1、 ATIS-0900003 ATIS Standard on - METRICS CHARACTERIZING PACKET-BASED NETWORK SYNCHRONIZATION ATIS is the leading technical planning and standards development organization committed to the rapid development of global, market-driven standards for the information, entertainment and communications indus
2、try. More than 200 companies actively formulate standards in ATIS 17 Committees, covering issues including: IPTV, Cloud Services, Energy Efficiency, IP-Based and Wireless Technologies, Quality of Service, Billing and Operational Support, Emergency Services, Architectural Platforms and Emerging Netwo
3、rks. In addition, numerous Incubators, Focus and Exploratory Groups address evolving industry priorities including Smart Grid, Machine-to-Machine, Networked Car, IP Downloadable Security, Policy Management and Network Optimization. ATIS is the North American Organizational Partner for the 3rd Genera
4、tion Partnership Project (3GPP), a member and major U.S. contributor to the International Telecommunication Union (ITU) Radio and Telecommunications Sectors, and a member of the Inter-American Telecommunication Commission (CITEL). ATIS is accredited by the American National Standards Institute (ANSI
5、). For more information, please visit . Notice of Disclaimer all components of the local oscillator clock noise below the corner frequency will be removed. Equivalently, all the components of the PDV post-PP below the corner frequency will be passed through to the clock output and all components of
6、the local oscillator noise above the corner frequency will be passed through to the clock output. This is a crucial observation. In the absence of packet pre-processing algorithms, the low-pass filtering action of the PLL is responsible for attenuating the entire PDV so that the clock output mask ca
7、n be complied with. In order to achieve this, it may be necessary to make the corner frequency of the PLL filter very small. The drawback is that the high-pass characteristic then allows more local oscillator noise to feed into the clock output. This implies that any attempt to reduce the bandwidth
8、of the PLL cannot be done arbitrarily; the quality of the local oscillator must be accounted for. A high-quality oscillator generates less noise that feeds into the clock output. Another way of expressing the same idea is via the notion of time-constant and stability. The reciprocal of the corner fr
9、equency is proportional to the time-constant of the PLL and represents the averaging time applied by the PLL. For proper operation, the oscillator must be stable and should not “move” or “drift” substantively in any time interval smaller than the time-constant. 4.2 Definition of the Time-Error Seque
10、nce Two families of metrics have been proposed for packet measurement data analysis. One has, as its basis, the TDEV metric and the other is linked to MTIE two metrics that are central to synchronization measurement analysis. The MTIE and TDEV metrics are also used in connection with synchronization
11、 equipment requirements and synchronization interface requirements with limits expressed in terms of MTIE and TDEV masks. However, experience has shown that there are other metrics for which normative limits are not defined yet are useful for purposes of analysis. Central to the idea of timing trans
12、fer is the notion of clock noise. The term clock noise as used here represents all impairments to the timing information, including jitter, wander, and other items such as packet delay variation that impact the regularity of timing events (such as clock edges or time-stamps). A simple depiction of w
13、hat constitutes clock noise in a conventional scenario is provided in Figure 4. A clock has a nominally periodic waveform and the deviation of actual from the nominal (ideal) is the clock noise (also called time error). The sequence x(n) is usually called the Time-Error (TE) sequence. For high rate
14、clocks it is permissible to divide the clock down to have a convenient sampling rate for the TE sequence provided the high rate clock does not have high-frequency jitter components (this is an application of the Nyquist theorem on sampling well known in discrete-time signal processing). In practice,
15、 measurements are made between the client clock output and a reference ATIS-0900003 8 clock. The reference clock may not be ideal but is known to be of much higher quality than the client clock. The time-error sequence is a fundamental quantity. All clock metrics are computed from the TE sequence x(
16、n) (also written as xn), which can be viewed as a sampled data signal with underlying sampling interval 0. NOTE The term wander is used in the telecommunications community to identify low-Fourier-frequency (i.e., 10Hz) jitter; the data-communications community does not make this distinction. Figure
17、4: Clock noise for conventional clock signals (periodic waveforms) For packet timing signals, the time-error sequence can be established in the following way. For specificity, consider the transfer of timing packets originating at the “source” (i.e., “master”) and terminating at the “slave” (i.e., “
18、client”). In the case of PTP (see 17), the rate of packets, f0, is determined via negotiation between master and slave and can be as high as f0= 128 packets/sec. Packets leaving the master occur, nominally, with a spacing of 0= 1/f0. From a signal processing perspective, the sampling rate is f0and a
19、n arbitrary mathematical-time origin for describing the times of departure from the master can be chosen. With this choice of time origin, the kthpacket departs the master at time t = k 0. In practice the kthpacket will depart at time Tk, which is only approximately equal to k 0. Note that in the ca
20、se of circuit emulation, the times of departure are considered to be exactly spaced by 0. The kthpacket then arrives at the slave at time Sk, where: kkkkkkTSeTS (Eq. 1) Where is the reference transit delay time and kis the transit delay time variation (i.e., packet delay variation or PDV). For calcu
21、lating metrics of the TDEV and MTIE family, the operation involves a differencing (as shown later) and consequently the reference transit delay time, , is moot since it is part of every term. Consequently, for purposes of calculating metrics of the TDEV and MTIE family, ekcan be used as the packet d
22、elay variation and used interchangeable with k. The same principle applies for packets that traverse the network from the slave to the master. If the (hypothetical) time-error signal x(t) is considered, then the sample of x(t) taken at the sampling instant Tkis none other than ek. That is, the seque
23、nce ek is equivalent to the time-error sequence but ATIS-0900003 9 on a non-uniform time grid. The normal time-error sequence, xk, is actually equivalent to the sequence generated by sampling the time-error signal x(t) on a uniform grid with sampling interval 0. 4.3 PDV Measurement In general, a PDV
24、 measurement involves comparing time instants on a sequence of packets, such as those of a packet timing signal, as they pass two points in the network. A configuration for performing such a measurement is shown in Figure 5 below (which is Figure 7 from G.8260). For each packet, a difference is comp
25、uted between the time instant taken at the point of origin and the time instant taken at the point of destination. Figure 5: Configuration for PDV measurement An ideal configuration for making this measurement places two references traceable to a common time standard at each of the two measurement p
26、oints. Such a configuration assesses not only the variation of packet delay, but also the packet transit time. In many circumstances, such as packet-based frequency synchronization, the focus is on variation of packet delay rather than absolute packet delay. In such a case, frequency references can
27、be employed for the references R and a common time reference is not required. The use of unstable or inaccurate references directly impact the PDV measurement and may lead to limitations regarding the length of the PDV measurement. If the references are frequency standards, packet delay variation ca
28、n be studied with the same precision as for the case where the references are time standards. If practical, a common frequency standard should be used for both R references. In other cases, separate primary reference clocks could be used. The probe function could be implemented as separate equipment
29、, or in the case where the first measurement point is at the source of the packet timing signal of interest, integrated into that equipment. In this case, the time instant could be delivered within the packet in the form of a timestamp. Similarly, in the case where the second measurement point is at
30、 the destination equipment, the probe function may be integrated into that equipment. Any inaccuracy of the timestamping function in the probes directly impacts the precision of the PDV measurement. In the case where packets are sent according to a schedule that is known in advance, such as packets
31、spaced by a uniform interval of time (e.g., CES), the relative origin timestamps are implicit, and the packet delay variation measurement can be performed with timestamping at the destination node. ATIS-0900003 10 4.4 Packet Selection Processes As will be shown later, packet selection can be incorpo
32、rated directly into the calculation or implemented as a pre-processing step prior to application of the formula for the metric. The packet selection techniques integrated into the calculation are useful in metrics that are intended to determine the fundamental limits set by the packet delay variatio
33、n. For example, in minTDEV, the least delay is selected from an increasing window as the value of is increased. On the other hand, pre-processing selects packets from fixed window lengths. Thus, the selection process resembles that of a practical packet clock in steady state operation. The Maximum A
34、verage Time Interval Error (MATIE) and Maximum Average Frequency Error (MAFE) calculations, for example, have been studied using pre-processing techniques. Selection processes that may be applied to either integrated or pre-processed metrics include (1) selection of packets with minimum delay, (2) s
35、election of packets between the minimum delay and some percentile or time range delay limit, and (3) selection of packets around a non-minimum delay within a percentile or time range delay limit. The effect of a packet selection process from the viewpoint of time error can be viewed in the following
36、 way. Suppose that the original packet interval is Pand the pre-processing involves non-overlapping windows of K samples. This is referred to as fixed-window pre-processing. If the original sampling interval is Pand the pre-processing window width is K samples, the new sequence has a sample interval
37、 of 0= K P. The packet delay selection could involve the selection of a single minimum, the averaging together of a small number of selected minimum values at the floor, or some other rule. The effect of this pre-processing is to synthesize a time-error sequence with sampling interval 0= K Pand samp
38、le values that are representative of the time error over the selection window. Note that the creation of a new time-error sequence with underlying sampling interval 0permits the application of all conventional metrics such as MTIE and TDEV on this time-error sequence. It should be noted that the win
39、dow duration associated with the packet selection processing (fixed-window processing) should not be confused with the observation interval, , that forms the independent variable for the metrics such as TDEV() and MTIE() among others. In the case of uniform time-error sampling, each fixed window con
40、tains the same number of samples, K. The case of non-uniform sampling is considered later. ATIS-0900003 11 Figure 6: Non-overlapping windows of width 0 (= K P) (fixed-window processing) 4.4.1 Pre-Processed Packet Selection With pre-processed packet selection, quantifying packet timing signals is car
41、ried in two steps: (1) Applying a specific packet selection procedure to select a specific subset of PDV noise samples, having similar delay properties, among the entire population of PDV noise samples. (2) Applying the required stability quantification algorithm (metric) over the selected group of
42、samples to estimate the achievable output clock quality. Figure 7: Preprocessed packet selection With pre-processed packet selection, quantifying packet timing signals is carried in two steps. In essence, an input packet TE sequence x(t) is subject to packet selection, which produces a new packet TE
43、 sequence x(t). In the case of pre-processed packet selection, the preliminary packet selection process is independent of the applied stability quantification. Thus, different combinations of the two might yield interesting properties and need to be looked into. Both need to be fully defined as each
44、 has significant influence on the resulting performance measurement. ATIS-0900003 12 4.4.2 Integrated Packet Selection With integrated packet selection, the packet selection is integrated into the metric calculation. Generally this involves replacing a full-population averaging calculation with a se
45、lection process that may or may not include averaging by itself. The minTDEV and minMAFE calculations defined below are examples of this. Figure 8: Integrated packet selection 4.4.3 Packet Selection Methods Some examples of packet selection methods are described in the sections that follow. The firs
46、t two minimum packet selection and percentile packet selection focus on packet data at the floor. The second two band packet selection and cluster range packet selection can be applied either at the floor or at some other region. The packet selection methods can be either applied as a pre-processing
47、 step or integrated into the calculation. The notion of “floor” is equivalent to the notion of minimum possible transit delay. That is, the transit delay experienced by a packet that has not experienced any waiting time of any sort in any of the network elements it has traversed between master and s
48、lave corresponds to the “floor delay”. Depending on loading and other considerations, it is possible that in any given finite window of observation a “lucky” packet with delay equal to this theoretical minimum may not be observed. The cluster method applies a time-aperture criterion rather than a pe
49、rcentage criterion to achieve the selection. That is, whereas band and percentile selection will choose a particular chosen fraction of the packets within the selection window, the time-aperture method chooses all the packets within a particular chosen transit delay range that is referred to here as cluster range. 4.4.3.1 Minimum Packet Selection Method The minimum packet selection method involves selecting a minimum within a section of data. This can be represented as follows: 1minmin nijiforxixj(Eq. 2)
