idnits 2.17.1 draft-morton-ippm-2679-bis-05.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 draft header indicates that this document obsoletes RFC2679, but the abstract doesn't seem to directly say this. It does mention RFC2679 though, so this could be OK. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year -- The document date (July 4, 2014) is 3584 days in the past. Is this intentional? Checking references for intended status: Proposed Standard ---------------------------------------------------------------------------- (See RFCs 3967 and 4897 for information about using normative references to lower-maturity documents in RFCs) -- Looks like a reference, but probably isn't: '1' on line 967 -- Looks like a reference, but probably isn't: '2' on line 970 -- Looks like a reference, but probably isn't: '4' on line 975 -- Looks like a reference, but probably isn't: '5' on line 978 -- Looks like a reference, but probably isn't: '3' on line 973 -- Looks like a reference, but probably isn't: '6' on line 980 -- Looks like a reference, but probably isn't: '7' on line 983 == Unused Reference: 'RFC2026' is defined on line 990, but no explicit reference was found in the text == Unused Reference: 'RFC2330' is defined on line 996, but no explicit reference was found in the text == Unused Reference: 'RFC2680' is defined on line 1003, but no explicit reference was found in the text == Unused Reference: 'RFC3432' is defined on line 1006, but no explicit reference was found in the text == Unused Reference: 'RFC4656' is defined on line 1010, but no explicit reference was found in the text == Unused Reference: 'RFC5357' is defined on line 1014, but no explicit reference was found in the text == Unused Reference: 'RFC5657' is defined on line 1018, but no explicit reference was found in the text == Unused Reference: 'RFC5835' is defined on line 1022, but no explicit reference was found in the text == Unused Reference: 'RFC6049' is defined on line 1025, but no explicit reference was found in the text == Unused Reference: 'ADK' is defined on line 1038, but no explicit reference was found in the text == Unused Reference: 'RFC3931' is defined on line 1049, but no explicit reference was found in the text ** Downref: Normative reference to an Informational RFC: RFC 2330 ** Obsolete normative reference: RFC 2679 (Obsoleted by RFC 7679) ** Obsolete normative reference: RFC 2680 (Obsoleted by RFC 7680) ** Downref: Normative reference to an Informational RFC: RFC 5835 ** Downref: Normative reference to an Informational RFC: RFC 6703 Summary: 5 errors (**), 0 flaws (~~), 12 warnings (==), 9 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Network Working Group G. Almes 3 Internet-Draft Texas A&M 4 Obsoletes: 2679 (if approved) S. Kalidindi 5 Intended status: Standards Track Ixia 6 Expires: January 5, 2015 M. Zekauskas 7 Internet2 8 A. Morton, Ed. 9 AT&T Labs 10 July 4, 2014 12 A One-Way Delay Metric for IPPM 13 draft-morton-ippm-2679-bis-05 15 Abstract 17 This memo (RFC 2679 bis) defines a metric for one-way delay of 18 packets across Internet paths. It builds on notions introduced and 19 discussed in the IPPM Framework document, RFC 2330; the reader is 20 assumed to be familiar with that document. 22 Requirements Language 24 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 25 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 26 document are to be interpreted as described in RFC 2119 [RFC2119]. 28 Status of This Memo 30 This Internet-Draft is submitted in full conformance with the 31 provisions of BCP 78 and BCP 79. 33 Internet-Drafts are working documents of the Internet Engineering 34 Task Force (IETF). Note that other groups may also distribute 35 working documents as Internet-Drafts. The list of current Internet- 36 Drafts is at http://datatracker.ietf.org/drafts/current/. 38 Internet-Drafts are draft documents valid for a maximum of six months 39 and may be updated, replaced, or obsoleted by other documents at any 40 time. It is inappropriate to use Internet-Drafts as reference 41 material or to cite them other than as "work in progress." 43 This Internet-Draft will expire on January 5, 2015. 45 Copyright Notice 47 Copyright (c) 2014 IETF Trust and the persons identified as the 48 document authors. All rights reserved. 50 This document is subject to BCP 78 and the IETF Trust's Legal 51 Provisions Relating to IETF Documents 52 (http://trustee.ietf.org/license-info) in effect on the date of 53 publication of this document. Please review these documents 54 carefully, as they describe your rights and restrictions with respect 55 to this document. Code Components extracted from this document must 56 include Simplified BSD License text as described in Section 4.e of 57 the Trust Legal Provisions and are provided without warranty as 58 described in the Simplified BSD License. 60 Table of Contents 62 1. RFC 2679 bis . . . . . . . . . . . . . . . . . . . . . . . . 3 63 2. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 4 64 2.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . 5 65 2.2. General Issues Regarding Time . . . . . . . . . . . . . . 6 66 3. A Singleton Definition for One-way Delay . . . . . . . . . . 7 67 3.1. Metric Name: . . . . . . . . . . . . . . . . . . . . . . 7 68 3.2. Metric Parameters: . . . . . . . . . . . . . . . . . . . 7 69 3.3. Metric Units: . . . . . . . . . . . . . . . . . . . . . . 7 70 3.4. Definition: . . . . . . . . . . . . . . . . . . . . . . . 7 71 3.5. Discussion: . . . . . . . . . . . . . . . . . . . . . . . 8 72 3.6. Methodologies: . . . . . . . . . . . . . . . . . . . . . 9 73 3.7. Errors and Uncertainties: . . . . . . . . . . . . . . . . 10 74 3.7.1. Errors or uncertainties related to Clocks . . . . . . 10 75 3.7.2. Errors or uncertainties related to Wire-time vs Host- 76 time . . . . . . . . . . . . . . . . . . . . . . . . 11 77 3.7.3. Calibration . . . . . . . . . . . . . . . . . . . . . 12 78 3.8. Reporting the metric: . . . . . . . . . . . . . . . . . . 14 79 3.8.1. Type-P . . . . . . . . . . . . . . . . . . . . . . . 14 80 3.8.2. Loss Threshold . . . . . . . . . . . . . . . . . . . 14 81 3.8.3. Calibration Results . . . . . . . . . . . . . . . . . 15 82 3.8.4. Path . . . . . . . . . . . . . . . . . . . . . . . . 15 83 4. A Definition for Samples of One-way Delay . . . . . . . . . . 15 84 4.1. Metric Name: . . . . . . . . . . . . . . . . . . . . . . 16 85 4.2. Metric Parameters: . . . . . . . . . . . . . . . . . . . 16 86 4.3. Metric Units: . . . . . . . . . . . . . . . . . . . . . . 16 87 4.4. Definition: . . . . . . . . . . . . . . . . . . . . . . . 16 88 4.5. Discussion: . . . . . . . . . . . . . . . . . . . . . . . 17 89 4.6. Methodologies: . . . . . . . . . . . . . . . . . . . . . 17 90 4.7. Errors and Uncertainties: . . . . . . . . . . . . . . . . 18 91 4.8. Reporting the metric: . . . . . . . . . . . . . . . . . . 18 92 5. Some Statistics Definitions for One-way Delay . . . . . . . . 18 93 5.1. Type-P-One-way-Delay-Percentile . . . . . . . . . . . . . 18 94 5.2. Type-P-One-way-Delay-Median . . . . . . . . . . . . . . . 19 95 5.3. Type-P-One-way-Delay-Minimum . . . . . . . . . . . . . . 19 96 5.4. Type-P-One-way-Delay-Inverse-Percentile . . . . . . . . . 20 97 6. Security Considerations . . . . . . . . . . . . . . . . . . . 20 98 7. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 21 99 8. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 21 100 9. Refetrences (temporary) . . . . . . . . . . . . . . . . . . . 21 101 10. References . . . . . . . . . . . . . . . . . . . . . . . . . 21 102 10.1. Normative References . . . . . . . . . . . . . . . . . . 21 103 10.2. Informative References . . . . . . . . . . . . . . . . . 22 104 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 23 106 1. RFC 2679 bis 108 The following text constitutes RFC 2769 bis proposed for advancement 109 on the IETF Standards Track. 111 [I-D.ietf-ippm-testplan-rfc2679] (now approved) provides the test 112 plan and results supporting [RFC2679] advancement along the standards 113 track, according to the process in [RFC6576]. The conclusions of 114 [I-D.ietf-ippm-testplan-rfc2679] list four minor modifications for 115 inclusion: 117 1. Section 6.2.3 of [I-D.ietf-ippm-testplan-rfc2679] asserts that 118 the assumption of post-processing to enforce a constant waiting 119 time threshold is compliant, and that the text of the RFC should 120 be revised slightly to include this point (see the last list item 121 of section 3.6, below). 123 2. Section 6.5 of [I-D.ietf-ippm-testplan-rfc2679] indicates that 124 Type-P-One-way-Delay-Inverse-Percentile statistic has been 125 ignored in both implementations, so it is a candidate for removal 126 or deprecation in RFC2679bis (this small discrepancy does not 127 affect candidacy for advancement) (see section 5.4, below). 129 3. The IETF has reached consensus on guidance for reporting metrics 130 in [RFC6703], and this memo should be referenced in RFC2679bis to 131 incorporate recent experience where appropriate (see the last 132 list item of section 3.6, section 3.8, and section 5 below). 134 4. There is currently one erratum with status "Held for document 135 update" for [RFC2679], and it appears this minor revision and 136 additional text should be incorporated in RFC2679bis (see section 137 5.1). 139 A small number of updates to the [RFC2679] text have been proposed 140 (by the current Editor) in the text below, principally to reference 141 key IPPM RFCs that were approved after [RFC2679]. 143 Section 5.4.4 of RFC 6390 suggests a common template for performance 144 metrics partially derived from previous IPPM and BMWG RFCs, but also 145 some new items. All of the RFC 6390 Normative points are covered, 146 but not quite in the same section names or orientation. Several of 147 the Informative points are covered. It is proposed to "grandfather- 148 in" bis RFCs w.r.t. RFC 6390 (keeping the familiar outline and 149 minimizing unnecessary differences), and focus efforts on applying 150 the template with new metric memos instead. 152 The publication of RFC 6921 suggests an area where this memo might be 153 updated. Packet transfer on Faster-Than-Light (FTL) networks could 154 result in negative delays and packet reordering, and both are covered 155 as possibilities in the current text (we note that this is an April 156 1st RFC). 158 2. Introduction 160 This memo defines a metric for one-way delay of packets across 161 Internet paths. It builds on notions introduced and discussed in the 162 IPPM Framework document, RFC 2330 [1]; the reader is assumed to be 163 familiar with that document. 165 This memo is intended to be parallel in structure to a companion 166 document for Packet Loss ("A One-way Packet Loss Metric for IPPM") 167 [2]. 169 Although RFC 2119 was written with protocols in mind, the key words 170 are used in this document for similar reasons. They are used to 171 ensure the results of measurements from two different implementations 172 are comparable, and to note instances when an implementation could 173 perturb the network. 175 The structure of the memo is as follows: 177 + A 'singleton' analytic metric, called Type-P-One-way-Delay, will be 178 introduced to measure a single observation of one-way delay. 180 + Using this singleton metric, a 'sample', called Type-P-One-way- 181 Delay-Poisson-Stream, will be introduced to measure a sequence of 182 singleton delays measured at times taken from a Poisson process. 184 + Using this sample, several 'statistics' of the sample will be 185 defined and discussed. This progression from singleton to sample to 186 statistics, with clear separation among them, is important. 188 Whenever a technical term from the IPPM Framework document is first 189 used in this memo, it will be tagged with a trailing asterisk. For 190 example, "term*" indicates that "term" is defined in the Framework. 192 2.1. Motivation 194 One-way delay of a Type-P* packet from a source host* to a 195 destination host is useful for several reasons: 197 + Some applications do not perform well (or at all) if end-to-end 198 delay between hosts is large relative to some threshold value. 200 + Erratic variation in delay makes it difficult (or impossible) to 201 support many real-time applications. 203 + The larger the value of delay, the more difficult it is for 204 transport-layer protocols to sustain high bandwidths. 206 + The minimum value of this metric provides an indication of the 207 delay due only to propagation and transmission delay. 209 + The minimum value of this metric provides an indication of the 210 delay that will likely be experienced when the path* traversed is 211 lightly loaded. 213 + Values of this metric above the minimum provide an indication of 214 the congestion present in the path. 216 The measurement of one-way delay instead of round-trip delay is 217 motivated by the following factors: 219 + In today's Internet, the path from a source to a destination may be 220 different than the path from the destination back to the source 221 ("asymmetric paths"), such that different sequences of routers are 222 used for the forward and reverse paths. Therefore round-trip 223 measurements actually measure the performance of two distinct paths 224 together. Measuring each path independently highlights the 225 performance difference between the two paths which may traverse 226 different Internet service providers, and even radically different 227 types of networks (for example, research versus commodity networks, 228 or ATM versus packet-over-SONET). 230 + Even when the two paths are symmetric, they may have radically 231 different performance characteristics due to asymmetric queueing. 233 + Performance of an application may depend mostly on the performance 234 in one direction. For example, a file transfer using TCP may depend 235 more on the performance in the direction that data flows (queue 236 occupation tends to grow in this direction, possibly dominating the 237 round-trip delay), rather than the direction in which 238 acknowledgements travel. 240 + In quality-of-service (QoS) enabled networks, provisioning in one 241 direction may be radically different than provisioning in the reverse 242 direction, and thus the QoS guarantees differ. Measuring the paths 243 independently allows the verification of both guarantees. 245 It is outside the scope of this document to say precisely how delay 246 metrics would be applied to specific problems. 248 2.2. General Issues Regarding Time 250 {Comment: the terminology below differs from that defined by ITU-T 251 documents (e.g., G.810, "Definitions and terminology for 252 synchronization networks" and I.356, "B-ISDN ATM layer cell transfer 253 performance"), but is consistent with the IPPM Framework document. 254 In general, these differences derive from the different backgrounds; 255 the ITU-T documents historically have a telephony origin, while the 256 authors of this document (and the Framework) have a computer systems 257 background. Although the terms defined below have no direct 258 equivalent in the ITU-T definitions, after our definitions we will 259 provide a rough mapping. However, note one potential confusion: our 260 definition of "clock" is the computer operating systems definition 261 denoting a time-of-day clock, while the ITU-T definition of clock 262 denotes a frequency reference.} 264 Whenever a time (i.e., a moment in history) is mentioned here, it is 265 understood to be measured in seconds (and fractions) relative to UTC. 267 As described more fully in the Framework document, there are four 268 distinct, but related notions of clock uncertainty: 270 synchronization* 272 measures the extent to which two clocks agree on what time it is. 273 For example, the clock on one host might be 5.4 msec ahead of the 274 clock on a second host. {Comment: A rough ITU-T equivalent is "time 275 error".} 277 accuracy* 279 measures the extent to which a given clock agrees with UTC. For 280 example, the clock on a host might be 27.1 msec behind UTC. {Comment: 281 A rough ITU-T equivalent is "time error from UTC".} 283 resolution* 284 measures the precision of a given clock. For example, the clock on 285 an old Unix host might tick only once every 10 msec, and thus have a 286 resolution of only 10 msec. {Comment: A very rough ITU-T equivalent 287 is "sampling period".} 289 skew* 291 measures the change of accuracy, or of synchronization, with time. 292 For example, the clock on a given host might gain 1.3 msec per hour 293 and thus be 27.1 msec behind UTC at one time and only 25.8 msec an 294 hour later. In this case, we say that the clock of the given host 295 has a skew of 1.3 msec per hour relative to UTC, which threatens 296 accuracy. We might also speak of the skew of one clock relative to 297 another clock, which threatens synchronization. {Comment: A rough 298 ITU-T equivalent is "time drift".} 300 3. A Singleton Definition for One-way Delay 302 3.1. Metric Name: 304 Type-P-One-way-Delay 306 3.2. Metric Parameters: 308 + Src, the IP address of a host 310 + Dst, the IP address of a host 312 + T, a time 314 + Tmax, a loss threshold waiting time 316 3.3. Metric Units: 318 The value of a Type-P-One-way-Delay is either a real number, or an 319 undefined (informally, infinite) number of seconds. 321 3.4. Definition: 323 For a real number dT, >>the *Type-P-One-way-Delay* from Src to Dst at 324 T is dT<< means that Src sent the first bit of a Type-P packet to Dst 325 at wire-time* T and that Dst received the last bit of that packet at 326 wire-time T+dT. 328 >>The *Type-P-One-way-Delay* from Src to Dst at T is undefined 329 (informally, infinite)<< means that Src sent the first bit of a 330 Type-P packet to Dst at wire-time T and that Dst did not receive that 331 packet (within the loss threshold waiting time, Tmax). 333 Suggestions for what to report along with metric values appear in 334 Section 3.8 after a discussion of the metric, methodologies for 335 measuring the metric, and error analysis. 337 3.5. Discussion: 339 Type-P-One-way-Delay is a relatively simple analytic metric, and one 340 that we believe will afford effective methods of measurement. 342 The following issues are likely to come up in practice: 344 + Real delay values will be positive. Therefore, it does not make 345 sense to report a negative value as a real delay. However, an 346 individual zero or negative delay value might be useful as part of a 347 stream when trying to discover a distribution of a stream of delay 348 values. 350 + Since delay values will often be as low as the 100 usec to 10 msec 351 range, it will be important for Src and Dst to synchronize very 352 closely. GPS systems afford one way to achieve synchronization to 353 within several 10s of usec. Ordinary application of NTP may allow 354 synchronization to within several msec, but this depends on the 355 stability and symmetry of delay properties among those NTP agents 356 used, and this delay is what we are trying to measure. A combination 357 of some GPS-based NTP servers and a conservatively designed and 358 deployed set of other NTP servers should yield good results, but this 359 is yet to be tested. 361 + A given methodology will have to include a way to determine whether 362 a delay value is infinite or whether it is merely very large (and the 363 packet is yet to arrive at Dst). As noted by Mahdavi and Paxson [4], 364 simple upper bounds (such as the 255 seconds theoretical upper bound 365 on the lifetimes of IP packets [5]) could be used, but good 366 engineering, including an understanding of packet lifetimes, will be 367 needed in practice. {Comment: Note that, for many applications of 368 these metrics, the harm in treating a large delay as infinite might 369 be zero or very small. A TCP data packet, for example, that arrives 370 only after several multiples of the RTT may as well have been lost.} 372 + If the packet is duplicated along the path (or paths) so that 373 multiple non-corrupt copies arrive at the destination, then the 374 packet is counted as received, and the first copy to arrive 375 determines the packet's one-way delay. 377 + If the packet is fragmented and if, for whatever reason, reassembly 378 does not occur, then the packet will be deemed lost. 380 3.6. Methodologies: 382 As with other Type-P-* metrics, the detailed methodology will depend 383 on the Type-P (e.g., protocol number, UDP/TCP port number, size, 384 precedence). 386 Generally, for a given Type-P, the methodology would proceed as 387 follows: 389 + Arrange that Src and Dst are synchronized; that is, that they have 390 clocks that are very closely synchronized with each other and each 391 fairly close to the actual time. 393 + At the Src host, select Src and Dst IP addresses, and form a test 394 packet of Type-P with these addresses. Any 'padding' portion of the 395 packet needed only to make the test packet a given size should be 396 filled with randomized bits to avoid a situation in which the 397 measured delay is lower than it would otherwise be due to compression 398 techniques along the path. Note that use of transport layer 399 encryption will counteract the deployment of network-based analysis 400 and may reduce the adoption of payload optimizations like 401 compression. 403 + At the Dst host, arrange to receive the packet. 405 + At the Src host, place a timestamp in the prepared Type-P packet, 406 and send it towards Dst. 408 + If the packet arrives within a reasonable period of time, take a 409 timestamp as soon as possible upon the receipt of the packet. By 410 subtracting the two timestamps, an estimate of one-way delay can be 411 computed. Error analysis of a given implementation of the method 412 must take into account the closeness of synchronization between Src 413 and Dst. If the delay between Src's timestamp and the actual sending 414 of the packet is known, then the estimate could be adjusted by 415 subtracting this amount; uncertainty in this value must be taken into 416 account in error analysis. Similarly, if the delay between the 417 actual receipt of the packet and Dst's timestamp is known, then the 418 estimate could be adjusted by subtracting this amount; uncertainty in 419 this value must be taken into account in error analysis. See the 420 next section, "Errors and Uncertainties", for a more detailed 421 discussion. 423 + If the packet fails to arrive within a reasonable period of time, 424 the one-way delay is taken to be undefined (informally, infinite). 425 Note that the threshold of 'reasonable' is a parameter of the 426 methodology. These points are examined in detail in [RFC6703], 427 including analysis preferences to assign undefined delay to packets 428 that fail to arrive with the difficulties emerging from the informal 429 "infinite delay" assignment, and an estimation of an upper bound on 430 waiting time for packets in transit. Further, enforcing a specific 431 constant waiting time on stored singletons of one-way delay is 432 compliant with this specification and may allow the results to serve 433 more than one reporting audience. 435 Issues such as the packet format, the means by which Dst knows when 436 to expect the test packet, and the means by which Src and Dst are 437 synchronized are outside the scope of this document. {Comment: We 438 plan to document elsewhere our own work in describing such more 439 detailed implementation techniques and we encourage others to as 440 well.} 442 3.7. Errors and Uncertainties: 444 The description of any specific measurement method should include an 445 accounting and analysis of various sources of error or uncertainty. 446 The Framework document provides general guidance on this point, but 447 we note here the following specifics related to delay metrics: 449 + Errors or uncertainties due to uncertainties in the clocks of the 450 Src and Dst hosts. 452 + Errors or uncertainties due to the difference between 'wire time' 453 and 'host time'. 455 In addition, the loss threshold may affect the results. Each of 456 these are discussed in more detail below, along with a section 457 ("Calibration") on accounting for these errors and uncertainties. 459 3.7.1. Errors or uncertainties related to Clocks 461 The uncertainty in a measurement of one-way delay is related, in 462 part, to uncertainties in the clocks of the Src and Dst hosts. In 463 the following, we refer to the clock used to measure when the packet 464 was sent from Src as the source clock, we refer to the clock used to 465 measure when the packet was received by Dst as the destination clock, 466 we refer to the observed time when the packet was sent by the source 467 clock as Tsource, and the observed time when the packet was received 468 by the destination clock as Tdest. Alluding to the notions of 469 synchronization, accuracy, resolution, and skew mentioned in the 470 Introduction, we note the following: 472 + Any error in the synchronization between the source clock and the 473 destination clock will contribute to error in the delay measurement. 474 We say that the source clock and the destination clock have a 475 synchronization error of Tsynch if the source clock is Tsynch ahead 476 of the destination clock. Thus, if we know the value of Tsynch 477 exactly, we could correct for clock synchronization by adding Tsynch 478 to the uncorrected value of Tdest-Tsource. 480 + The accuracy of a clock is important only in identifying the time 481 at which a given delay was measured. Accuracy, per se, has no 482 importance to the accuracy of the measurement of delay. When 483 computing delays, we are interested only in the differences between 484 clock values, not the values themselves. 486 + The resolution of a clock adds to uncertainty about any time 487 measured with it. Thus, if the source clock has a resolution of 10 488 msec, then this adds 10 msec of uncertainty to any time value 489 measured with it. We will denote the resolution of the source clock 490 and the destination clock as Rsource and Rdest, respectively. 492 + The skew of a clock is not so much an additional issue as it is a 493 realization of the fact that Tsynch is itself a function of time. 494 Thus, if we attempt to measure or to bound Tsynch, this needs to be 495 done periodically. Over some periods of time, this function can be 496 approximated as a linear function plus some higher order terms; in 497 these cases, one option is to use knowledge of the linear component 498 to correct the clock. Using this correction, the residual Tsynch is 499 made smaller, but remains a source of uncertainty that must be 500 accounted for. We use the function Esynch(t) to denote an upper 501 bound on the uncertainty in synchronization. Thus, |Tsynch(t)| <= 502 Esynch(t). 504 Taking these items together, we note that naive computation Tdest- 505 Tsource will be off by Tsynch(t) +/- (Rsource + Rdest). Using the 506 notion of Esynch(t), we note that these clock-related problems 507 introduce a total uncertainty of Esynch(t)+ Rsource + Rdest. This 508 estimate of total clock-related uncertainty should be included in the 509 error/uncertainty analysis of any measurement implementation. 511 3.7.2. Errors or uncertainties related to Wire-time vs Host-time 513 As we have defined one-way delay, we would like to measure the time 514 between when the test packet leaves the network interface of Src and 515 when it (completely) arrives at the network interface of Dst, and we 516 refer to these as "wire times." If the timings are themselves 517 performed by software on Src and Dst, however, then this software can 518 only directly measure the time between when Src grabs a timestamp 519 just prior to sending the test packet and when Dst grabs a timestamp 520 just after having received the test packet, and we refer to these two 521 points as "host times". 523 We note that some systems perform host time stamping on the network 524 interface hardware, in an attempt to minimize the difference from 525 wire times. 527 To the extent that the difference between wire time and host time is 528 accurately known, this knowledge can be used to correct for host time 529 measurements and the corrected value more accurately estimates the 530 desired (wire time) metric. 532 To the extent, however, that the difference between wire time and 533 host time is uncertain, this uncertainty must be accounted for in an 534 analysis of a given measurement method. We denote by Hsource an 535 upper bound on the uncertainty in the difference between wire time 536 and host time on the Src host, and similarly define Hdest for the Dst 537 host. We then note that these problems introduce a total uncertainty 538 of Hsource+Hdest. This estimate of total wire-vs-host uncertainty 539 should be included in the error/uncertainty analysis of any 540 measurement implementation. 542 3.7.3. Calibration 544 Generally, the measured values can be decomposed as follows: 546 measured value = true value + systematic error + random error 548 If the systematic error (the constant bias in measured values) can be 549 determined, it can be compensated for in the reported results. 551 reported value = measured value - systematic error 553 therefore 555 reported value = true value + random error 557 The goal of calibration is to determine the systematic and random 558 error generated by the instruments themselves in as much detail as 559 possible. At a minimum, a bound ("e") should be found such that the 560 reported value is in the range (true value - e) to (true value + e) 561 at least 95 percent of the time. We call "e" the calibration error 562 for the measurements. It represents the degree to which the values 563 produced by the measurement instrument are repeatable; that is, how 564 closely an actual delay of 30 ms is reported as 30 ms. {Comment: 95 565 percent was chosen because (1) some confidence level is desirable to 566 be able to remove outliers, which will be found in measuring any 567 physical property; (2) a particular confidence level should be 568 specified so that the results of independent implementations can be 569 compared; and (3) even with a prototype user-level implementation, 570 95% was loose enough to exclude outliers.} 571 From the discussion in the previous two sections, the error in 572 measurements could be bounded by determining all the individual 573 uncertainties, and adding them together to form 575 Esynch(t) + Rsource + Rdest + Hsource + Hdest. 577 However, reasonable bounds on both the clock-related uncertainty 578 captured by the first three terms and the host-related uncertainty 579 captured by the last two terms should be possible by careful design 580 techniques and calibrating the instruments using a known, isolated, 581 network in a lab. 583 For example, the clock-related uncertainties are greatly reduced 584 through the use of a GPS time source. The sum of Esynch(t) + Rsource 585 + Rdest is small, and is also bounded for the duration of the 586 measurement because of the global time source. 588 The host-related uncertainties, Hsource + Hdest, could be bounded by 589 connecting two instruments back-to-back with a high-speed serial link 590 or isolated LAN segment. In this case, repeated measurements are 591 measuring the same one-way delay. 593 If the test packets are small, such a network connection has a 594 minimal delay that may be approximated by zero. The measured delay 595 therefore contains only systematic and random error in the 596 instrumentation. The "average value" of repeated measurements is the 597 systematic error, and the variation is the random error. 599 One way to compute the systematic error, and the random error to a 600 95% confidence is to repeat the experiment many times - at least 601 hundreds of tests. The systematic error would then be the median. 602 The random error could then be found by removing the systematic error 603 from the measured values. The 95% confidence interval would be the 604 range from the 2.5th percentile to the 97.5th percentile of these 605 deviations from the true value. The calibration error "e" could then 606 be taken to be the largest absolute value of these two numbers, plus 607 the clock-related uncertainty. {Comment: as described, this bound is 608 relatively loose since the uncertainties are added, and the absolute 609 value of the largest deviation is used. As long as the resulting 610 value is not a significant fraction of the measured values, it is a 611 reasonable bound. If the resulting value is a significant fraction 612 of the measured values, then more exact methods will be needed to 613 compute the calibration error.} 615 Note that random error is a function of measurement load. For 616 example, if many paths will be measured by one instrument, this might 617 increase interrupts, process scheduling, and disk I/O (for example, 618 recording the measurements), all of which may increase the random 619 error in measured singletons. Therefore, in addition to minimal load 620 measurements to find the systematic error, calibration measurements 621 should be performed with the same measurement load that the 622 instruments will see in the field. 624 We wish to reiterate that this statistical treatment refers to the 625 calibration of the instrument; it is used to "calibrate the meter 626 stick" and say how well the meter stick reflects reality. 628 In addition to calibrating the instruments for finite one-way delay, 629 two checks should be made to ensure that packets reported as losses 630 were really lost. First, the threshold for loss should be verified. 631 In particular, ensure the "reasonable" threshold is reasonable: that 632 it is very unlikely a packet will arrive after the threshold value, 633 and therefore the number of packets lost over an interval is not 634 sensitive to the error bound on measurements. Second, consider the 635 possibility that a packet arrives at the network interface, but is 636 lost due to congestion on that interface or to other resource 637 exhaustion (e.g. buffers) in the instrument. 639 3.8. Reporting the metric: 641 The calibration and context in which the metric is measured MUST be 642 carefully considered, and SHOULD always be reported along with metric 643 results. We now present four items to consider: the Type-P of test 644 packets, the threshold of infinite delay (if any), error calibration, 645 and the path traversed by the test packets. This list is not 646 exhaustive; any additional information that could be useful in 647 interpreting applications of the metrics should also be reported (see 648 [RFC6703] for extensive discussion of reporting considerations for 649 different audiences). 651 3.8.1. Type-P 653 As noted in the Framework document [1], the value of the metric may 654 depend on the type of IP packets used to make the measurement, or 655 "type-P". The value of Type-P-One-way-Delay could change if the 656 protocol (UDP or TCP), port number, size, or arrangement for special 657 treatment (e.g., IP precedence or RSVP) changes. The exact Type-P 658 used to make the measurements MUST be accurately reported. 660 3.8.2. Loss Threshold 662 In addition, the threshold (or methodology to distinguish) between a 663 large finite delay and loss MUST be reported. 665 3.8.3. Calibration Results 667 + If the systematic error can be determined, it SHOULD be removed 668 from the measured values. 670 + You SHOULD also report the calibration error, e, such that the true 671 value is the reported value plus or minus e, with 95% confidence (see 672 the last section.) 674 + If possible, the conditions under which a test packet with finite 675 delay is reported as lost due to resource exhaustion on the 676 measurement instrument SHOULD be reported. 678 3.8.4. Path 680 Finally, the path traversed by the packet SHOULD be reported, if 681 possible. In general it is impractical to know the precise path a 682 given packet takes through the network. The precise path may be 683 known for certain Type-P on short or stable paths. If Type-P 684 includes the record route (or loose-source route) option in the IP 685 header, and the path is short enough, and all routers* on the path 686 support record (or loose-source) route, then the path will be 687 precisely recorded. This is impractical because the route must be 688 short enough, many routers do not support (or are not configured for) 689 record route, and use of this feature would often artificially worsen 690 the performance observed by removing the packet from common-case 691 processing. However, partial information is still valuable context. 692 For example, if a host can choose between two links* (and hence two 693 separate routes from Src to Dst), then the initial link used is 694 valuable context. {Comment: For example, with Merit's NetNow setup, a 695 Src on one NAP can reach a Dst on another NAP by either of several 696 different backbone networks.} 698 4. A Definition for Samples of One-way Delay 700 Given the singleton metric Type-P-One-way-Delay, we now define one 701 particular sample of such singletons. The idea of the sample is to 702 select a particular binding of the parameters Src, Dst, and Type-P, 703 then define a sample of values of parameter T. The means for 704 defining the values of T is to select a beginning time T0, a final 705 time Tf, and an average rate lambda, then define a pseudo-random 706 Poisson process of rate lambda, whose values fall between T0 and Tf. 707 The time interval between successive values of T will then average 1/ 708 lambda. 710 {Comment: Note that Poisson sampling is only one way of defining a 711 sample. Poisson has the advantage of limiting bias, but other 712 methods of sampling might be appropriate for different situations. 714 We encourage others who find such appropriate cases to use this 715 general framework and submit their sampling method for 716 standardization.} 718 >>> Editor proposal: Add ref to RFC 3432 Periodic sampling above. 720 4.1. Metric Name: 722 Type-P-One-way-Delay-Poisson-Stream 724 4.2. Metric Parameters: 726 + Src, the IP address of a host 728 + Dst, the IP address of a host 730 + T0, a time 732 + Tf, a time 734 + lambda, a rate in reciprocal seconds 736 4.3. Metric Units: 738 A sequence of pairs; the elements of each pair are: 740 + T, a time, and 742 + dT, either a real number or an undefined number of seconds. 744 The values of T in the sequence are monotonic increasing. Note that 745 T would be a valid parameter to Type-P-One-way-Delay, and that dT 746 would be a valid value of Type-P-One-way-Delay. 748 4.4. Definition: 750 Given T0, Tf, and lambda, we compute a pseudo-random Poisson process 751 beginning at or before T0, with average arrival rate lambda, and 752 ending at or after Tf. Those time values greater than or equal to T0 753 and less than or equal to Tf are then selected. At each of the times 754 in this process, we obtain the value of Type-P-One-way-Delay at this 755 time. The value of the sample is the sequence made up of the 756 resulting pairs. If there are no such pairs, the 757 sequence is of length zero and the sample is said to be empty. 759 4.5. Discussion: 761 The reader should be familiar with the in-depth discussion of Poisson 762 sampling in the Framework document [1], which includes methods to 763 compute and verify the pseudo-random Poisson process. 765 We specifically do not constrain the value of lambda, except to note 766 the extremes. If the rate is too large, then the measurement traffic 767 will perturb the network, and itself cause congestion. If the rate 768 is too small, then you might not capture interesting network 769 behavior. {Comment: We expect to document our experiences with, and 770 suggestions for, lambda elsewhere, culminating in a "best current 771 practices" document.} 773 Since a pseudo-random number sequence is employed, the sequence of 774 times, and hence the value of the sample, is not fully specified. 775 Pseudo-random number generators of good quality will be needed to 776 achieve the desired qualities. 778 The sample is defined in terms of a Poisson process both to avoid the 779 effects of self-synchronization and also capture a sample that is 780 statistically as unbiased as possible. {Comment: there is, of course, 781 no claim that real Internet traffic arrives according to a Poisson 782 arrival process.} The Poisson process is used to schedule the delay 783 measurements. The test packets will generally not arrive at Dst 784 according to a Poisson distribution, since they are influenced by the 785 network. 787 All the singleton Type-P-One-way-Delay metrics in the sequence will 788 have the same values of Src, Dst, and Type-P. 790 Note also that, given one sample that runs from T0 to Tf, and given 791 new time values T0' and Tf' such that T0 <= T0' <= Tf' <= Tf, the 792 subsequence of the given sample whose time values fall between T0' 793 and Tf' are also a valid Type-P-One-way-Delay-Poisson-Stream sample. 795 4.6. Methodologies: 797 The methodologies follow directly from: 799 + the selection of specific times, using the specified Poisson 800 arrival process, and 802 + the methodologies discussion already given for the singleton Type- 803 P-One-way-Delay metric. 805 Care must, of course, be given to correctly handle out-of-order 806 arrival of test packets; it is possible that the Src could send one 807 test packet at TS[i], then send a second one (later) at TS[i+1], 808 while the Dst could receive the second test packet at TR[i+1], and 809 then receive the first one (later) at TR[i]. 811 >>> Editor proposal: Add ref to RFC 4737 Reordering metric above. 813 4.7. Errors and Uncertainties: 815 In addition to sources of errors and uncertainties associated with 816 methods employed to measure the singleton values that make up the 817 sample, care must be given to analyze the accuracy of the Poisson 818 process with respect to the wire-times of the sending of the test 819 packets. Problems with this process could be caused by several 820 things, including problems with the pseudo-random number techniques 821 used to generate the Poisson arrival process, or with jitter in the 822 value of Hsource (mentioned above as uncertainty in the singleton 823 delay metric). The Framework document shows how to use the Anderson- 824 Darling test to verify the accuracy of a Poisson process over small 825 time frames. {Comment: The goal is to ensure that test packets are 826 sent "close enough" to a Poisson schedule, and avoid periodic 827 behavior.} 829 4.8. Reporting the metric: 831 You MUST report the calibration and context for the underlying 832 singletons along with the stream. (See "Reporting the metric" for 833 Type-P-One-way-Delay.) 835 5. Some Statistics Definitions for One-way Delay 837 Given the sample metric Type-P-One-way-Delay-Poisson-Stream, we now 838 offer several statistics of that sample. These statistics are 839 offered mostly to be illustrative of what could be done. See 840 [RFC6703] for additional discussion of statistics that are relevant 841 to different audiences. 843 5.1. Type-P-One-way-Delay-Percentile 845 Given a Type-P-One-way-Delay-Poisson-Stream and a percent X between 846 0% and 100%, the Xth percentile of all the dT values in the Stream. 847 In computing this percentile, undefined values are treated as 848 infinitely large. Note that this means that the percentile could 849 thus be undefined (informally, infinite). In addition, the Type-P- 850 One-way-Delay-Percentile is undefined if the sample is empty. 852 Example: suppose we take a sample and the results are: 854 Stream1 = < 855 857 859 861 863 865 > 867 Then the 50th percentile would be 110 msec, since 90 msec and 100 868 msec are smaller and 500 msec and 'undefined' are larger. See 869 Section 11.3 of [1] for computing percentiles. 871 Note that if the possibility that a packet with finite delay is 872 reported as lost is significant, then a high percentile (90th or 873 95th) might be reported as infinite instead of finite. 875 5.2. Type-P-One-way-Delay-Median 877 Given a Type-P-One-way-Delay-Poisson-Stream, the median of all the dT 878 values in the Stream. In computing the median, undefined values are 879 treated as infinitely large. As with Type-P-One-way-Delay- 880 Percentile, Type-P-One-way-Delay-Median is undefined if the sample is 881 empty. 883 As noted in the Framework document, the median differs from the 50th 884 percentile only when the sample contains an even number of values, in 885 which case the mean of the two central values is used. 887 Example: suppose we take a sample and the results are: 889 Stream2 = < > 892 Then the median would be 105 msec, the mean of 100 msec and 110 msec, 893 the two central values. 895 5.3. Type-P-One-way-Delay-Minimum 897 Given a Type-P-One-way-Delay-Poisson-Stream, the minimum of all the 898 dT values in the Stream. In computing this, undefined values are 899 treated as infinitely large. Note that this means that the minimum 900 could thus be undefined (informally, infinite) if all the dT values 901 are undefined. In addition, the Type-P-One-way-Delay-Minimum is 902 undefined if the sample is empty. 904 In the above example, the minimum would be 90 msec. 906 5.4. Type-P-One-way-Delay-Inverse-Percentile 908 Note: This statistic is deprecated in this version of the memo 909 because of lack of use. 911 Given a Type-P-One-way-Delay-Poisson-Stream and a time duration 912 threshold, the fraction of all the dT values in the Stream less than 913 or equal to the threshold. The result could be as low as 0% (if all 914 the dT values exceed threshold) or as high as 100%. Type-P-One-way- 915 Delay-Inverse-Percentile is undefined if the sample is empty. 917 In the above example, the Inverse-Percentile of 103 msec would be 918 50%. 920 6. Security Considerations 922 Conducting Internet measurements raises both security and privacy 923 concerns. This memo does not specify an implementation of the 924 metrics, so it does not directly affect the security of the Internet 925 nor of applications which run on the Internet. However, 926 implementations of these metrics must be mindful of security and 927 privacy concerns. 929 There are two types of security concerns: potential harm caused by 930 the measurements, and potential harm to the measurements. The 931 measurements could cause harm because they are active, and inject 932 packets into the network. The measurement parameters MUST be 933 carefully selected so that the measurements inject trivial amounts of 934 additional traffic into the networks they measure. If they inject 935 "too much" traffic, they can skew the results of the measurement, and 936 in extreme cases cause congestion and denial of service. 938 The measurements themselves could be harmed by routers giving 939 measurement traffic a different priority than "normal" traffic, or by 940 an attacker injecting artificial measurement traffic. If routers can 941 recognize measurement traffic and treat it separately, the 942 measurements will not reflect actual user traffic. If an attacker 943 injects artificial traffic that is accepted as legitimate, the loss 944 rate will be artificially lowered. Therefore, the measurement 945 methodologies SHOULD include appropriate techniques to reduce the 946 probability measurement traffic can be distinguished from "normal" 947 traffic. Authentication techniques, such as digital signatures, may 948 be used where appropriate to guard against injected traffic attacks. 950 The privacy concerns of network measurement are limited by the active 951 measurements described in this memo. Unlike passive measurements, 952 there can be no release of existing user data. 954 7. IANA Considerations 956 This memo makes no requests of IANA. 958 8. Acknowledgements 960 Special thanks are due to Vern Paxson of Lawrence Berkeley Labs for 961 his helpful comments on issues of clock uncertainty and statistics. 962 Thanks also to Garry Couch, Will Leland, Andy Scherrer, Sean Shapira, 963 and Roland Wittig for several useful suggestions. 965 9. Refetrences (temporary) 967 [1] Paxson, V., Almes, G., Mahdavi, J. and M. Mathis, "Framework for 968 IP Performance Metrics", RFC 2330, May 1998. 970 [2] Almes, G., Kalidindi, S. and M. Zekauskas, "A One-way Packet 971 Loss Metric for IPPM", RFC 2680, September 1999. 973 [3] Mills, D., "Network Time Protocol (v3)", RFC 1305, April 1992. 975 [4] Mahdavi J. and V. Paxson, "IPPM Metrics for Measuring 976 Connectivity", RFC 2678, September 1999. 978 [5] Postel, J., "Internet Protocol", STD 5, RFC 791, September 1981. 980 [6] Bradner, S., "Key words for use in RFCs to Indicate Requirement 981 Levels", BCP 14, RFC 2119, March 1997. 983 [7] Bradner, S., "The Internet Standards Process -- Revision 3", BCP 984 9, RFC 2026, October 1996. 986 10. References 988 10.1. Normative References 990 [RFC2026] Bradner, S., "The Internet Standards Process -- Revision 991 3", BCP 9, RFC 2026, October 1996. 993 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 994 Requirement Levels", BCP 14, RFC 2119, March 1997. 996 [RFC2330] Paxson, V., Almes, G., Mahdavi, J., and M. Mathis, 997 "Framework for IP Performance Metrics", RFC 2330, May 998 1998. 1000 [RFC2679] Almes, G., Kalidindi, S., and M. Zekauskas, "A One-way 1001 Delay Metric for IPPM", RFC 2679, September 1999. 1003 [RFC2680] Almes, G., Kalidindi, S., and M. Zekauskas, "A One-way 1004 Packet Loss Metric for IPPM", RFC 2680, September 1999. 1006 [RFC3432] Raisanen, V., Grotefeld, G., and A. Morton, "Network 1007 performance measurement with periodic streams", RFC 3432, 1008 November 2002. 1010 [RFC4656] Shalunov, S., Teitelbaum, B., Karp, A., Boote, J., and M. 1011 Zekauskas, "A One-way Active Measurement Protocol 1012 (OWAMP)", RFC 4656, September 2006. 1014 [RFC5357] Hedayat, K., Krzanowski, R., Morton, A., Yum, K., and J. 1015 Babiarz, "A Two-Way Active Measurement Protocol (TWAMP)", 1016 RFC 5357, October 2008. 1018 [RFC5657] Dusseault, L. and R. Sparks, "Guidance on Interoperation 1019 and Implementation Reports for Advancement to Draft 1020 Standard", BCP 9, RFC 5657, September 2009. 1022 [RFC5835] Morton, A. and S. Van den Berghe, "Framework for Metric 1023 Composition", RFC 5835, April 2010. 1025 [RFC6049] Morton, A. and E. Stephan, "Spatial Composition of 1026 Metrics", RFC 6049, January 2011. 1028 [RFC6576] Geib, R., Morton, A., Fardid, R., and A. Steinmitz, "IP 1029 Performance Metrics (IPPM) Standard Advancement Testing", 1030 BCP 176, RFC 6576, March 2012. 1032 [RFC6703] Morton, A., Ramachandran, G., and G. Maguluri, "Reporting 1033 IP Network Performance Metrics: Different Points of View", 1034 RFC 6703, August 2012. 1036 10.2. Informative References 1038 [ADK] Scholz, F. and M. Stephens, "K-sample Anderson-Darling 1039 Tests of fit, for continuous and discrete cases", 1040 University of Washington, Technical Report No. 81, May 1041 1986. 1043 [I-D.ietf-ippm-testplan-rfc2679] 1044 Ciavattone, L., Geib, R., Morton, A., and M. Wieser, "Test 1045 Plan and Results Supporting Advancement of RFC 2679 on the 1046 Standards Track", draft-ietf-ippm-testplan-rfc2679-03 1047 (work in progress), September 2012. 1049 [RFC3931] Lau, J., Townsley, M., and I. Goyret, "Layer Two Tunneling 1050 Protocol - Version 3 (L2TPv3)", RFC 3931, March 2005. 1052 Authors' Addresses 1054 Guy Almes 1055 Texas A&M 1057 Sunil Kalidindi 1058 Ixia 1060 Matt Zekauskas 1061 Internet2 1063 Email: matt@internet2.edu 1065 Al Morton (editor) 1066 AT&T Labs 1067 200 Laurel Avenue South 1068 Middletown, NJ 07748 1069 USA 1071 Phone: +1 732 420 1571 1072 Fax: +1 732 368 1192 1073 Email: acmorton@att.com 1074 URI: http://home.comcast.net/~acmacm/