T Rec G.8260 201508 I!!pdf e
T Rec G.8260 201508 I!!pdf e
T Rec G.8260 201508 I!!pdf e
ITU-T G.8260
TELECOMMUNICATION (08/2015)
STANDARDIZATION SECTOR
OF ITU
Summary
Recommendation ITU-T G.8260 provides the definitions, terminology and abbreviations used in
ITU-T Recommendations on timing and synchronization in packet networks.
History
Edition Recommendation Approval Study Group Unique ID*
1.0 ITU-T G.8260 2010-08-12 15 11.1002/1000/10907
2.0 ITU-T G.8260 2012-02-13 15 11.1002/1000/11521
2.1 ITU-T G.8260 (2012) Amd. 1 2013-08-29 15 11.1002/1000/12016
2.2 ITU-T G.8260 (2012) Amd. 2 2014-05-14 15 11.1002/1000/12189
3.0 ITU-T G.8260 2015-08-13 15 11.1002/1000/12545
3.1 ITU-T G.8260 (2015) Amd. 1 2016-04-13 15 11.1002/1000/12808
Keywords
Frequency, phase and time, packet delay variation, synchronization definitions.
____________________
* To access the Recommendation, type the URL http://handle.itu.int/ in the address field of your web
browser, followed by the Recommendation's unique ID. For example, http://handle.itu.int/11.1002/1000/11
830-en.
NOTE
In this Recommendation, the expression "Administration" is used for conciseness to indicate both a
telecommunication administration and a recognized operating agency.
Compliance with this Recommendation is voluntary. However, the Recommendation may contain certain
mandatory provisions (to ensure, e.g., interoperability or applicability) and compliance with the
Recommendation is achieved when all of these mandatory provisions are met. The words "shall" or some other
obligatory language such as "must" and the negative equivalents are used to express requirements. The use of
such words does not suggest that compliance with the Recommendation is required of any party.
ITU 2016
All rights reserved. No part of this publication may be reproduced, by any means whatsoever, without the prior
written permission of ITU.
1 Scope
This Recommendation provides the definitions, terminology and abbreviations used in
Recommendations on frequency, phase and time synchronization in packet networks. It includes
mathematical definitions for various synchronization stability and quality metrics for packet
networks, and also provides background information on the nature of packet timing systems and the
impairments created by packet networks.
Ethernet physical layer methods for synchronization are based on traditional time division
multiplexing (TDM) physical layer synchronization and therefore most of the definitions related to
these methods are covered by [ITU-T G.810]. Additional definitions are included in this
Recommendation.
2 References
The following ITU-T Recommendations and other references contain provisions which, through
reference in this text, constitute provisions of this Recommendation. At the time of publication, the
editions indicated were valid. All Recommendations and other references are subject to revision;
users of this Recommendation are therefore encouraged to investigate the possibility of applying the
most recent edition of the Recommendations and other references listed below. A list of the currently
valid ITU-T Recommendations is regularly published. The reference to a document within this
Recommendation does not give it, as a stand-alone document, the status of a Recommendation.
[ITU-T G.810] Recommendation ITU-T G.810 (1996), Definitions and terminology for
synchronization networks.
[ITU-T G.811] Recommendation ITU-T G.811 (1997), Timing characteristics of primary
reference clocks.
[ITU-T G.8261] Recommendation ITU-T G.8261/Y.1361 (2013), Timing and
synchronization aspects in packet networks.
[ITU-T G.8261.1] Recommendation ITU-T G.8261.1/Y.1361.1 (2012), Packet delay variation
network limits applicable to packet-based methods (Frequency
synchronization).
[ITU-T G.8263] Recommendation ITU-T G.8263/Y.1363 (2012), Timing characteristics of
packet-based equipment clocks.
[ITU-T Y.1413] Recommendation ITU-T Y.1413 (2004), TDM-MPLS network
interworking User plane interworking.
[IEEE 1588] IEEE Standard 1588-2008, IEEE standard for a precision clock
synchronization protocol for networked measurement and control systems.
3 Definitions
3.1.6 packet-based method: Timing distribution method (for frequency or time or phase) where
the timing information is associated with packets.
The frequency can be recovered using two-way or one-way protocols.
Time and phase information is recovered with a two-way protocol in order to compensate for
the transfer delay from packet master clock to packet slave clock.
3.1.17 primary reference time clock (PRTC): A reference time generator that provides a reference
timing signal traceable to an internationally recognized time standard [e.g., Co-ordinated Universal
Time (UTC)].
3.1.20 time error: The difference between the time of a clock and the time indicated by the time
standard. A model for expressing the time error of a clock is described in clause I.3 of [ITU-T G.810].
constant time error: With reference to the time error model provided in clause I.3 of
[ITU-T G.810], the constant time error (cTE) of a synchronized clock is the term x0.
constant time error estimate: Given a time error sequence {x(n); n = 0, 1 (N 1)}, an
estimate of the constant time error is the average of the first M samples of the time error
sequence. M is obtained from the observation interval providing the minimum value for
TDEV as computed for the given time error sequence. If a frequency offset is present, then a
linear regression method in accordance with Appendix II of [ITU-T G.823] can be applied.
Considerations for measurement data containing transients is for further study.
NOTE In some cases ,due to the frequency components of the noise of the signal being measured,
it might be difficult to identify a stable, consistent observation interval. These cases must be addressed
case by case.
dynamic time error: With reference to the time error model provided in clause I.3 of
[ITU-T G.810], the dynamic time error (dTE) of a synchronized clock is the random noise
component, i.e.,
5 Conventions
No conventions are used in this Recommendation.
A packet timing signal is similar in concept. The frequency is encoded as a series of time-critical
packets in a network, known as event packets. While the transmission medium is different (packets
on a network as opposed to signals on a wire), the packets still contain significant instants (normally
the front edge of the packet), with a defined ideal position in time. The variation of the significant
instants around their ideal position is termed packet delay variation (PDV). This is shown in Figure 6.
Some of the causes and characteristics of PDV and other impairments that may be introduced by the
packet network are discussed in clause 10 of [ITU-T G.8261].
Some packet timing signals may be periodic (e.g., circuit emulation packets containing constant bit
rate data), and for these the ideal position in time is implicitly given by the packet rate. Other packet
timing signals are not periodic (e.g., PTP or NTP), and for these the ideal position in time is given by
a timestamp embedded in the packet data. It is important to note that both periodic and non-periodic
packet timing signals are still time domain signals. It is the position in time of the packets that is
significant, not the contents of the packets.
It can be shown that dafwd and darev can be used interchangeably with dfwd and drev in the formulas that
derive time error from delay measurements.
Requirements on the accuracy of the reference (R) are driven by the characteristics of the packet
timing signal, and in many cases might exceed those for studying network PDV. If the packet timing
signal is derived directly from a primary reference, reference (R) would need to be a primary
reference, ideally one with greater stability. Further, in cases where the device under study takes an
external reference directly or is traceable to an external reference, the optimal configuration is for
both the device and the packet timestamper (PT) in Figure 10 to share the same reference (R).
NOTE This appendix contains information related to ongoing studies on the definition of suitable PDV
metrics. The text below is for information only, and may be revised in a future version of the Recommendation.
It must not be used as normative text, nor as an implied specification of a packet slave clock.
I.1 Introduction
With the telecommunications industry evolving and rapidly adopting packet technology, much
emphasis has been placed on addressing packet synchronization and timing, including the use of
measurement data to assist in specifying the performance of packet-based clocks.
Physical layer timing signal stability quantities, including metrics such as maximum time interval
error (MTIE) and time deviation (TDEV), have been used extensively and are central to
synchronization measurement analysis. For a packet clock, the level of stability at the clock's packet
network input has a direct bearing on the stability of the clock output.
In terms of the packet metrics, the goal of a first category of PDV metrics, introduced in clause I.4,
is to formulate packet-based stability quantities (metrics) that will provide a means of estimating the
physical-based stability quantities for the packet clock output. This is illustrated in Figure I.1.
Physical
Packet layer timing
Packet interface interface
PEC slave
network
A second category of PDV metrics is also introduced in clause I.5. The goal of this second category
is not directly to provide an estimation of the physical-based stability quantities for the packet clock
output, but simply to study the population of timing packets within a certain delay window range.
PDV measurement guidelines are provided in clause 6.5.
For packet measurement data analysis, packet selection is added as an important component to the
analysis. Indeed, in order to reduce the input PDV noise, the packet slave clock implementations
generally use only a subset of the received timing packets.
Therefore, a first simple approach to analyse the PDV as received by a packet slave clock can be to
display the measured PDV in the form of a histogram. It generally provides useful information about
the population of packets in different delay regions, and is in some cases sufficient to analyse the
network conditions. Figure I.2 shows an example plot of the measured PDV and the corresponding
histogram.
1k
90.0
s/div
100
10
180
s 1
0.000 1.00 hours/div 10.36 180.00 bins 2048 Measurements 596919 1.1700
hours hours s ms
G.8260(10)_FI.2
In a second approach, mathematical tools (called "metrics" in this appendix) can be applied on a given
PDV measurement to analyse it more in detail. Those metrics generally use only a subset of the
packets. The packet selection can be either integrated into the calculation or performed as a pre-
processing step. For example, the packet selection can focus on the minimum packet delay floor or
more generally on some other region of packet delays.
With regard to the packet selection just discussed, it is important to point out the link between the
methods of packet measurement data analysis described here and packet clock algorithms as they
exist in actual equipment. For both, packet selection is important for optimization given the realities
of packet delay variation.
However, it is important to mention that due to the proprietary nature of most of the packet slave
clock implementations today, especially regarding the packet selection criteria, the packet selection
used by a given PDV metric may not correspond to the criteria used in the packet slave clock of
interest. Therefore, there can be some discrepancies between the information provided by a given
PDV metric and the real performances achieved by a packet slave clock.
Methods for alignment of the results provided by the PDV metrics with the performance of the packet
slave clock are still under study. Alignment may involve the specification of some minimum common
behavior in the packet selection criteria in the packet slave clock implementations.
Moreover, it is important to mention that PDV metrics compute an estimate of achievable
performance through the use of PDV sample information only, and do not consider the effects of
internal oscillator noise in a packet slave clock. Non-negligible differences between the estimate and
the actual performance of a packet slave clock may sometimes be observed because of this effect. In
order to take oscillator noise into account, the noise generation components of a packet slave clock
are considered in [ITU-T G.8263].
While metrics can provide the basis for setting equipment requirements and network limits, their
value as general analysis tools leading to insight into particular sets of measurement data should not
be overlooked. For time division multiplexing (TDM) synchronization measurements, normative
limits have been applied to the MTIE and TDEV calculations, but other metrics such as Allan
deviation (ADEV) and modified Allan deviation (MDEV), while not associated with normative
telecom limits, have great utility as analysis tools.
is the forward transit delay and T2(i) is expressed relative to the master clock. Assuming the time
instants are determined by synchronized clocks, this forward transit delay is composed of a fixed,
albeit unknown, delay and a random delay component that is the result of queuing delays in the
intermediate network elements. If the application is frequency synchronization, the fixed (unknown)
delay is not relevant and can be ignored, leaving the random delay component as the principal cause
of time error.
The same principle applies for packets that traverse the network from the slave to the master.
Denoting the packet rate in the reverse direction by fR = 1/R, the jth packet will depart at time T3 ( j )
which is approximately equal to jR.
(I-2)
d rev ( j ) T4 ( j ) T3 ( j )
is the reverse packet delay that is composed of a fixed, unknown, delay and a random delay
component. T3(i) and T4(i) are expressed relative to the same clock.
If the forward delay were known a priori, then the master and slave clock could be time synchronized
using Equation I-1. However, if the delay is not known a priori, and the slave takes the master time
to be T1(i) at the instant it receives the forward packet, the (hypothetical) time-error at this instant will
be T1(i) T2(i) = dfwd(i) (i.e., the slave's estimate of the master time minus the actual master time),
because the actual time relative to the master clock that the forward packet is received at the slave is
T2(i). Then, if the (hypothetical) forward packet time-error signal xF(t) is considered, the sample of
the forward packet time-error signal taken at the sampling instant T1 (i ) is:
(I-3)
xF t dfwd (i)
Similarly, if the reverse packet time-error signal xR(t) is considered, then the sample of the reverse
packet time-error signal taken at the sampling instant T3 ( j ) is:
xR t drev ( j )
(I-4)
The sign is reversed compared to the forward time error signal, because now the slave takes the master
time to be T4(i) when it is actually T3(j); the time error is T4(j) T3(j) = drev(j).
This means that increasing forward delay corresponds to increasingly negative time error in the slave,
whereas increasing reverse delay corresponds to increasingly positive time error. That is, the
sequences { dfwd (i) } and { drev ( j ) } are equivalent to packet time error sequences, but on a
non-uniform time grid. Note that delay measures the difference between two time instants, while time
error measures the difference between two clocks at the same instant if the packet were used to set or
estimate the time of the slave clock without any filtering. The normal packet time error sequences,
In essence, an input packet time error sequence xt xi0 is subject to packet selection, which
produces a new packet time error sequence. The input packet time error sequence is divided into time
windows of equal length m0 where m is the number of samples in the selection window. A fraction
of packets is selected from each window in a similar manner and the information is combined so that
each window produces a single value to the new packet time error sequence x' t x' i s . The sample
interval of the selected-packet time error sequence is s m 0 . Unless otherwise specified, the time
errors of the selected packets in each selection window are averaged to produce a single delay value.
When a metric is computed using pre-processed packet selection, the name of the metric is prepended
by the term pktselected, e.g., pktselectedTDEV (see clause I.4.2.1).
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 that require investigation. Both need to be fully defined as each has significant influence
on the resulting performance measurement.
I.3.1.2 Pre-processed packet filtering
As described above, an input packet time error sequence x(t) that is subject to packet selection,
produces a new selected-packet time error sequence x'(t). Additionally that new packet time error
sequence x'(t) may be subsequently filtered to create a filtered-packet time error sequence y(t). This
flow is shown below in Figure I.5.
The selected-packet time error sequence x'(t) may be filtered by applying an averaging function in
line with the clock bandwidth, with averaging time related to the window length of the packet
The filtered-packet time error sequence y(t) may be used to compute TIE and applied to traditional
synchronization metrics defined in [ITU-T G.810] such as MTIE and TDEV. When TIE or a metric
are computed using y(t), the term pktfiltered is prepended to TIE or the name of the metric,
respectively. As illustrated in Figure I.5, the pktfiltered operation includes both packet selection and
bandwidth filtering.
I.3.1.3 Integrated packet selection
With integrated packet selection, the packet selection and sometimes also the filtering steps are
integrated into the metric calculation, as shown in Figure I.6. Generally this involves replacing a full
population averaging calculation with a selection process that may or may not itself include
averaging.
____________________
1 The time constant of a PLL, also known as its characteristic response time, provides an indication of the
duration of the effects on the output of the PLL due to a given input. This is why it is important that the
selection window is properly chosen in order to get a significant number of samples during this period of
time. Note that the time constant c is related to the 3 dB bandwidth of the PLL f3 dB, by the following
relationship: c = 1/(2f3 dB)
Each of the indices a and b is determined by rounding to find the closest index to the desired percentile
value. The additional constraint is that both indices have a minimum value of the first index and a
maximum value of the last index. Further, at least one point within the data set must be selected.
Thus, for example, a set of ten points with the percentile values set to 0% and 2% (0.02), both a and
b would be set to the minimum index so that at least a single point would be selected.
I.3.2.4 Cluster range packet selection method
The cluster range packet selection method uses a time- and/or phase-bounded range rather than
indices based on percentiles (probabilities) to perform the packet selection. This selection method
involves the selection of a group of one or more packets that are closely related with respect to their
transit time. The location of the cluster may be made based on various criteria, for example, packets
at the floor or from some other region observed in the window interval, or the location of the cluster
may be based on other criteria or information outside the interval. The cluster of packets could then
be processed in a variety of ways to generate a single value for that interval, such as the mean transit
time of all packets within the cluster.
Figure I.7 shows an example of a packet delay sequence, zooming in on an example of a packet cluster
for a single window interval.
where in Equation I-10, ( ) is the indicator function that expresses the selection mechanism in a
mathematical manner and is given by:
1 for x(nK i)P a(n)
n, i 2 (I-11)
0 otherwise
When combined with packet selection, which is performed on the forward and reverse sequences
independently, the packet-selected time-error sequence of the forward path, xF'(t), is combined with
the packet-selected time-error sequence of the reverse path, xR'(t), to create the packet-selected
two-way time error sequence xC'(t).
xR ' (ns ) xF ' (ns )
xC ' (ns ) (I-12b)
2
where s is the packet selection window width.
Figure I.10 shows that when the combination operation is preceded by packet selection, the
packet-selected two-way time error sequence is produced. This can itself be optionally combined with
bandwidth filtering and/or stability metrics. The sequence xC'(t) is referred to as packet-selected
two-way time error (pktSelected2wayTE).
Figure I.10 Two-way time error including packet selection and filtering
Note that the combination operation could be performed after bandwidth filtering as applied to each
packet selection output separately, with the same results.
Subsequently, xC'(t) for two-way flows (or xC(t) if there is no packet selection) may be substituted
into the various metrics in the same manner as x'(t) for one-way flows. When used in this way, the
prefix ''2way'' denotes the fact that the metric is computed on a two-way flow, e.g., ''2wayTDEV'',
''pktSelected2wayMAFE'', or ''pktFiltered2wayMTIE''.
x xi
1
Xn max in (I-13)
1 k N 2n 1 n i k
where xmin k is as defined in equation (I-8) for the forward time error sequence, or (I-8a) for the
reverse time error sequence, n0 is the observation window length, n is the number of samples in the
window, 0 is the sample interval, N is the number of samples in the data set, and k is incremented for
sliding the window.
Estimator formula
minMATIE(n0) may be estimated by:
minMATIEn 0 max xmin k n xmin k
1 k N 2 n 1 (I-17)
for n = 1, 2, ..., integer part (N/2)
The above is a point estimate, and is obtained for measurements (i.e., samples xi of the packet time
error sequence, which represent the data values) over a single measurement period (see Figure II.1 of
[ITU-T G.810]). Estimates of minMATIE (for specified N, = n0, and ), and their respective
degrees of statistical confidence, may be obtained from measured data if measurements are made for
multiple measurement periods (see clause II.5 of [ITU-T G.810]).
I.4.1.3 pktfilteredTIE
pktfilteredTIE is the TIE of the filtered-packet time error sequence, substituted into the formula
defined in [ITU-T G.810].
pktfilteredTIE(t,) = y(t + ) y(t) (I-18)
I.4.1.4 pktfilteredMTIE
pktfilteredMTIE is the MTIE of the filtered-packet time error sequence, obtained from the appropriate
formula given in [ITU-T G.810] for the definition or estimator.
Definition
pktfilteredMTIE() is defined as a specified percentile, , of the random variable:
(I-19)
Y max max y(t ) min yt ,
0t 0 T t 0 t t 0 t 0 t t 0
where T is the measurement interval and is the observation interval.
Estimator formula
pktfilteredMTIE(n0) may be estimated by:
N
for n = 1, 2, ..., integer part
3
where the angle brackets denote ensemble average.
Estimator formula
minTDEV(n) may be estimated by:
N 3n1
minTDEVn0 xmin i 2n 2 xmin i n xmin i
1 2
(I-25)
6N 3n 1 i1
N
for n = 1, 2, ..., integer part
3
x _ band () bandTDEV() 1
6 x
band_ avg i 2n 2 xband _ avg i 2
_ avg i n xband (I-27)
N 3 n1
bandTDEVn0 _ avg i 2n 2 xband
_ avg i n xband
_ avg i 2 (I-28)
1
6N 3n 1 i1
xband
N
for n = 1, 2, ..., integer part
3
Usage
The bandTDEV calculation has the flexibility, in comparison to minTDEV and percentileTDEV, of
being able to select a region of packet delay values away from the floor. Thus, if the population of
packet delay values at the floor is noisier than the population immediately above, bandTDEV indices
could be selected to focus analysis on that region.
Some of the comments on minTDEV usage apply here, but bandTDEV can apply effectively to
distributions other than one-sided distributions slanted towards the packet with the minimum delay.
It is particularly effective for packet delay distributions with a well-populated mode somewhere in
the packet delay distribution.
Pros and cons
Like minTDEV and percentileTDEV, bandTDEV gives information on network packet delay noise
processes but is not optimal for frequency offset characterization.
Like TDEV, minTDEV, and percentileTDEV, bandTDEV is sensitive to systematic effects which
could mask noise components.
The definition of the precise aspects that create the potential sensitivities listed above and the
subsequent method of handling these when applying this metric are for further study.
For clusterTDEV the average of the values in the cluster range as per Equation I-29 is substituted for
the mean value in the defining equation for TDEV. ClusterTDEV(n0) can then be defined as:
(I-30)
x _ cluster () clusterTDEV() 1 xcluster
i 2n 2 xcluster
i n xcluster
i
2
6
where the angle brackets denote ensemble average.
Estimator formula
clusterTDEV(n0) may be estimated by:
N 3n 1 , (I-31)
clusterTDEVn0 [ xcluster i 2n 2 xcluster i n xcluster i ]
1
2
6N 3n 1 j 1
N
for n = 1, 2, ..., integer part
3
Usage
The clusterTDEV calculation has the flexibility of being able to select a region of packet delay values
away from the floor. Thus, if the population of packet delay values at the floor is noisier than the
population immediately above, clusterTDEV indices could be selected to focus analysis on that
region. Generally speaking, clusterTDEV provides a quantitative measure of stability of transit delays
that are in a pre-determined band based on a phase and/or time range centred at a value that is
determined by a chosen selection rule.
Some of the comments on minTDEV usage apply to clusterTDEV as well, but clusterTDEV can
apply effectively to distributions other than one-sided distributions slanted towards the minimum
delayed packet. It is particularly effective for packet delay distributions with a well-populated mode
somewhere in the packet delay distribution.
Pros and cons
Like minTDEV, clusterTDEV gives information on network packet delay noise processes but is not
suitable for frequency offset characterization.
Like TDEV and minTDEV, clusterTDEV is sensitive to systematic effects which could mask noise
components.
Unlike TDEV and like minTDEV, clusterTDEV is sensitive to frequency offsets. Frequency offsets
may be more difficult to ascertain precisely when neither a well-populated floor nor ceiling exists.
Two additional parameters, the cluster range and the cluster rule, must be selected for clusterTDEV.
I.4.2.1.5 pktfilteredTDEV
pktfilteredTDEV is the TDEV of the filtered-packet time error sequence, obtained from the
appropriate formula given in [ITU-T G.810] for the definition or estimator.
Definition
pktfilteredTDEV(n0) is defined as:
2
1 n
pktfilteredTDEV(n0) = 2
yi 2n 2 yi n yi , (I-33)
6n i 1
MATIE( n 0 )
(I-41)
b MAFE( n 0 )
n 0
where WPM(k) is a white noise sequence with power (variance) 2. It can be shown that
whereas a closed form expression for MATIE does not exist it can be approximated by:
1 (I-43)
MATIE( n 0 ) 4 2 0
n
0
Equation I-43 assumes that the noise distribution is approximately Gaussian, based on the
central limit theorem. The maximum value is approximated as four times the standard
deviation. This is represented by the factor of 4 in Equation I-43. For a Gaussian distribution,
four standard deviations corresponds to the upper 6.33 105 quantile (2-sided).
In terms of MAFE:
3 (I-44)
MAFE( n 0 )
MATIE()
4 2 0
2
Thus the impact of white PM on values of MATIE/MAFE becomes less important for large
values of the observation interval . When plotted on a log-log scale, the graph of MATIE
appears as a straight line of slope 0.5 (1.5 for MAFE).
c. Flicker phase noise
In this case
(I-45)
x(k ) FPM (k )
Generally speaking, flicker lies between white noise and random walk. If C is a constant that
is related to the strength of the flicker random process, then we can write (note that this is not
a formal proof.):
MATIE( n 0 ) 4 C (I-46)
MAFE( n 0 ) 4 C ( 1)
The factor of 4 represents four standard deviations as described in the white phase noise
description above.
d. Random walk phase noise
In this case
(I-47)
x(k ) WFM(k )
and it can be shown that for large n, we can write
2 (I-48)
MATIE( n 0 ) 4 n 0 0.5
3 0
2
MAFE( n 0 ) 4 n 0 0.5
3 0
6 M 2i 1
pktfilteredFFO(t; Ms)
M s
yi m1 (M 2 1) M 1 , m =1,2,3, ., N M + 1 (I-51)
i 1
where:
t = m s
Figure I.13a Packet-selected two-way time error sequence and derived values
x ' (i)
1
X nmax max C (I-52)
1k N n1 n i k
for n = 1, 2, ..., N
where xC ' (i) is the packet-selected two-way time error (which represents time error and is a random
sequence), nS is the observation window length, n is the number of packet selection windows in the
observation window and consequently the number of pre-processed samples in the observation
window, S is the packet selection window length and consequently the time interval between delay
samples after the pre-processing step of packet selection, N is the total number of pre-processed
samples, and k is incremented for sliding the observation window. maxATE describes the maximum
value of average packet-selected two-way time error over an observation interval of length nS.
Similarly, minATE(n S) (''Minimum Average Time Error'') is defined as a specified percentile, , of
the random variable:
nk 1
x ' (i)
1
X nmin min C (I-53)
1k N n1 n i k
for n = 1, 2, ..., N
where the variables are defined as above. minATE describes the minimum value of average
packet-selected two-way time error over an observation interval of length nS.
Finally, ppATE(n S) (''Peak-to-peak Average Time Error'') is defined as a specified percentile, , of
the random variable:
nk 1 nk 1
for n = 1, 2, ..., N
where the variables are defined as above. ppATE describes the peak-to-peak value of average
packet-selected two-way time error over an observation interval of length nS.
Estimator formulas
For calculating maxATE the sliding window size is varied by sequencing n and determining the
maximum two-way time error for each value of n, thus creating maxATE as a function of sliding
averaging window width. In a similar fashion, minATE is calculated by determining the minimum
two-way time error as a function of sliding averaging window width. ppATE is calculated through a
point-by-point subtraction of maxATE minus minATE elements.
1 nk 1
maxATE nS max xC ' (i), (I-55)
1k N n1 n i k
1 nk 1
minATE n S min xC ' (i), (I-56)
1k N n1 n i k
ppATE nS = maxATE nS - minATE nS (I-57)
for n = 1, 2, ..., N
where S is the time interval between delay samples after packet selection. Thus S is the same as the
selection window size. N is the total number of samples and xC ' (i) is the packet-selected two-way
PDV limits specified in terms of these metrics are considered as met if at least m packets, or
alternately at least p% of packets, are observed for any window interval of t s within a fixed cluster
range starting at the observed floor delay and having a size . If fewer than m packets are observed,
or alternately less than p% of packets, then the PDV limit is considered as not met.
This process can be described in the following way:
Let x[i] represent the measured latency of timing packet i, where 0 i < N. That is, there are N packets
in the measurement data set. Let the nominal time between timing packets be represented by P. Let
represent the cluster range and let W represent the window interval in units of time, which can also
be expressed as K samples, where K = W/P. K represents the (nominal) number of packets transmitted
in the window interval.
NOTE It is assumed that the packet rate of the timing flow is nominally constant. The case for a variable rate
of packet transmission is for further study.
Define the minimum observed delay as:
(I-59)
d min min x[i]
0i N
The observed dmin given by Equation I-59 is an estimate of the absolute minimum latency that a packet
may experience. If a better estimate of the absolute minimum latency is available, for example from
previous measurement data, that alternate value may be used. In all cases, Equations I-60 to I-65 are
valid for choices of minimum delay less than or equal to the observed dmin.
Then, define the indicator function which performs floor packet selection:
1; if x[i] d min
F (i, ) for 0 i N (I-60)
0; otherwise
Note that Equation I-60 assumes that packet delay is always greater than dmin.
The convention followed in Equations I-61, I-62 and I-63 is that sample index n is associated with
the end of the window. That is, the floor packet metrics are based on complete windows and
consequently values of n less than (K 1) are not defined.
Then define the Floor Packet Count (FPC) sequence with parameters n, W (W = KP) and :
Define the Floor Packet Rate (FPR) sequence with parameters n, W and :
FPC(n,W , )
FPR(n,W , ) for ( K 1) n N (I-62)
W
Define the Floor Packet Per cent (FPP) sequence with parameters n, W and :
FPP(n, W , ) P FPC(n, W , ) 100 % for ( K 1) n N (I-63)
W
The floor packet per cent is applicable to defining network limits. That is, the network performance
is acceptable if
min FPP(n,W , ) p% (I-64)
( K 1) n N
where the network acceptance criterion is p%, and the parameters W and are provided in the
appropriate Recommendation, for example [ITU-T G.8261.1].
The floor packet rate (equivalently floor packet count) is a suitable metric for identifying the slave
clock tolerance limit. That is, the slave clock must meet its specified output performance if
min FPC(n,W , ) m (I-65)
( K 1) n N
where the parameters m, W and are provided in the appropriate recommendation as applicable.
Equations I-62, I-63, I-64 and I-65 are general and appropriate for sliding window approaches.
Jumping and overlapping window calculations can be obtained by sub-sampling the sliding window
samples.
For the jumping window case, estimates are derived every K samples. That is, the jumping window
samples are simply the sliding window estimates under-sampled by a factor of K. Over the full
measurement interval, there are M = (N/K) jumping window samples and consequently the index for
the jumping window sequence ranges from 0 through (M 1).
The jumping window approach is suitable when network conditions are stationary and spectral, and
probability density parameters do not change rapidly. The sliding windows may be more appropriate,
for example, for short term transient or rapidly changing events.
NOTE 1 This category of PDV metrics requires a long enough measurement period such that the observed
floor delay would give a good enough estimation of the absolute floor delay. The minimum measurement
period depends on the type of network considered. Long measurement periods, for instance over one or several
days, should be favored in order to study diurnal PDV effects.
NOTE 2 Like minTDEV and MAFE, these metrics may be sensitive to a small number of low-lying outliers.
The definition of the precise aspect that creates the potential sensitivity and the subsequent method of handling
this when applying this metric is for further study.
NOTE 3 This category of PDV metrics is sensitive to non-stationary network conditions as described in
clause I.3.3 that produce floor delay steps of significant amplitude, which may occur for instance during
network re-routing events. The handling of floor delay steps is for further study.
NOTE 4 These metrics are mainly intended to be used as post-processing metrics. The use of these metrics
for real-time processing is for further study.
NOTE 5 These metrics can be used to study the PDV noise produced independently by the forward or the
reverse direction of a packet timing flow. Consideration of the combined effect of both directions is for further
study.
The value of the floor packet metrics in Equations I-61 to I-63 at time n are then calculated using
dmin(n) instead of dmin, as shown in Equation I-66.
As an example for a specific implementation, the floor delay at time n can be iteratively estimated
according to the following algorithm:
1. Denote d(n) as the minimum packet delay of the most recent observation window;
2. Compare d(n) to the current estimate of the observed floor delay dmin(n)
a. If d(n) < dmin(n) dmin(n) = d(n)
b. Otherwisedmin(n) remains unchanged (I-67)
The progressive floor determination method continually refines the estimate of the observed floor
delay value during the measurement period. At each Floor Population metric computation point n,
the observed floor delay is estimated as the smallest delay value, dmin(n), in the measurement period
up to (and including) the window over which the metric is computed. The running value of this
Series E Overall network operation, telephone service, service operation and human factors
Series F Non-telephone telecommunication services
Printed in Switzerland
Geneva, 2016