 or 1 mS. 861 6.3. Test Qualifications 863 Things to monitor before, during and after a test. 865 6.3.1. Verify the Traffic Generation Accuracy 867 for most tests, failing to accurately generate the test traffic 868 indicates an inconclusive tests, since it has to be presumed that the 869 error in traffic generation might have affected the test outcome. To 870 the extent that the network itself had an effect on the the traffic 871 generation (e.g. in the standing queue tests) the possibility exists 872 that allowing too large of error margin in the traffic generation 873 might introduce feedback loops that comprise the vantage independents 874 properties of these tests. 876 Parameters: 878 Maximum Data Rate Error The permitted amount that the test traffic 879 can be different than specified for the current test. This is a 880 symmetrical bound. 881 Maximum Data Rate Overage The permitted amount that the test traffic 882 can be above than specified for the current test. 883 Maximum Data Rate Underage The permitted amount that the test 884 traffic can be less than specified for the current test. 886 6.3.2. Verify the absence of cross traffic 888 The proper treatment of cross traffic is different for different 889 subpaths. In general when testing infrastructure which is associated 890 with only one subscriber, the test should be treated as inconclusive 891 it that subscriber is active on the network. However, for shared 892 infrastructure, the question at hand is likely to be testing if 893 provider has sufficient total capacity. In such cases the presence 894 of cross traffic due to other subscribers is explicitly part of the 895 network conditions and its effects are explicitly part of the test. 897 Note that canceling tests due to load on subscriber lines may 898 introduce sampling errors for testing other parts of the 899 infrastructure. For this reason tests that are scheduled but not run 900 due to load should be treated as a special case of "inconclusive". 902 Use a passive packet or SNMP monitoring to verify that the traffic 903 volume on the subpath agrees with the traffic generated by a test. 904 Ideally this should be performed before during and after each test. 906 The goal is provide quality assurance on the overall measurement 907 process, and specifically to detect the following measurement 908 failure: a user observes unexpectedly poor application performance, 909 the ISP observes that the access link is running at the rated 910 capacity. Both fail to observe that the user's computer has been 911 infected by a virus which is spewing traffic as fast as it can. 913 Parameters: 914 Maximum Cross Traffic Data Rate The amount of excess traffic 915 permitted. Note that this will be different for different tests. 917 One possible method is an adaptation of: www-didc.lbl.gov/papers/ 918 SCNM-PAM03.pdf D Agarwal etal. "An Infrastructure for Passive 919 Network Monitoring of Application Data Streams". Use the same 920 technique as that paper to trigger the capture of SNMP statistics for 921 the link. 923 6.3.3. Additional test preconditions 925 Send pre-load traffic as needed to activate radios with a sleep mode, 926 or other "reactive network" elements (term defined in 927 [draft-morton-ippm-2330-update-01]). 929 Use the procedure above to confirm that the pre-test background 930 traffic is low enough. 932 7. Single Property Tests 934 7.1. Basic Data and Loss Rate Tests 936 We propose several versions of the loss rate test. All are rate 937 controlled at or below the target_data_rate. The first, performed at 938 constant full data rate, is intrusive and recommend for infrequent 939 testing, such as when a service is first turned up or as part of an 940 auditing process. The second, background loss rate, is designed for 941 ongoing monitoring for change is subpath quality. 943 7.1.1. Loss Rate at Paced Full Data Rate 945 Confirm that the observed run length is at least the 946 target_run_lenght while sending at the target_rate. This test 947 implicitly confirms that sub_path has sufficient raw capacity to 948 carry the target_data_rate. This version of the loss rate test 949 relies on timers to schedule data transmission at a true constant bit 950 rate (CBR). 952 Test Parameters: 953 Run Length Same as target_run_lenght 954 Data Rate Same as target_data_rate 955 Maximum Cross Traffic A specified small fraction of 956 target_data_rate. 958 Note that target_run_lenght and target_data_rate parameters MUST NOT 959 be derated. If the default parameters are too stringent an alternate 960 model as described in Appendix A can be used to compute 961 target_run_lenght. 963 The test traffic is sent using the procedures in Section 6.1.1 at 964 target_data_rate with a burst size of 1, subject to the 965 qualifications in Section 6.3. The receiver accumulates packet 966 delivery statistics as described in Section 6.2 to score the outcome: 968 Pass: it is statistically significantly that the observed run length 969 is larger than the target_run_length. 971 Fail: it is statistically significantly that the observed run length 972 is smaller than the target_run_length. 974 Inconclusive: The test failed to meet the qualifications defined in 975 Section 6.3 or neither test was statistically significant. 977 7.1.2. Loss Rate at Full Data Windowed Rate 979 Confirm that the observed run length is at least the 980 target_run_lenght while sending at the target_rate. This test 981 implicitly confirms that sub_path has sufficient raw capacity to 982 carry the target_data_rate. This version of the loss rate test 983 relies on a fixed window to self clock data transmission into the 984 network. This is more authentic. 986 Test Parameters: 987 Run Length Same as target_run_lenght 988 Data Rate Same as target_data_rate 989 Maximum Cross Traffic A specified small fraction of 990 target_data_rate. 992 Note that target_run_lenght and target_data_rate parameters MUST NOT 993 be derated. If the default parameters are too stringent an alternate 994 model as described in Appendix A can be used to compute 995 target_run_lenght. 997 The test traffic is sent using the procedures in Section 6.1.1 at 998 target_data_rate with a burst size of 1, subject to the 999 qualifications in Section 6.3. The receiver accumulates packet 1000 delivery statistics as described in Section 6.2 to score the outcome: 1002 Pass: it is statistically significantly that the observed run length 1003 is larger than the target_run_length. 1005 Fail: it is statistically significantly that the observed run length 1006 is smaller than the target_run_length. 1008 Inconclusive: The test failed to meet the qualifications defined in 1009 Section 6.3 or neither test was statistically significant. 1011 7.1.3. Background Loss Rate Tests 1013 The background loss rate is a low rate version of the target rate 1014 test above, designed for ongoing monitoring for changes in subpath 1015 quality without disrupting users. It should be used in conjunction 1016 with the above full rate test because it may be subject to false 1017 results under some conditions, in particular it may false pass 1018 changes in underlying link properties (e.g. a modem retraining to an 1019 out of contract lower rate). 1021 Parameters: 1022 Run Length Same as target_run_lenght 1023 Data Rate Some small fraction of target_data_rate, such as 1%. 1025 Once the preconditions described in Section 6.3 are met, the test 1026 data is sent at the prescribed rate with a burst size of 1. The 1027 receiver accumulates packet delivery statistics and the procedures 1028 described in Section 6.2.1 and Section 6.3 are used to score the 1029 outcome: 1031 Pass: it is statistically significantly that the observed run length 1032 is larger than the target_run_length. 1034 Fail: it is statistically significantly that the observed run length 1035 is smaller than the target_run_length. 1037 Inconclusive: Neither test was statistically significant or there was 1038 excess cross traffic during the test. 1040 7.2. Standing Queue tests 1042 These test confirm that the bottleneck is well behaved across the 1043 onset of queueing. For conventional bottlenecks this will be from 1044 the onset of queuing to the point where there is a full target_pipe 1045 of standing data. Well behaved generally means lossless for 1046 target_run_length, followed by a small number of losses to signal to 1047 the transport protocol that it should slow down. Losses that are too 1048 early can prevent the transport from averaging above the target_rate. 1049 Losses that are too late indicate that the queue might be subject to 1050 bufferbloat and subject other flows to excess queuing delay. Excess 1051 losses (more than half of of target_pipe) make loss recovery 1052 problematic for the transport protcol. 1054 These tests can also observe some problems with channel acquisition 1055 systems, especially at the onset of persistent queueing. Details 1056 TBD. 1058 7.2.1. Congestion Avoidance 1060 Use the procedure in Section 6.1.2.1 to sweep the window (rate) from 1061 below link_pipe up to beyond target_pipe+link_pipe. Depending on 1062 events that happen during the scan, score the link. Identify the 1063 power_point=MAX(rate/RTT) as the start of the test. 1065 Fail if first loss is too early (loss rate too high) on repeated 1066 tests or if the losses are more than half of the outstanding data. (a 1067 capacity test) 1069 7.2.2. Buffer Bloat 1071 Use the procedure in Section 6.1.2.1 to sweep the window (rate) from 1072 below link_pipe up to beyond target_pipe+link_pipe. Depending on 1073 events that happen during the scan, score the link. Identify the 1074 "power point:MAX(rate/RTT) as the start of the test (should be 1075 window=target_pipe) 1077 Fail if first loss is too late (insufficient AQM and subject to 1078 bufferbloat - an engineering test). NO THEORY 1080 7.2.3. Duplex Self Interference 1082 Use the procedure in Section 6.1.2.1 to sweep the window (rate) from 1083 below link_pipe up to beyond target_pipe+required_queue. Depending 1084 on events that happen during the scan, score the link. Identify the 1085 "power point:MAX(rate/RTT) as the start of the test (should be 1086 window=target_pipe) @@@ add required_queue and power_point 1088 Fail if RTT is non-monotonic by more than a small number of packet 1089 times (channel allocation self interference - engineering) IS THIS 1090 SUFFICIENT? 1092 7.3. Slowstart tests 1094 These tests mimic slowstart: data is sent at slowstart_rate (twice 1095 subpath_rate). They are deemed inconclusive if the elapsed time to 1096 send the data burst is not less than half of the (extrapolated) time 1097 to receive the ACKs. (i.e. sending data too fast is ok, but sending 1098 it slower than twice the actual bottleneck rate is deemed 1099 inconclusive). Space the bursts such that the average ACK rate is 1100 equal to or faster than the target_data_rate. 1102 These tests are not useful at burst sizes smaller than the sender 1103 interface rate tests, since the sender interface rate tests are more 1104 strenuous. If it is necessary to derate the sender interface rate 1105 tests, then the full window slowstart test (un-derated) would be 1106 important. 1108 7.3.1. Full Window slowstart test 1110 Send (target_pipe_size+required_queue)*derate bursts must have fewer 1111 than one loss per target_run_length*derate. Note that these are the 1112 same parameters as the Sender Full Window burst test, except the 1113 burst rate is at slowestart rate, rather than sender interface rate. 1114 SHOULD derate=1. 1116 Otherwise TCP will exit from slowstart prematurely, and only reach a 1117 full target_pipe_size window by way of congestion avoidance. 1119 This is a capacity test: cross traffic may cause premature losses. 1121 7.3.2. Slowstart AQM test 1123 Do a continuous slowstart (date rate = slowstart_rate), until first 1124 loss, and repeat, gathering statistics on the last delivered packet's 1125 RTT and window size. Fail if too large (NO THEORY for value). 1127 This is an engineering test: It would be best performed on a 1128 quiescent network or testbed, since cross traffic might cause a false 1129 pass. 1131 7.4. Sender Rate Burst tests 1133 These tests us "sender interface rate" bursts. Although this is not 1134 well defined it should be assumed to be current state of the art 1135 server grade hardware (often 10Gb/s today). (load) 1137 7.4.1. Sender TCP Send Offload (TSO) tests 1139 If MIN(target_pipe_size, 42) packet bursts meet target_run_lenght 1140 (Not derated!). 1142 Otherwise the link will interact badly with modern server NIC 1143 implementations, which as an optimization to reduce host side 1144 interactions (interrupts etc) accept up to 64kB super packets and 1145 send them as 42 seperate packets on the wire side.cc (load) 1147 7.4.2. Sender Full Window burst test 1149 target_pipe_size*derate bursts have fewer than one loss per 1150 target_run_length*derate. 1152 Otherwise application pauses will cause unwarranted losses. Current 1153 standards permit TCP to send a full cwnd burst following an 1154 application pause. (Cwnd validation in not required, but even so 1155 does not take effect until the pause is longer than RTO). 1157 NB: there is no model here for what is good enough. derate=1 is 1158 safest, but may be unnecessarily conservative for some applications. 1159 Some application, such as streaming video need derate=1 to be 1160 efficient when the application pacing quanta is larger than cwnd. 1161 (load) 1163 8. Combined Tests 1165 These tests are more efficient from a deployment/operational 1166 perspective, but may not be possible to diagnose if they fail. 1168 8.1. Sustained burst test 1170 Send target_pipe_size sender interface rate bursts every target_RTT, 1171 verify that the observed run length meets target_run_length. Key 1172 observations: 1173 o This test is RTT invariant, as long as the tester can generate the 1174 required pattern. 1175 o The subpath under test is expected to go idle for some fraction of 1176 the time: (link_rate-target_rate)/link_rate. Failing to do so 1177 suggests a problem with the procedure. 1178 o This test is more strenuous than the slowstart tests: they are not 1179 needed if the link passes underated sender interface rate burst 1180 tests. 1181 o This test could be derated by reducing both the burst size and 1182 headway (same average data rate). 1183 o A link that passes this test is likely to be able to sustain 1184 higher rates (close to link_rate) for paths with RTTs smaller than 1185 the target_RTT. Offsetting this performance underestimation is 1186 the rationale behind permitting derating in general. 1187 o This test should be implementable with standard instrumented TCP, 1188 [RFC 4898] using a specialized measurement application at one end 1189 and a minimal service at the other end [RFC 863, RFC 864]. It may 1190 require tweaks to the TCP implementation. 1191 o This test is efficient to implement, since it does not require 1192 per-packet timers, and can make maximal use of TSO in modern NIC 1193 hardware. 1194 o This test is not totally sufficient: the standing window 1195 engineering tests are also needed to be sure that the link is well 1196 behaved at and beyond the onset of congestion. 1197 o I believe that this test can be proven to be the one capacity test 1198 to supplant them all. 1200 Example 1202 To confirm that a 100 Mb/s link can reliably deliver single 10 1203 MByte/s stream at a distance of 50 mS, test the link by sending 346 1204 packet bursts every 50 mS (10 MByte/s payload rate, assuming a 1500 1205 Byte IP MTU and 52 Byte TCP/IP headers). These bursts are 4196288 1206 bits on the wire (assuming 16 bytes of link overhead and framing) for 1207 an aggregate test data rate of 8.4 Mb/s. 1209 To pass the test using the most conservative TCP model for a single 1210 stream the observed run length must be larger than 179574 packets. 1212 This is the same as less than one loss per 519 bursts (1.5*346) or 1213 every 26 seconds. 1215 Note that this test potentially cause transient 346 packet queues at 1216 the bottleneck. 1218 9. Calibration 1220 If using derated metrics, or when something goes wrong, the results 1221 must be calibrated against a traditional BTC. The preferred 1222 diagnostic follow-up to calibration issues is to run open end-to-end 1223 measurements on an open platform, such as Measurement Lab 1224 [http://www.measurementlab.net/] 1226 10. Acknowledgements 1228 Ganga Maguluri suggested the statistical test for measuring loss 1229 probability in the target run length. 1231 Meredith Whittaker for improving the clarity of the communications. 1233 11. Informative References 1235 [RFC2330] Paxson, V., Almes, G., Mahdavi, J., and M. Mathis, 1236 "Framework for IP Performance Metrics", RFC 2330, 1237 May 1998. 1239 [RFC4737] Morton, A., Ciavattone, L., Ramachandran, G., Shalunov, 1240 S., and J. Perser, "Packet Reordering Metrics", RFC 4737, 1241 November 2006. 1243 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion 1244 Control", RFC 5681, September 2009. 1246 [RFC5835] Morton, A. and S. Van den Berghe, "Framework for Metric 1247 Composition", RFC 5835, April 2010. 1249 [RFC6049] Morton, A. and E. Stephan, "Spatial Composition of 1250 Metrics", RFC 6049, January 2011. 1252 [I-D.morton-ippm-lmap-path] 1253 Bagnulo, M., Burbridge, T., Crawford, S., Eardley, P., and 1254 A. Morton, "A Reference Path and Measurement Points for 1255 LMAP", draft-morton-ippm-lmap-path-00 (work in progress), 1256 January 2013. 1258 [MSMO97] Mathis, M., Semke, J., Mahdavi, J., and T. Ott, "The 1259 Macroscopic Behavior of the TCP Congestion Avoidance 1260 Algorithm", Computer Communications Review volume 27, 1261 number3, July 1997. 1263 [WPING] Mathis, M., "Windowed Ping: An IP Level Performance 1264 Diagnostic", INET 94, June 1994. 1266 [Pathdiag] 1267 Mathis, M., Heffner, J., O'Neil, P., and P. Siemsen, 1268 "Pathdiag: Automated TCP Diagnosis", Passive and Active 1269 Measurement , June 2008. 1271 [BScope] Broswerscope, "Browserscope Network tests", Sept 2012, 1272 . 1274 [Rtool] R Development Core Team, "R: A language and environment 1275 for statistical computing. R Foundation for Statistical 1276 Computing, Vienna, Austria. ISBN 3-900051-07-0, URL 1277 http://www.R-project.org/", , 2011. 1279 [StatQC] Montgomery, D., "Introduction to Statistical Quality 1280 Control - 2nd ed.", ISBN 0-471-51988-X, 1990. 1282 [CVST] Krueger, T. and M. Braun, "R package: Fast Cross- 1283 Validation via Sequential Testing", version 0.1, 11 2012. 1285 Appendix A. Model Derivations 1287 This appendix describes several different ways to calculate 1288 target_run_length and the implication of the chosen calculation. 1290 Rederive MSMO97 under two different assumptions: target_rate = 1291 link_rate and target_rate < 2 * link_rate. 1293 Show equivalent derivation for CUBIC. 1295 Commentary on the consequence of the choice. 1297 Appendix B. old text 1299 This entire section is contains scraps of text to be moved, removed 1300 or absorbed elsewhere in the document 1302 B.1. An earlier document 1304 Step 0: select target end-to-end parameters: a target rate and target 1305 RTT. The primary test will be to confirm that the link quality is 1306 sufficient to meet the specified target rate for the link under test, 1307 when extended to the target RTT by an ideal network. The target rate 1308 must be below the actual link rate and nominally the target RTT would 1309 be longer than the link RTT. There should probably be a convention 1310 for the relationship between link and target rates (e.g. 85%). 1312 For example on a 10 Mb/s link, the target rate might be 1 MBytes/s, 1313 at an RTT of 100 mS (a typical continental scale path). 1315 Step 1: On the basis of the target rate and RTT and your favorite TCP 1316 performance model, compute the "required run length", which is the 1317 required number of consecutive non-losses between loss episodes. The 1318 run length resembles one over the loss probability, if clustered 1319 losses only count as a single event. Also select "test duration" and 1320 "test rate". The latter would nominally the same as the target rate, 1321 but might be different in some situations. There must be 1322 documentation connecting the test rate, duration and required run 1323 length, to the target rate and RTT selected in step 0. 1325 Continuing the above example: Assuming a 1500 Byte MTU. The 1326 calculated model loss rate for a single TCP stream is about 0.01% (1 1327 loss in 1E4 packets). 1329 Step 2, the actual measurement proceeds as follows: Start an 1330 unconstrained bulk data flow using any modern TCP (with large buffers 1331 and/or autotuning). During the first interval (no rate limits) 1332 observe the slowstart (e.g. tcpdump) and measure: Peak burst size; 1333 link clock rate (delivery rate for each round); peak data rate for 1334 the fastest single RTT interval; fraction of segments lost at the end 1335 of slowstart. After the flow has fully recovered from the slowstart 1336 (details not important) throttle the flow down to the test rate (by 1337 clamping cwnd or application pacing at the sender or receiver). 1338 While clamped to the test rate, observe the losses (run length) for 1339 the chosen test duration. The link passes the test if the slowstart 1340 ends with less than approximately 50% losses and no timeouts, the 1341 peak rate is at least the target rate, and the measured run length is 1342 better than the required run length. There will also need to be some 1343 ancillary metrics, for example to discard tests where the receiver 1344 closes the window, invalidating the slowstart test. [This needs to 1345 be separated into multiple subtests] 1347 Optional step 3: In some cases it might make sense to compute an 1348 "extrapolated rate", which is the minimum of the observed peak rate, 1349 and the rate computed from the specified target RTT and the observed 1350 run length by using a suitable TCP performance model. The 1351 extrapolated rate should be annotated to indicate if it was run 1352 length or peak rate limited, since these have different predictive 1353 values. 1355 Other issues: 1357 If the link RTT is not substantially smaller than the target RTT and 1358 the actual run length is close to the target rate, a standards 1359 compliant TCP implementation might not be effective at accurately 1360 controlling the data rate. To be independent of the details of the 1361 TCP implementation, failing to control the rate has to be treated as 1362 a spoiled measurement, not a infrastructure failure. This can be 1363 overcome by "stiffening" TCP by using a non-standard congestion 1364 control algorithm. For example if the rate controlling by clamping 1365 cwnd then use "relentless TCP" style reductions on loss, and lock 1366 ssthresh to the cwnd clamp. Alternatively, implement an explicit 1367 rate controller for TCP. In either case the test must be abandoned 1368 (aborted) if the measured run length is substantially below the 1369 target run length. 1371 If the test is run "in situ" in a production environment, there also 1372 needs to be baseline tests using alternate paths to confirm that 1373 there are no bottlenecks or congested links between the test end 1374 points and the link under test. 1376 It might make sense to run multiple tests with different parameters, 1377 for example infrequent tests with test rate equal to the target rate, 1378 and more frequent, less disruptive tests with the same target rate 1379 but the test rate equal to 1% of the target rate. To observe the 1380 required run length, the low rate test would take 100 times longer to 1381 run. 1383 Returning to the example: a full rate test would entail sending 690 1384 pps (1 MByte/s) for several tens of seconds (e.g. 50k packets), and 1385 observing that the total loss rate is below 1:1e4. A less disruptive 1386 test might be to send at 6.9 pps for 100 times longer, and observing 1388 B.2. End-to-end parameters from subpaths 1390 [This entire section needs to be overhauled and should be skipped on 1391 a first reading. The concepts defined here are not used elsewhere.] 1393 The following optional parameters apply for testing generalized end- 1394 to-end paths that include subpaths with known specific types of 1395 behaviors that are not well represented by simple queueing models: 1397 Bottleneck link clock rate: This applies to links that are using 1398 virtual queues or other techniques to police or shape users 1399 traffic at lower rates full link rate. The bottleneck link clock 1400 rate should be representative of queue drain times for short 1401 bursts of packets on an otherwise unloaded link. 1402 Channel hold time: For channels that have relatively expensive 1403 channel arbitration algorithms, this is the typical (maximum?) 1404 time that data and or ACKs are held pending acquiring the channel. 1405 While under heavy load, the RTT may be inflated by this parameter, 1406 unless it is built into the target RTT 1407 Preload traffic volume: If the user's traffic is shaped on the basis 1408 of average traffic volume, this is volume necessary to invoke 1409 "heavy hitter" policies. 1410 Unloaded traffic volume: If the user's traffic is shaped on the 1411 basis of average traffic volume, this is the maximum traffic 1412 volume that a test can use and stay within a "light user" 1413 policies. 1415 Note on a ConEx enabled network [ConEx], the word "traffic" in the 1416 last two items should be replaced by "congestion" i.e. "preload 1417 congestion volume" and "unloaded congestion volume". 1419 B.3. Per subpath parameters 1421 [This entire section needs to be overhauled and should be skipped on 1422 a first reading. The concepts defined here are not used elsewhere.] 1424 Some single parameter tests also need parameter of the subpath. 1426 subpath RTT: RTT of the subpath under test. 1427 subpath link clock rate: If different than the Bottleneck link clock 1428 rate 1430 B.4. Version Control 1432 Formatted: Fri Jun 21 18:23:29 PDT 2013 1434 Authors' Addresses 1436 Matt Mathis 1437 Google, Inc 1438 1600 Amphitheater Parkway 1439 Mountain View, California 93117 1440 USA 1442 Email: mattmathis@google.com 1443 Al Morton 1444 AT&T Labs 1445 200 Laurel Avenue South 1446 Middletown, NJ 07748 1447 USA 1449 Phone: +1 732 420 1571 1450 Email: acmorton@att.com 1451 URI: http://home.comcast.net/~acmacm/