idnits 2.17.1 draft-ietf-ippm-npmps-02.txt: ** The Abstract section seems to be numbered Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- ** Looks like you're using RFC 2026 boilerplate. This must be updated to follow RFC 3978/3979, as updated by RFC 4748. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- ** The document seems to lack a 1id_guidelines paragraph about the list of Shadow Directories. ** The document is more than 15 pages and seems to lack a Table of Contents. == No 'Intended status' indicated for this document; assuming Proposed Standard Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- ** 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.) ** The document seems to lack separate sections for Informative/Normative References. All references will be assumed normative when checking for downward references. ** There are 25 instances of too long lines in the document, the longest one being 3 characters in excess of 72. ** There is 1 instance of lines with control characters in the document. ** The abstract seems to contain references ([2], [1]), which it shouldn't. Please replace those with straight textual mentions of the documents in question. Miscellaneous warnings: ---------------------------------------------------------------------------- == Line 52 has weird spacing: '...vant to appli...' == Line 66 has weird spacing: '... metric using...' == Line 278 has weird spacing: '...han the round...' == Line 297 has weird spacing: '...ization and t...' -- The document seems to lack a disclaimer for pre-RFC5378 work, but may have content which was first submitted before 10 November 2008. If you have contacted all the original authors and they are all willing to grant the BCP78 rights to the IETF Trust, then this is fine, and you can ignore this comment. If not, you may need to add the pre-RFC5378 disclaimer. (See the Legal Provisions document at https://trustee.ietf.org/license-info for more information.) -- The document date (July 2000) is 8678 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) ** Downref: Normative reference to an Informational RFC: RFC 2330 (ref. '1') ** Obsolete normative reference: RFC 2679 (ref. '2') (Obsoleted by RFC 7679) -- Possible downref: Non-RFC (?) normative reference: ref. '3' -- Possible downref: Non-RFC (?) normative reference: ref. '5' Summary: 11 errors (**), 0 flaws (~~), 5 warnings (==), 4 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 1 Network Working Group V. Raisanen 2 INTERNET-DRAFT Nokia 3 Expiration Date: January 2001 G. Grotefeld 4 Motorola 5 July 2000 7 Network performance measurement for periodic streams 8 10 1. Status of this Memo 12 This document is an Internet-Draft and is in full conformance with 13 all provisions of Section 10 of RFC2026. 15 Internet-Drafts are working documents of the Internet Engineering 16 Task Force (IETF), its areas, and its working groups. Note that 17 other groups may also distribute working documents as Internet- 18 Drafts. 20 Internet-Drafts are draft documents valid for a maximum of six months 21 and may be updated, replaced, or obsoleted by other documents at any 22 time. It is inappropriate to use Internet-Drafts as reference 23 material or to cite them other than as "work in progress." 25 The list of current Internet-Drafts can be accessed at 26 http://www.ietf.org/ietf/1id-abstracts.txt 28 The list of Internet-Draft shadow directories can be accessed at 29 http://www.ietf.org/shadow.html 31 This memo provides information for the Internet community. This 32 memo does not specify an Internet standard of any 33 kind. Distribution of this memo is unlimited. 35 2. Abstract 37 This document describes some of the issues associated with 38 application-level measurements of network performance for periodic 39 streams. An example application would be the testing of Dst-Src routes 40 for use as bearer for multimedia streams. In this document, 41 the reader is assumed to be familiar with the terminology of the 42 Framework for IP Performance Metrics RFC 2330 [1]. This document is 43 parallel to A One-way Delay Metric for IPPM RFC 2679 [2]. A sample 44 metric is described that is suitable for application-level measurement 45 for streaming multimedia over IP. Using such a measurement, 46 transmission service of a network is probed with a traffic stream 47 similar to that of the application of interest, which is likely to be 48 very dissimilar to the Poisson inter-arrival interval described in [2]. 50 3. Introduction 52 This document discusses concepts relevant to application-level 53 performance measurements of an IP network. The original driver for 54 this work is Quality of Service of interactive periodic streams such 55 as multimedia conference over IP, but the idea of application-level 56 measurement may have a wider scope. In the following, interactive 57 multimedia traffic is used as an example to illustrate the concept. 59 A constant bit-rate (CBR), or nearly CBR, streaming (hereinafter 60 called periodic) multimedia bit stream may be simulated by 61 transmitting uniformly sized packets (or mostly uniformly sized 62 packets) at regular intervals through the network to be evaluated. 63 The "mostly uniformly sized packets" may be found in applications 64 that may use smaller packets during a portion of the stream (e.g. 65 digitally coded voice during silence periods). As noted in the 66 framework document [1], a sample metric using regularly spaced 67 singleton tests has some limitations when considered from a 68 general measurement point of view: only part of the network 69 performance spectrum is sampled. However, from the point of view of 70 application-level performance, this is actually good news as 71 explained below. 73 IP delivery service measurements have been discussed within the 74 International Telecommunications Union (ITU). A framework for IP 75 service level measurements (with references to the framework for IP 76 performance [1]) that is intended to be suitable for service planning 77 has been approved as I.380 [3]. The emphasis in the ITU 78 recommendation is on passive measurements, though not explicitly 79 forbidding active measurements. The present contribution proposes a 80 method that is usable both for service planning and end-user testing 81 purposes, and is based on active measurements. 83 3.1 Terminology 85 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 86 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 87 document are to be interpreted as described in RFC 2119 [4]. 88 Although RFC 2119 was written with protocols in mind, the key words 89 are used in this document for similar reasons. They are used to 90 ensure the results of measurements from two different implementations 91 are comparable, and to note instances when an implementation could 92 perturb the network. 94 3.2 Considerations related to delay 96 For interactive multimedia sessions, end-to-end delay is an 97 important factor. Too large a delay reduces the quality of the 98 multimedia session as perceived by the participants. One approach for 99 managing end-to-end delays on an Internet path involving 100 heterogeneous link layer technologies is to use per-domain delay 101 quotas (e.g. 50 ms for a particular IP domain). The 50 ms would 102 then be included into a calculation of an end-to-end delay bound. A 103 practical implementation of such as scheme ought to address issues 104 like possibility of asymmetric delays in a route in different 105 directions, and sensitivity of an application to delay variations in 106 a given domain. There are several alternatives as to which kind of 107 derivative delay metric one ought to use in managing end-to-end QoS. 108 This question, although very interesting, is not within the scope of 109 this draft and is not discussed further here. 111 In the following, a methodology and metric are presented for 112 measuring media stream transport QoS in an IP domain. The 113 measurement results may be used in derivative metrics such as 114 average and maximum delays. A metric is presented that is a standard 115 way for performing a measurement irrespective of the possible QoS 116 mechanism utilized in the core network. As an example, for a QoS 117 mechanism without hard guarantees, measurements may be used to 118 ascertain that the "best" class gets the service that has been 119 promised for the traffic class in question. Moreover, an operator 120 could study the quality of a cheap, low-guarantee service 121 implemented using possible slack bandwidth in other classes. Such 122 measurements could be made either in studying the feasibility of a 123 new service, or on a regular basis. 125 The present draft seeks to formalize the measurements in such a way 126 that interoperable results are achieved. 128 3.3 Protocol level issues 130 Fig.1 illustrates measurements on multiple protocol levels that 131 are relevant to this draft. The major focus of the present draft 132 is on transport quality evaluation from application point of 133 view. However, to properly account for quality effects of, e.g., 134 operating system and codec on packet voice, it is beneficial to be 135 able to measure quality at IP level [5]. Link layer monitoring 136 provides a way of accounting for link layer characteristics such 137 as bit error rates. 139 --------------- 140 | application | 141 --------------- 142 | transport | <-- 143 --------------- 144 | network | <-- 145 --------------- 146 | link | <-- 147 --------------- 148 | physical | 149 --------------- 151 Fig. 1: Different possibilities for performing measurements: a 152 protocol view. Above, "application" refers to all layers above 153 L4 and is not used in the OSI sense. 155 In general, the results of measurements may be influenced by 156 individual application requirements/responses related to the 157 following issues: 159 + Lost packets: Applications may have varying tolerance to lost 160 packets. Another consideration is the distribution of lost 161 packets (i.e. random or bursty). 162 + Long delays: Many applications will consider packets delayed 163 longer than a certain value to be equivalent to lost packets 164 (i.e. real time applications). 165 + Duplicate packets: Some applications may be perturbed if 166 duplicate packets are received. 167 + Out of sequence: Some applications may be perturbed if 168 packets are received out of sequence. This may be in addition 169 to the possibility of exceeding the "long" delay threshold as a 170 result of being out of sequence. 171 + Corrupt packet header: Most applications will probably treat a 172 packet with a corrupt header as equivalent to a lost packet. 173 + Corrupt packet payload: Some applications (e.g. digital voice 174 codecs) may accept corrupt packet payload. In some cases, the 175 packet payload may contain application specific forward error 176 correction (FEC) that can compensate for some level of 177 corruption. 178 + Spurious packet: Dst may receive spurious packets (i.e. packets 179 that are not part of the metric). Many applications may be 180 perturbed by spurious packets. 182 Depending, e.g., on the observed protocol level, some issues listed 183 above may be indistinguishable from others by the application, it 184 may be important to preserve the distinction for the operators of 185 Src, Dst, and/or the intermediate network(s). 187 Because of the possible errors listed above, in most cases it is 188 recommended to use a packet identifier for each packet generated at 189 Src. Identifiers for the metric sample may be those used by the 190 underlying transport layer (e.g. RTP sequence number) or the same 191 identifiers used by an application if the application to be modeled 192 by the metric uses an identifier. The possibility of identifier 193 roll-over (reuse if intentional) during a metric collected over 194 a "long" (application dependent) time should be observed. 196 If the application does not use an identifier, it may still be 197 useful to add identifiers to the packets in the metric sample to 198 help identify possible anomalies such as out of sequence packets. 199 This would be most useful in the case where the application 200 expects to receive packets in sequence, but has no capability to 201 identify the sequence of packets received at Dst. 203 3.4 Application-level measurement 205 In what follows, a metric is proposed for application-level network 206 performance measurement. In effect, the metric is an emulation of 207 periodic multimedia stream performance. The justification for using 208 realistic application metrics in the measurement: 210 + The results of the measurement are automatically relevant to the 211 performance as perceived by the application in question. 212 + All the packets in the measurement contribute to accuracy of the 213 estimation of performance variation at timescale that is 214 important to the multimedia application (packetization 215 interval). 216 + Effects of elastic traffic (TCP) on measurement packets are 217 different for a sustained stream than for single packets during 218 overloading situations as discussed in [3]. 220 3.5 Measurement types 222 Delay measurements can be one-way [2,3], paired one-way, or 223 round-trip [6]. Accordingly, the measurements may be performed 224 either with synchronized or unsynchronized Src/Dst host clocks. 225 Different possibilities are listed below. 227 The reference measurement setup for all measurement types is 228 shown in Fig. 2. 230 ----------------< IP >-------------------- 231 | | | | 232 ------- ------- -------- -------- 233 | Src | | MP | | MP | | Dst | 234 ------- |(Src)| |(Dst) | -------- 235 ------- -------- 237 Fig. 2: Example setup for the metric usage. 239 An example of the use of the metric is a setup with a source host 240 (Src), a destination host (Dst), and corresponding measurement 241 points (MP(Src) and MP(Dst)) as shown in Figure 2. Separate equipment 242 for measurement points may be used if having Src and/or Dst conduct 243 the measurement may significantly affect the delay performance to be 244 measured. MP(Src)should be placed/measured close to the egress point 245 of packets from Src. MP(Dst) should be placed/measure close to 246 the ingress point of packets for Dst. "Close" is defined as a 247 distance sufficiently small so that application-level performance 248 characteristics measured (such as delay) can be expected to follow 249 the corresponding performance characteristic between Src and Dst to 250 an adequate accuracy. 252 The test setup just described fulfills two important criteria: 253 1) Test is made with realistic stream metrics, emulating - for example - 254 a full-duplex Voice over IP (VoIP) call. 255 2) Either one-way or round-trip characteristics may be obtained. 257 It is also possible to have intermediate measurement points between 258 MP(Src) and MP(Dst), but that is beyond the scope of this document. 260 3.5.1 One way measurement 262 In the interests of specifying metrics that are as generally usable 263 as possible, application-level measurements based on one-way delays 264 are used in the example metrics. The implication of application-level 265 measurement for bi-directional applications such as interactive 266 multimedia conferencing is discussed below. 268 Performing a single one-way measurement only yields information on 269 network behavior in one direction. Moreover, the stream at the 270 network transport level does not emulate accurately a full-duplex 271 multimedia connection. 273 3.5.2 Paired one way measurement 275 Paired one way delay refers to two multimedia streams: Src to Dst 276 and Dst to Src for the same Src and Dst. By way of example, for 277 some applications, the delay performance of each one way path is 278 more important than the round trip delay. This is the case for 279 delay-limited signals such as VoIP. Possible reasons for the 280 difference between one-way delays is different routing of streams 281 from Src to Dst vs. Dst to Src. 283 For example, a paired one way measurement may show that Src to Dst 284 has an average delay of 30ms while Dst to Src has an average delay 285 of 120ms. To a round trip delay measurement, this example would 286 look like an average of 150ms delay. Without the knowledge of the 287 asymmetry, we might miss a problem that the application at either 288 end may have with delays averaging more than 100ms. 290 Moreover, paired one way delay measurement emulates a full-duplex 291 VoIP call more accurately than a single one-way measurement only. 293 3.5.3 Round trip measurement 295 From the point of view of periodic multimedia streams, 296 round-trip measurements have two advantages: they avoid the need of 297 host clock synchronization and they allow for a simulation of 298 full-duplex connections. The former aspect means that a measurement 299 is easily performed, since no special equipment or NTP setup is 300 needed. The latter property means that measurement streams are 301 transmitted in both directions. Thus, the measurement provides 302 information on quality of service as experienced by appropriate 303 application. 305 The downsides of round-trip measurement are the need for more 306 bandwidth than an one-way test and more complex accounting of 307 packet loss. Moreover, the stream that is returning towards the 308 original sender may be more bursty than the one on the first "leg" of 309 the round-trip journey. The last issue, however, means in practice 310 that returning stream experiences worse QoS than the other one, and 311 the performance estimates thus obtained are pessimistic ones. The 312 possibility of asymmetric routing and queuing must be taken into 313 account during analysis of the results. 315 Please note that with suitable arrangements, round-trip measurements 316 may be performed using paired one way measurements. 318 4 Sample metric for multimedia stream simulation 320 The sample metric presented here is similar to the sample metric 321 Type-P-One-way-Delay-Poisson-Stream presented in [2]. "Singletons", as 322 defined in [1] and [2] are not directly used in this document because 323 certain key results (such as duplicate or out of sequence packets) 324 cannot be identified in the context of a singleton, but only as part 325 of a sample. 327 4.1 Metric name 329 Type-P-One-way-Delay-Periodic-Stream 331 4.2 Metric parameters 333 4.2.1 Global metric parameters 335 These parameters are applicable to the metrics collected in the 336 following sections (4.2.2, 4.2.3, and 4.2.4). 338 + Src, the IP address of a host 339 + Dst, the IP address of a host 340 + T0, a time, for starting to generate packets and taking 341 measurements for a sample 342 + Tf, a time, greater than T0, for stopping generation of packets 343 for a sample 344 + periodic packet interval incT, a time duration 345 + packet size p(j), the number of bytes in each packet of Type-P of 346 size j 347 + dTloss, a time interval, used for determining if a packet should 348 be considered lost 349 + Tcons, a time interval [optional] 351 While a number of applications will use one packet size (j = 1), 352 other applications may use packets of different sizes (j > 1). 353 Especially in cases of congestion, it may be useful to have 354 packets smaller than the maximum or predominant size of packets 355 in the periodic stream. 357 4.2.2 Metrics collected at MP(Src) 359 + Tstamp(Src)[i], for each packet [i], the time of the packet as 360 measured at MP(Src) 361 + PktID [i], for each packet [i], an identification number for the 362 the packet sent from Src to Dst 363 + PktSiTy [i], for each packet [i], the packet size and/or type. 364 Some applications may use packets of different size, either 365 because of application requirements or in response to IP 366 performance experienced. 368 4.2.3 Metrics collected at MP (Dst) 370 + dTstop, a time interval, used to add to time Tf to determine when to 371 stop collecting metrics for a sample 372 + Tstamp(Dst)[i], for each packet [i], the time of the packet as 373 measured at MP(Dst) 374 + PktID [i], for each packet [i], an identification number for the 375 the packet received at Dst from Src. This identification number 376 may be corrupted. 377 + PktSiTy [i], for each packet [i], the packet size and/or type. 378 Some applications may use packets of different size, either 379 because of application requirements or in response to IP 380 performance experienced. 381 + PktStatus [i], for each packet [i], the status of the packet 382 received. Possible status includes: OK, packet header corrupt, 383 packet payload corrupt, spurious, duplicate 385 4.2.4 Metrics resulting when metrics collected at MP(Src) and MP(Dst) 386 are merged 388 These parameters are only available as a complete set when the 389 parameters from the preceding sections (4.2.1, 4.2.2, and 4.2.3 are 390 combined. 392 + Tstamp(Src)[i], for each packet [i], the time of the packet as 393 measured at MP(Src). This entry may be blank or noted as N/A 394 for spurious packets received at MP(Dst) 395 + Tstamp(Dst)[i], for each packet [i], the time of the packet as 396 measured at MP(Dst). This entry may be blank or noted as N/A 397 for packets not received at MP(Dst), received with corrupt 398 packet headers, or for duplicate packets received at MP(Dst). 399 + PktID [i], for each packet [i], an identification number for the 400 the packet received. This identification number may be corrupted 401 for certain packets received at MP (Dst). 402 + PktSiTy [i], for each packet [i], the packet size and/or type. 403 + PktStatus [i], for each packet [i], the status of the packet 404 received. Possible status includes: OK, packet header corrupt, 405 packet payload corrupt, spurious, duplicate, out of sequence. 407 + Delay [i], for each packet [i], the time interval Tstamp(Dst)[i] - 408 Tstamp(Src)[i]. For the following conditions, it will not be 409 possible to be able to compute delay: 410 Spurious: There will be no Tstamp(Src)[i] time 411 Not received: There will be no Tstamp (Dst) [i] 412 Corrupt packet header: There will be no Tstamp (Dst) [i] 413 Duplicate: Only the first non-corrupt copy of the packet 414 received at Dst should have Delay [i] computed. 415 + SDV[i] [optional] , for each packet [i] except the first one: 416 momentary delay variation between successive packets, i.e., the 417 time interval Delay[i] - Delay [i-1]. SDV[i] may be negative, 418 zero, or positive. Delay for both packets i and i+1 must be 419 calculable according to the definition above or SDV[i] is 420 undefined. 422 4.3 High level description of the procedure to collect a sample 424 Beginning on or after time T0, Type-P packets are generated 425 by Src and sent to Dst until time Tf is reached with a nominal 426 interval between the first bit of successive packets of incT as 427 measured at MP(Src). incT may be nominal due to a number of reasons: 428 variation in packet generation at Src, clock issues (see section 4.6), 429 etc. 431 MP(Src) records the following information only for packets with 432 timestamps between and including T0 and Tf: timestamp, 433 packet identifier, and packet size/type of each packet sent from Src 434 to Dst that is part of the sample. 436 MP (Dst) records the following information only for packets with 437 time stamps between T0 and (Tf+ dTstop): timestamp, packet identifier, 438 packet size/type, and received status of each packet received from 439 Src at Dst that is part of the sample. Optionally, at a time Tf + 440 Tcons, the data from MP(Src) and MP(Dst) are consolidated to derive 441 the results of the sample metric. 443 To prevent stopping data collection too soon, dTcons should be greater 444 than or equal to dTstop. Conversely, to keep data collection 445 reasonably efficient, dTstop should be some reasonable time interval 446 (seconds/minutes/hours), even if dTloss is infinite or extremely long. 448 4.4 Discussion 450 The sample metric thus defined is intended to probe the delays and 451 the delay variation as experienced by multimedia streams of 452 an application. Subsequently, the delay is assumed to be measured at 453 transport layer level. Since a range of packet sizes and nominal 454 interval between packets is used, the method probes only a specific 455 time scale of network QoS variations. 457 There are a number of factors that should be taken into account when 458 collecting a sample metric of Type-P-One-way-Delay-Periodic-Stream. 460 + T0 and (Tf + dTloss) should specify a long enough time interval to 461 represent a reasonable use of the application under test (e.g. do 462 not provide only a 100 ms time interval for a phone call) 464 + T0 and (Tf + dTloss) should specify a time interval that is not 465 excessively long compared to the usage of the application under test 466 (e.g. do not provide a one week continuous phone call) 468 + The nominal interval between packets (incT) and the packet size(s) 469 (p(j)) should not define an equivalent bit rate that is in excess 470 of the capacity of the egress port of Src, the ingress port of Dst, 471 or the carrying capacity of the intervening network(s). There may 472 be exceptional cases to test the response of the application to 473 overload conditions in the transport networks, but these cases 474 should be strictly controlled. 476 + Real delay values will be positive. Therefore, it does not make 477 sense to report a negative value as a real delay. However, an 478 individual zero or negative delay value might be useful as part of 479 a stream when trying to discover a distribution of the delay values 480 of a stream. 482 + Depending on measurement topology, delay values may be as low as 483 100 usec to 10 msec, whereby it may be important for Src and Dst to 484 synchronize very closely. GPS systems afford one way to achieve 485 synchronization to within several 10s of usec. Ordinary application 486 of NTP may allow synchronization to within several msec, but this 487 depends on the stability and symmetry of delay properties among those 488 NTP agents used, and this delay is what we are trying to measure. A 489 combination of some GPS-based NTP servers and a conservatively 490 designed and deployed set of other NTP servers should yield good 491 results, but this is yet to be tested. 493 + A given methodology will have to include a way to determine 494 whether packet was lost or whether delay is merely very large (and 495 the packet is yet to arrive at Dst). The global metric parameter 496 dTloss defines a time interval such that delays larger than dTloss 497 are interpreted as losses. 498 {Comment: Note that, for many applications of these metrics, the 499 harm in treating a large delay as infinite might be zero or very 500 small. A TCP data packet, for example, that arrives only after 501 several multiples of the RTT may as well have been lost.} 503 4.5 Additional Methodology Aspects 505 As with other Type-P-* metrics, the detailed methodology will depend 506 on the Type-P (e.g., protocol number, UDP/TCP port number, size, 507 precedence). 509 4.6 Errors and uncertainties 511 The description of any specific measurement method should include an 512 accounting and analysis of various sources of error or uncertainty. 513 The Framework document [1] provides general guidance on this point, 514 but we note here the following specifics related to delay metrics: 516 + Errors or uncertainties due to uncertainties in the clocks of the 517 MP(Src) and MP(Dst) measurement points. 519 + Errors or uncertainties due to the difference between 'wire time' 520 and 'host time'. 522 4.6.1. Errors or uncertainties related to Clocks 524 The uncertainty in a measurement of one-way delay is related, in 525 part, to uncertainties in the clocks of MP(Src) and MP(Dst). In 526 the following, we refer to the clock used to measure when the packet 527 was measured at MP(Src) as the MP(Src) clock and we refer to the 528 clock used to measure when the packet was received at MP(Dst) as the 529 MP(Dst) clock. Alluding to the notions of synchronization, accuracy, 530 resolution, and skew, we note the following: 532 + Any error in the synchronization between the MP(Src) clock and 533 the MP(Dst) clock will contribute to error in the delay 534 measurement. We say that the MP(Src) clock and the MP(Dst) 535 clock have a synchronization error of Tsynch if the MP(Src) clock 536 is Tsynch ahead of the MP(Dst) clock. Thus, if we know the 537 value of Tsynch exactly, we could correct for clock 538 synchronization by adding Tsynch to the uncorrected value of 539 Tstamp(Dst)[i] - Tstamp(Src) [i]. 541 + The accuracy of a clock is important only in identifying the time 542 at which a given delay was measured. Accuracy, per se, has no 543 importance to the accuracy of the measurement of delay. When 544 computing delays, we are interested only in the differences 545 between clock values, not the values themselves. 547 + The resolution of a clock adds to uncertainty about any time 548 measured with it. Thus, if the MP(Src) clock has a resolution of 549 10 msec, then this adds 10 msec of uncertainty to any time value 550 measured with it. We will denote the resolution of the source 551 clock and the MP(Dst) clock as ResMP(Src) and ResMP(Dst), 552 respectively. 553 + The skew of a clock is not so much an additional issue as it is a 554 realization of the fact that Tsynch is itself a function of time. 555 Thus, if we attempt to measure or to bound Tsynch, this needs to 556 be done periodically. Over some periods of time, this function 557 can be approximated as a linear function plus some higher order 558 terms; in these cases, one option is to use knowledge of the 559 linear component to correct the clock. Using this correction, the 560 residual Tsynch is made smaller, but remains a source of 561 uncertainty that must be accounted for. We use the function 562 Esynch(t) to denote an upper bound on the uncertainty in 563 synchronization. Thus, |Tsynch(t)| <= Esynch(t). 565 Taking these items together, we note that naive computation 566 Tstamp(Dst)[i] - Tstamp(Src) [i] will be off by Tsynch(t) +/- 567 (ResMP(SRc) + ResMP(Dst)). Using the notion of Esynch(t), we note 568 that these clock-related problems introduce a total uncertainty of 569 Esynch(t)+ Rsource + Rdest. This estimate of total clock-related 570 uncertainty should be included in the error/uncertainty analysis of 571 any measurement implementation. 573 4.6.2. Errors or uncertainties related to Wire-time vs Host-time 575 As we have defined one-way periodic delay, we would like to measure 576 the time between when a packet is measured and time-stamped at 577 MP(Src) and when it arrives and is time-stamped at MP(Dst) and we 578 refer to these as "wire times." If the timings are themselves 579 performed by software on Src and Dst, however, then this software can 580 only directly measure the time between when Src generates the packet 581 just prior to sending the test packet and when Dst has started to 582 process the packet after having received the test packet, and we refer 583 to these two points as "host times". 585 To the extent that the difference between wire time and host time is 586 accurately known, this knowledge can be used to correct for wire time 587 measurements and the corrected value more accurately estimates the 588 desired (host time) metric. 590 To the extent, however, that the difference between wire time and 591 host time is uncertain, this uncertainty must be accounted for in an 592 analysis of a given measurement method. We denote by Hsource an 593 upper bound on the uncertainty in the difference between wire time 594 of MP(Src) and host time on the Src host, and similarly define Hdest 595 for the difference between the host time on the Dst host and the wire 596 time of MP(Dst). We then note that these problems introduce a total 597 uncertainty of Hsource+Hdest. This estimate of total wire-vs-host 598 uncertainty should be included in the error/uncertainty analysis of 599 any measurement implementation. 601 4.6.3. Calibration 603 Generally, the measured values can be decomposed as follows: 605 measured value = true value + systematic error + random error 607 If the systematic error (the constant bias in measured values) can be 608 determined, it can be compensated for in the reported results. 610 reported value = measured value - systematic error 612 therefore 614 reported value = true value + random error 616 The goal of calibration is to determine the systematic and random 617 error generated by the instruments themselves in as much detail as 618 possible. At a minimum, a bound ("e") should be found such that the 619 reported value is in the range (true value - e) to (true value + e) 620 at least 95 percent of the time. We call "e" the calibration error 621 for the measurements. It represents the degree to which the values 622 produced by the measurement instrument are repeatable; that is, how 623 closely an actual delay of 30 ms is reported as 30 ms. {Comment: 95 624 percent was chosen due to reasons discussed in [2], briefly 625 summarized as (1) some confidence level is desirable to be able to 626 remove outliers, which will be found in measuring any physical 627 property; (2) a particular confidence level should be specified so 628 that the results of independent implementations can be compared.} 630 From the discussion in the previous two sections, the error in 631 measurements could be bounded by determining all the individual 632 uncertainties, and adding them together to form 634 Esynch(t) + ResMP(Src) + ResMP(Dst) + Hsource + Hdest. 636 However, reasonable bounds on both the clock-related uncertainty 637 captured by the first three terms and the host-related uncertainty 638 captured by the last two terms should be possible by careful design 639 techniques and calibrating the instruments using a known, isolated, 640 network in a lab. 642 For example, the clock-related uncertainties are greatly reduced 643 through the use of a GPS time source. The sum of Esynch(t) + 644 ResMP(Src) + ResMP(Dst) is small, and is also bounded for the 645 duration of the measurement because of the global time source. 647 The host-related uncertainties, Hsource + Hdest, could be bounded by 648 connecting two instruments back-to-back with a high-speed serial link 649 or isolated LAN segment. In this case, repeated measurements are 650 measuring the same one-way delay. 652 If the test packets are small, such a network connection has a 653 minimal delay that may be approximated by zero. The measured delay 654 therefore contains only systematic and random error in the 655 instrumentation. The "average value" of repeated measurements is the 656 systematic error, and the variation is the random error. 658 One way to compute the systematic error, and the random error to a 659 95% confidence is to repeat the experiment many times - at least 660 hundreds of tests. The systematic error would then be the median. 661 The random error could then be found by removing the systematic error 662 from the measured values. The 95% confidence interval would be the 663 range from the 2.5th percentile to the 97.5th percentile of these 664 deviations from the true value. The calibration error "e" could then 665 be taken to be the largest absolute value of these two numbers, plus 666 the clock-related uncertainty. {Comment: as described, this bound is 667 relatively loose since the uncertainties are added, and the absolute 668 value of the largest deviation is used. As long as the resulting 669 value is not a significant fraction of the measured values, it is a 670 reasonable bound. If the resulting value is a significant fraction 671 of the measured values, then more exact methods will be needed to 672 compute the calibration error.} 674 Note that random error is a function of measurement load. For 675 example, if many paths will be measured by one instrument, this might 676 increase interrupts, process scheduling, and disk I/O (for example, 677 recording the measurements), all of which may increase the random 678 error in measured singletons. Therefore, in addition to minimal load 679 measurements to find the systematic error, calibration measurements 680 should be performed with the same measurement load that the 681 instruments will see in the field. 683 We wish to reiterate that this statistical treatment refers to the 684 calibration of the instrument; it is used to "calibrate the meter 685 stick" and say how well the meter stick reflects reality. 687 4.7 Reporting the metric 689 The calibration and context in which the metric is measured MUST be 690 carefully considered, and SHOULD always be reported along with metric 691 results. We now present five items to consider: the Type-P of test 692 packets, the threshold of delay equivalent to loss, error 693 calibration, the path traversed by the test packets, and background 694 conditions at Src, Dst, and the intervening networks during a sample. 695 This list is not exhaustive; any additional information that could be 696 useful in interpreting applications of the metrics should also be 697 reported. 699 4.7.1. Type-P 701 As noted in the Framework document [1], the value of the metric may 702 depend on the type of IP packets used to make the measurement, or 703 "type-P". The value of Type-P-One-way-Periodic-Delay could change 704 if the protocol (UDP or TCP), port number, size, or arrangement for 705 special treatment (e.g., IP precedence or RSVP) changes. The exact 706 Type-P used to make the measurements MUST be accurately reported. 708 4.7.2. Threshold for delay equivalent to loss 710 In addition, the threshold for delay equivalent to loss (or 711 methodology to determine this threshold) MUST be reported. 713 4.7.3. Calibration results 715 + If the systematic error can be determined, it SHOULD be removed 716 from the measured values. 718 + You SHOULD also report the calibration error, e, such that the 719 true value is the reported value plus or minus e, with 95% 720 confidence (see the last section.) 722 + If possible, the conditions under which a test packet with finite 723 delay is reported as lost due to resource exhaustion on the 724 measurement instrument SHOULD be reported. 726 4.7.4. Path 728 The path traversed by the packets SHOULD be reported, if possible. 729 In general it is impractical to know the precise path a given packet 730 takes through the network. The precise path may be known for 731 certain Type-P packets on short or stable paths. If Type-P includes 732 the record route (or loose-source route) option in the IP header, 733 and the path is short enough, and all routers* on the path support 734 record (or loose-source) route, then the path will be precisely 735 recorded. 737 This may be impractical because the route must be short enough, 738 many routers do not support (or are not configured for) record route, 739 and use of this feature would often artificially worsen the 740 performance observed by removing the packet from common-case 741 processing. However, partial information is still valuable context. 742 For example, if a host can choose between two links* (and hence two 743 separate routes from Src to Dst), then the initial link used is 744 valuable context. {Comment: For example, with Merit's NetNow setup, 745 a Src on one NAP can reach a Dst on another NAP by either of several 746 different backbone networks.} 748 4.7.5 Background conditions 750 In many cases, the results of a sample may be influenced by conditions 751 at Src, Dst, and/or any intervening networks. Some things that may 752 affect the results of a sample include: traffic levels and/or bursts 753 during the sample, link and/or host failures, etc. Information about 754 the background conditions may only be available by non-Internet means 755 (e.g. phone calls, television) and may only become available days after 756 samples are taken. 758 4.8 Single sample vs. a "sample of samples" 760 Because this metric represents a periodic stream as one sample, there 761 may be value in running multiple tests using this metric to collect 762 a "sample of samples". For example, it may be more appropriate to 763 test 1,000 two-minute VoIP calls rather than a single 2,000 minute 764 VoIP call. When considering collection of a sample of samples, issues 765 like the interval between samples (e.g. Poisson vs. periodic, time of 766 day/day of week), composition of samples (e.g. equal (Tf-T0 duration, 767 different packet sizes), and network considerations (e.g. run different 768 samples over different intervening link-host combinations) should be 769 taken into account. For items like the interval between samples, 770 the pattern of use of the application being measured should be 771 considered. 773 4.9 Statistics based on Type-P-One-way-Delay-Periodic-Stream 775 4.9.1 Statistics calculable from one sample 777 As a metric based on a sample representative of certain 778 applications, some general purpose statistics (e.g. median and 779 percentile) may be less applicable than ways to characterize the 780 range of delay values recorded during the sample metrics. 782 Example, a sample metric generates 100 packets as measured at MP(Src) 783 with the following measurements at MP(Dst) 785 + 80 packets received with delay [i] <= 20 ms 786 + 8 packets received with delay [i] > 20 ms 787 + 5 packets received with corrupt packet headers 788 + 4 packets from MP(Src) with no matching packet recorded 789 at MP(Dst) (effectively lost) 790 + 3 packets received with corrupt packet payload and 791 and delay [i] <= 20 ms 792 + 2 packets that duplicate one of the 80 packets received 793 correctly in the first line 795 For this example, packets are considered acceptable if they are 796 received with less than or equal to 20ms delays and without corrupt 797 packet headers or packet payload. In this case, the percentage 798 of acceptable packets is 80/100 = 80%. 800 For a different application which will accept packets with corrupt 801 packet payload and no delay bound (so long as the packet is received), 802 the percentage of acceptable packets is (80+8+3)/100 = 91%. 804 4.9.2 Statistics calculable from multiple samples 806 For computing statistics, a "sample of samples" series of 807 measurements may be performed. As discussed in section 4.8, under 808 these conditions, general purpose statistics (e.g. median, percentile, 809 etc.) may be more relevant as a more statistically significant 810 number of packets are used. 812 5. Security Considerations 814 5.1 Denial of Service Attacks 816 This metric generates a periodic stream of packets from one host (Src) 817 to another host (Dst) through intervening networks. This metric 818 could be abused for denial of service attacks directed at Dst and/or 819 the intervening network(s). 821 Administrators of Src, Dst, and the intervening network(s) should 822 establish bilateral or multi-lateral agreements regarding the timing, 823 size, and frequency of collection of sample metrics. Use of this 824 metric in excess the terms agreed between the participants MAY BE 825 cause for immediate rejection or discard of packets or other 826 escalation procedures defined between the affected parties. 828 5.2 User data confidentiality 830 This metric generates packets for a sample metric, rather than 831 taking samples based on user data. Thus, this metric does not 832 threaten user data confidentiality. 834 5.3 Interference with the metric 836 It may be possible to identify that a certain packet or stream of 837 packets are part of a sample metric. With that knowledge at Dst 838 and/or the intervening networks, it is possible to change the 839 processing of the packets (e.g. increasing or decreasing delay) 840 that may distort the measured performance. It may also be 841 possible to generate additional packets that appear to be part of 842 the sample metric. These additional packets are likely to perturb 843 the results of the sample measurement. 845 6. Acknowledgements 847 The authors wish to thank the chairs of the IPPM WG for comments 848 that have made the present draft clearer and more focused. Howard 849 Stanislevic and Al Morton ahave presented useful comments and 850 questions. The authors have also built on the substantial 851 foundations laid by the authors of the framework for IP 852 performance [1]. 854 7. References 856 [1] V.Paxson, G.Almes, J.Mahdavi, and M.Mathis: Framework for IP 857 Performance Metrics, IETF RFC 2330, May 1998. 858 [2] G.Almes, S.Kalidindi, and M.Zekauskas: A one-way delay metric 859 for IPPM, IETF RFC 2679, September 1999. 860 [3] International Telecommunications Union recommendation I.380, 861 February 1999. 862 [4] S. Bradner: Key words for use in RFCs to Indicate Requirement 863 Levels, RFC 2119, March 1997. 864 [5] ETSI TIPHON document TS-101329-5 (to be published in July). 865 [6] G.Almes, S.Kalidindi, and M.Zekauskas: A round-trip delay 866 metric for IPPM, IETF RFC 2681. 868 8. Authors' contact information 870 Vilho Raisanen 871 P.O. Box 407 872 Communication Systems Laboratory 873 Nokia Research Center 874 FIN-00045 Nokia Group 875 Finland 876 Phone +358 9 4376 1 877 Fax. +358 9 4376 6852 879 Glenn Grotefeld 880 Motorola, Inc. 881 1303 E. Algonquin Road 882 4th Floor 883 Schaumburg, IL 60196 884 USA 885 Phone +1 847 576-5992 886 Fax +1 847 538-7455 888 EXPIRES JANUARY 2001