ISPCS 2009 International IEEE Symposium on Precision Clock
Synchronization for Measurement, Control and Communication
Brescia, Italy, October 12-16, 2009
Packet Delay Variation Management
for a better IEEE1588V2 performance
Dinh Thai BUI, Arnaud DUPAS, Michel LE PALLEC
Semantic and Autonomic Technologies (SAT) Department
Alcatel-Lucent Bell Labs France (ALBLF)
Nozay, FRANCE
Dinh_Thai.Bui@alcatel-lucent.com, Arnaud.Dupas@alcatel-lucent.com, Michel.Le_Pallec@alcatel-lucent.com
Abstract— Designed for the distribution of time, IEEE 1588V2
also called as Precision Time Protocol V2 (PTPV2) is often
referred to as a key candidate for synchronization over Packet
Switched Networks (PSNs). Unfortunately, as the IETF's
Network Time Protocol and unlike the ITU-T's Synchronous
Ethernet, the performance of this protocol strongly depends on
Packet Delay Variations (PDVs) produced by the telecom
network environment. Such behavior represents a strong issue,
particularly considering the severe synchronization requirements
of the mobile backhaul segment. This article focuses on different
manners to control and to fight out PDVs. After discussing PDV-
management principles, via the illustration of some existing
mechanisms, we propose a new distributed ranging scheme which
allows for optimizing the overall PTPV2 synchronization
performance. Finally, considered as a critical mobile backhaul
use case, IEEE 1588V2 over Digital Subscriber Line (DSL) is
investigated regarding PDV concerns.
Keywords: synchronization protocols, 1588V2, mobile
backhaul, delay, delay variation, DSL
I. INTRODUCTION
IEEE 1588V2 offers the possibility to distribute both
frequency and time. Even intimately intertwined, these two
functionalities are studied separately at the ITU-T Study Group
15 Question 13 which aims at defining IEEE1588V2 telecom
profiles.
Taking time distribution as an example, PTPV2 can be
deployed either in a one-way or in a two-way signaling
approach. The former approach implies an “on-path” correction
(see Transparent Clock description in section IV.A) of the time
offset induced by the path delay. Alternatively - and this why it
is often preferred - the two way approach relies on an offset
delay correction based on a mean-delay knowledge (half of the
round-trip delay). This approach is thus conceptually
dependent on the delay asymmetry between the Master-to-
Slave and Slave-to-Master synchronization signalization
directions.
This delay asymmetry has two components, a fixed and a
variable one. The fixed and deterministic component is mainly
driven by the link delay. This component can generally be
determined by simple computation means (e.g. optical
transmission index for Gigabit Passive Optical Network) or by
using advanced methodologies [1]. The variable character of
978-1-4244-4392-5/09/$25.00 ©2009 IEEE
the delay asymmetry is mainly driven by PDVs generated by
communication paths. While a deterministic value of the delay
asymmetry can still allow for a correct computation of the time
offset, the random component, essentially induced by PDVs,
leads to the time offset uncertainty and error.
Thus, the overall PTPV2 performance greatly depends on
PDVs generated by the PSN in each communication direction.
After a brief description of different PDV components, we
illustrate some principles which enable the management of
PDVs. Such methods are presented following the hierarchical
synchronization architecture: emission profiles are firstly
discussed followed by a presentation of network-based
approaches. Then, Slave clock mechanisms are depicted.
Following this “prior art” description, a new distributed
ranging mechanism combining the functionalities of the three
aforementioned approaches (Master/Network/Slave-based) is
presented.
Finally, IEEE1588V2-over-DSL is investigated as an
illustration of a difficult PDV use case.
II. PDV COMPONENT DESCRIPTION
Delay and thus related delay variation can be divided into
three different components [2] [3].
- The transmission delay is the result of the velocity of the
signal between two endpoints of a transmission link and the
distance between these two endpoints. The transmission delay
variation is due to many factors depending on the transmission
technology employed (e.g. wireless, wire-line, etc). For
example on a copper link, temperature variations can modify
the transmission delay by shrinking or extending the
transmission distance.
- The processing delay is the delay resulting from the
processing of a timing packet within a Network Element (NE).
It strongly depends on the NE hardware and software. For
example, the processing delay variation can result from a
specific CPU (Central Processing Unit) configuration: a same
CPU on a given port could be in charge of in-going and out-
going packets. Consequently the traffic load in one direction
can have a detrimental effect on the other direction.
 - The buffering or queuing delay is the total amount of
waiting time of the timing packet within different buffers or
queues in a given NE before being processed and finally
transmitted to the next NE on the communication path.
Buffering delay variation is partially due to the “competition”
(see chapter IV.B) between timing packet flows and other
packet flows, and between different timing packet flows
themselves according to their relative arrival time at different
waiting queues and the priority policies implemented.
These three components have different PDV ranges.
Usually, the first component generates a PDV in the range of
sub-microseconds (especially considering temperature
dependence) and the second one in the [1-10] micro-second
range. The last component is often referenced as the main PDV
generator with a [10-10000] micro-second range.
III. EMISSION PROFILES
This approach consists in defining the “profile” to be
implemented at the 1588V2 Master level. Reference [6] gives a
definition of what is a profile: “The set of allowed Precision
Time Protocol features applicable to a device”. Here we
particularly focus on the message emission profile which is the
first tool for managing the PDV as described in the following
examples.
1) Timing packet rate
Timing message rate is one of the main parameters
enabling to dampen PDV effects. Indeed, a higher number of
samples (timing packets) per time unit allow for improving the
time or frequency accuracy at the Slave level. Figure 1
(Frequency accuracy vs PDV at different packet rates), based
on [4] equations in a first order approach, illustrates this
purpose taking a common 5 mHz Phase Locked Loop (PLL)
filter. One can observe, for instant, a factor 10 gain in term of
frequency accuracy with an emission rate increase from 2
messages/s to 128 messages/s (note that IEEE 1588V2
specifies a maximum message rate of 128 messages/s).
Targeting an often-mentioned 15 ppb frequency accuracy at the
network level for wireless requirements (e.g. WCDMA, LTE),
the [10-100] µs PDV range appears to be the recommended
limits.
1E-05
PLL cutoff frequency = 5 mHz
1E-06
Frequency accuracy
1E-07
1E-08
15 ppb wireless requirement
2 packets/s
1E-09 16 packets/s
128 packets/s
1024 packets/s
1E-10
1 10 100 1000 10000
Packet Delay Variation (µs)
Figure 1 - Frequency accuracy according to the timing packet rate for a
5 mHz PLL (Phase Locked Loop) filtering
The counterpart of this approach is its impact on the
bandwidth consumption, which is especially critical for low
bit-rate communication paths. It also implies an additional
CPU power consumption at the Master level. Moreover, packet
correlation has to be taken into account for high timing
message rate [5].
2) Randomization:
As it comes to syntonization, one often makes use of a
periodic PTPV2 stream (i.e. one-way periodic Sync messages).
In this case, it is demonstrated in [8] that PTPV2 streams with
similar transmission frequencies compete against each other
leading to additional PDVs. This phenomenon is often referred
to as a “beating effect”. In order to solve such an issue, a
simple but powerful method consists in randomizing the
transmission instant of PTPV2 messages around a given mean
value.
However, this technique breaks out the periodicity of the
PTPV2 stream and could lead to some incompatibility with
certain Slave clocks. Indeed, these latter could make use of the
periodicity of the PTPV2 stream in their filtering/ranging
process as described in chapter V.B.
IV. NETWORK-BASED APPROACHES
A. Specified IEEE 1588 V2 mechanisms
The IEEE has specified hardware components to correct
PDVs [6].
Boundary Clocks (BCs) allows for segmenting a large
synchronization network into small areas (within a PTP
domain) where PDVs can be engineered and/or controlled
within specified bounds. BCs serve as means to recover the
reference time (or frequency) as accurately as possible before
distributing the latter into next (per-PDV) areas of the
synchronization hierarchy.
End-to-End Transparent Clocks (E2E TC) allows for
correcting transit delays of the network elements whereas Peer-
to-Peer Transparent Clocks (P2P TC) allow for compensating
both transit delays and link delays. The first one allows to
record, within the message “correction field”, its (cumulative)
resident time within each NE and allows for the Slave clock to
compensate PDVs produced by queuing and processing
processes. The second one measures not only the NE resident
time but also link delays. Both information allows the Slave to
completely correct uncertainties due to PDV and thus to reach
accurate time/frequency references at the Slave level.
It is noted that there are no syntonization requirement on
TCs. Indeed, considering transit and link delays in the order of
milliseconds, it clearly appears that low cost local clocks with
some ppm (part per million) accuracy are sufficient to meet
measurement requirements (measurement error in the order of
nanoseconds). Moreover in the case of link delay measurement,
it also appears that P2P TCs cannot solve by themselves the
link delay asymmetry and that specific mechanisms are
necessary for such a purpose (for instance a delay asymmetry
ratio could be provisioned via the synchronization management
plane).
 Whilst TCs represent a very effective solution to fight out
PDVs, they represent some non negligible challenges for
network operators considering a full deployment. The first one
is the cost related to the implementation of such devices within
the NEs along all PTPV2 signalization paths. The second
important one is rather a technical challenge linked to the fact
that TCs represent an infringement to the layer separation
principle. Indeed, for precision purpose, TCs require timing
monitoring operations (e.g. hardware time stamping) being
realized at the physical layer while reporting the monitoring
results (e.g. packet resident time) at the PTPV2 packet layer.
Nevertheless, those constraints have to be re-visited in the
particular scenario of “green field” deployments targeting the
distribution of both accurate time and frequency references.
B. Quality Of Service (QoS) approach
The principle here consists in differentiating timing
signalization flows from other packet flows using QoS-build-
in information within the packet header. This allows for
prioritizing synchronization packets with regards to data
packets, and avoiding a direct competition of the latters
against the formers. QoS technique usually relies on parallel
buffers which filling depends on sheduling policies.
This technique allows to substantially reduce the PDV
induced by queuing effect for a given packet flow, especially
for high traffic load conditions, but without suppressing this
component completely [7].
Indeed, the method does not fully prevent low priority data
packets from competing with high priority timing packets.
Even with separated queuing buffers, QoS technique cannot
avoid competition at the physical port when it comes to the
serialization of packets onto the line. For instance, the timing
packet waiting times, induced by already departing Best Effort
packets (from another queue) onto the transmission line
(physical layer level), can greatly vary depending on the data
packet length and its relative arrival time at the physical output
medium. This phenomenon is often referred to as the (data)
“jumbo packet” phenomenon.
Another phenomenon that QoS technique cannot solve is
the competition between timing packet flows with a same QoS
priority. In case of periodic timing flows having the similar
transmission frequencies, such a competition produces the
“beating effect” [8].
C. Delay Equalization
Delay equalization is a well-known technique especially
implemented within ATM switches for the interconnection of
different SDH/SONET transmission lines. The idea consists in
adding a time label in order to track the node delay
experienced by ATM cells. At the output port, this makes it
possible to “equalize”, via a buffer, node delays originating
from the dismantling of the same (SDH/SONET) input
payload. Delay equalization is quite powerful to fight out
PDVs related to both processing and queuing, as different
packets pertaining to the same flow undergo the same node
delay.
However, in this specific scenario (retiming at the bit level),
delay equalization requires the distribution of a precise clock
across all the NE ports. This pre-requisite supposes a high
accurate, thus high cost, local clock especially if the latter is
left into a free-run or holdover mode.
V. SLAVE CLOCK MECHANISM
A. Packet Clock Models
Timing packet clock generally has two-stage architecture as
depicted by the Figure 2 below:
Figure 2 – Packet Clock Model
The first stage is a packet filtering stage designed for
selecting the most relevant packets while reducing PDV
effects. The packet filtering or selection process is also called
as “ranging” process.
The second stage aims at disciplining a local oscillator or
wall clock according to the filtered “packet signal” resulting
from the first stage.
B. Ranging Mechanisms
When syntonization is considered, the PTPV2 Slave deals
with periodic one-way PTPV2 Sync messages. The ranging
process in this case often consists in assessing the inter-arrival
times between successive Sync messages by averaging over a
large number of samples. Two methods are considered here
making the assumption that delays are independent and
identically distributed: the Moving Average (MA) and the
Exponentially Weighted Moving Average (EWMA) methods
which respectively computes means or weighted means on a
limited set of stored inter-arrival times. Although conceptually
simple with basic components in its implementation, the MA
method has the disadvantage of requiring a large number of
registers to achieve low jitter levels in the recovered clock
signal (impacting the convergence time). The EWMA partially
solves this issue at the expense of being very sensitive to
packet delay correlation. Indeed, the aforementioned
assumptions on delay distribution are not realistic as this latter
is observed to be more or less correlated with time. And it has
been demonstrated that EWMA method becomes rapidly
inefficient as the correlation between successive packets
increases [11][12]. It is also noted that both MA and EWMA
performance is predicted while assuming a time-invariant
(stationary) packet delay distribution which is not always the
case in real network.
Recent works are thus more focused on ranging
mechanisms which selects packets experiencing minimum
(alternatively maximum) packet delays [13][14] within an
observed window in order to dampen PDV effects. New related
metrics, such as MinTDEV [15], etc, have been defined for this
purpose. A performance comparison between these
mechanisms has been provided through [16], but no decision
has been taken within standardization organization in terms of
recommendation.
It is noted that the ranging process can give different results
regarding a time or a frequency distribution purpose. Indeed if
‘p’ is the probability to select a packet with a minimum delay
for a single transmission direction (such as the one-way
approach frequency distribution) the probability to select
packets experiencing minimum delays in the two transmission
directions (such as the two-way signaling approach) is ‘p²’ [9].
This implies strong constraints on the local oscillator
supporting the ranging process (during the ranging time all
packets have to be selected with the same frequency reference).
In such a case the timing packet rate has to be carefully
configured.
VI. NEW DISTRIBUTED RANGING MECHANISM
A. Principle description
This new approach relies on a distributed selection/filtering
of relevant (low delay) timing packets along the
synchronization path thanks to a given synchronization
requirement at the receiver level. The idea particularly enables
to increase the timing message rate with a limited impact on the
consumed bandwidth. In a simple configuration, each filter
(including the receiver) has the ability to trigger an increase of
the timing source message rate according to a provisionable
timing message rate threshold. Figure 3 below depicts the main
principles of this approach.
Figure 3 – Ranging Mechanism Distributed along the Synchronization Path
In such a representation, “transparent nodes” are
intermediate nodes which do not implement any PTPV2
features. Filtering modules have the ability to recognize
PTPV2 packets and to proceed to the ranging process.
For each aforementioned filter module, a ranging process is
applied in order to select packets demonstrating low PDV
(often based on packets experiencing minimum delays). If the
number of these selected packets is lower than a certain
provisioned threshold (e.g. tm1 for filter 1), the filter module
requests the Master for increasing the message emission
frequency. In the opposite way, a second threshold can be
provisioned so that the filter module can ask the Master for
decreasing the message emission frequency when the number
of selected packets is found to be above this second threshold.
This method guarantees that the Slave receives sufficient
relevant packets during its filtering process in order to feed its
second stage and to discipline its local oscillator or wall clock.
One possible way to change the Master message emission
frequency is to use a management message with a
PORT_DATA-SET management Type Length Value (TLV)
data field (this latter includes other TLVs defining different
message emission frequency values, such as logSyncInterval).
Alternatively, filter modules can make use of the
REQUEST_UNICAST_TRANSMISSION TLV to modify the
Master message emission frequency when unicast profile is
deployed.
In order to optimize the consumed bandwidth by
downstream NEs, non-selected synchronization messages
having their PDV range over a pre-defined bound can be
dropped and thus not forwarded towards the Slave.
Facing heavy traffic load conditions, the presented
approach enables to tune the individual timing packet filtering
ratio in order to ensure the best synchronization performance in
a degraded state, for which the synchronization requirement at
the Slave level is not reachable. This “less bad state” approach
could be relevant is some congestion scenarios for which the
synchronization topology reconfiguration time plays a
significant role.
From an implementation perspective, the timing packet
filtering functionality could be integrated either into a Network
Element, or into an external module placed onto the path
between two network nodes.
It is to be noted that these distributed filters can take
advantage of an external syntonization for a precise monitoring
of the selection process but similarly to the TC functionality,
this does not appear critical (some ppm frequency accuracy
should be sufficient to cover stringent synchronization
requirements)
B. Discussions on the mechanism benefits
In order to illustrate the benefits of the distributed ranging
mechanism, two representative PDV examples are taken (PDV
here is defined with regards to the minimum delay) with related
Probability Density Functions (PDF) as depicted in Figure 4.
These PDF examples, reflecting “high” and “low” traffic load
conditions as defined by the ITU-T G.8261 recommendation
[8], arbitrary follow a Gamma density function already
discussed within the ITU-T [10]. For low and high traffic
loads, the computed standard deviations are respectively 7.1 µs
and 20 µs ( α .β 2 with regards to Figure 4 examples).
An arbitrarily configured PDV threshold filtering of 10µs
per filter module is considered. This scenario particularly
means that, without any filter module, timing packet VII. DSL USE CASE
demonstrating a PDV over this 10 µs reference value have a Attractive DSL technologies are used to backhaul data and
very poor probability to be retained by the first filtering stage signalization of the wireless networks. For such a purpose, they
of the end receiver packet clock. The following discussions have to face some syntonization requirements, such as the
take the first filtering module as an example but one can distribution of a frequency reference within accuracy better
reiterate the arguments through successive filtering modules than 50ppb (base station syntonization requirement). Meeting
downstream. this requirement, SH-DSL has been the first main deployed
DSL technology. Network Time Reference (NTR) has been
specified for accurately distributing a frequency reference at
Gamma Law: the physical layer.
1E+0 Unfortunately, even if it is specified, NTR has not really
Filtered area been deployed in commercial DSLAMs. Moreover, in some
Probability Density Function

1E-1
High traffic load
configurations, such as pseudo-wire based backhaul (Figure 5),
60% Low traffic load
the DSLAM could be considered as transparent node in terms
1E-2 of syntonization and NTR can not be used consequently.

1E-3
98% α =4, β =10
40%
1E-4
2% α =2, β =5
1E-5
0 20 40 60 80 100
10 µs filtering Packet Delay Variation (µs)

Figure 5 - Pseudo-wired based Mobile backhaul scenario

Figure 4: examples of PDV Probability Density Functions for high and low
traffic load conditions
To face this issue (and for lowering the syntonization cost),
Within this configuration, a low traffic load implies a low equipment vendor try to monitor syntonization at the packet
PDV range and a better probability (e.g. 60%) to have an layer by taking advantage of timing protocols such as PTPV2.
arriving timing message selected by the first filter module. In
this case, the number of messages selected should be sufficient Unfortunately, in DSL environments, the “syntonization
(above the provisioned threshold) and thus no feedback noise” in the downstream (Master-to-Slave) direction implies
mechanism towards the sender (e.g. master clock) is needed in very hard constraints on the Slave features. Indeed, as referred
order to increase the transmission rate. In this scenario, one can to by Symmetricom [4], the observed PDV is very high -
observe a 40% timing-message bandwidth saving at the usually in hundreds of microsecond range even under no load
downstream link next to the first filter module. conditions (see [17]).
At high traffic load condition, the probability to have a
timing message selected by the first module is strongly Mean Delay Mean Delay Stddev PDV Stddev PDV
decreased (e.g. down to 2%). Thus the retained messages are Fwd Rev Fwd Rev
not to be of sufficient number with regards to the module
provisioned threshold. As described in the previous chapter, Switch 21.9 µs 21.3 µs 100 ns 100 ns
one possible policy in this case consists in loosing
synchronization constraints (e.g. degraded state) in order to VDSL 1.76 ms 1.54 ms 104 µs 166 µs
avoid traffic load increase. Alternatively it can be decided to SHDSL 3.60 ms 3.55 ms 89 µs 109 µs
trigger the feedback mechanism for an increase of the timing
message transmission rate. This second process has to be ADSL2+ 15.67 ms 14.65 ms 289 µs 196 µs
carefully calibrated according to the congestion state margin.
Indeed such a mechanism can significantly deteriorate the node
state with a final performance improvement lower than TABLE I. DELAY AND PDV MEASUREMENTS OF SYMMETRICOM FOR XDSL
expected; linked for instance to packet loss. Congestion control TECHNOLOGIES IN NO LOAD CONDITIONS [17]
and/or rerouting are consequently other techniques to be
considered for solving such an issue. Nevertheless, they have
to be well engineered by network operators with regards to:
- the priority allocated to the synchronization traffic;
- the policies which drive topology changes (avoiding
congestion points), as they usually involve signal transients
impacting the synchronization performance.
 Moreover, due to its periodic time-slotted structure (i.e. 250
µs symbol period), a DSL link can lead to a “beating effect” Authors would like to thank Jean Loup Ferrant and Nicolas
with a periodic PTPV2 emission profile as related in [8]. The Le Sauze for their valuable comments and constant support.
“beating effect” imposes a “sawtooth” delay profile onto the
PTPV2 stream as illustrated in Figure 6 below, under no load
condition.
[1]
[2]
[3]
[4]
[5]
[6]
[7]
Figure 6 – Systematic delay variation caused by DSL timeslots [8]
[8]
The “beating effect” usually implies a strong time [9]
dependency which puts more constraints on the Slave clock
(e.g. stability). On the whole, and as discussed in III.1), such
PDV profiles imply high cost low-pass filters with cut-off [10]
frequencies within the milliHertz and even sub-milliHertz
range or require to go beyond the specified 128-messages-per- [11]
second limit for the PTPV2 packet rate. Ranging algorithms
discussed in V.B should allow for relaxing clock constraints. [12]
But if one is targeting a low-cost and robust PTPV2-based
frequency and time distribution, an important work at the DSL
physical layer seems to be the only way out. [13]
VIII. CONCLUSION [14]
There are numerous manners to manage PDVs within a
telecom network. A limited set of powerful approaches have [15]
been presented with their strengths and weaknesses. Stating
that all improvements performed at the Slave level to manage [16]
PDV could not meet all deployment scenarios and application
requirements, network-based solutions have to be considered [17]
with careful studies on deployment cost. In this context, a new
distributed mechanism is proposed. This new mechanism
particularly enables to increase PTPV2 message rate - and thus
to improve PTPV2 performance for a given PDV profile - with
a limited impact on the available network bandwidth.
As an illustration to the whole PDV monitoring discussion,
IEEE1588V2 over DSL has been investigated. A simple
analysis of this difficult use case demonstrates the need to
define a set of compatible and optimal procedures for an
efficient use of PTPV2. For such, monitoring the PDV at the
Master side by an appropriate emission profile appears to be an
attractive approach. This latter will likely require to adapt
Slave algorithms in order to manage packet correlation aspects.
Finally, without any obvious packet-based solution, the
need to directly manage the DSL physical layer for a lower
PDV is particularly growing. This work is to be addressed by
the ITU-T.
ACKNOWLEDGMENT
REFERENCES
An Enhanced IEEE 1588 Time Synchronization Algorithm for
Asymmetric Communication Link using Block Burst Transmission
IETF RFC5481 - Packet Delay Variation Applicability Statement –
March 2009.
ITU-T SG15/Q13 WD35 - Effect of routers’architecture on timing
distribution - San Jose, March 2009.
K. Shenoi, Symmetricom - Sync for Support of Legacy Services – ITSF
London, October 2005.
Q. Li, D.L. Mills – The Implication of Short-Range Dependency on
Delay Variation Measurement. IEEE Network Computing and
Applications. 2003.
IEEE Standard for a Precision Clock Synchronization Protocol for
Networked Measurement and Control Systems – 24 July 2008.
ITU-T SG15/Q13 WD38 - Additional PDV measurements on Ethernet
platform, Edinburgh, June 2009
ITU-T G.8261/Y.1361 - Timing and synchronization aspects in packet
networks – April 2008.
Y. Stein et al, Rad Data Communications - Delivering Better Time-of-
Day Using Synchronous Ethernet and 1588 – ITSF London, November
2007.
ITU-T SG15/Q13 C327 - Modelling and simulation of long-tailed packet
delay variation (PDV) over a multi-switch network, May 2007.
J. Aweya, D.Y. Montuno, M. Ouellette, K. Felske – Clock recovery
based on packet inter-arrival time averaging. Computer Communications
29. 2006.
J. Aweya, D.Y. Montuno, M. Ouellette, K. Felske – Analysis of a clock-
recovery technique for circuit emulation services over packet networks.
International Journal of Communication Systems. 2008.
S.B. Moon, P. Skelly, D. Towsley – Estimation and Removal of Clock
Skew from Network Delay Measurements. Proc. IEEE Infocom, March
1999.
L. Zhang, Z. Liu, C.H. Xia – Clock Synchronization Algorithms for
Network Measurements, Proc IEEE Infocom, June 2002.
ITU-T SG15/Q13 C363 - Definition of Minimum TDEV (minTDEV) –
Geneva, June 2007.
ITU-T SG15/Q13 WD36 - Packet selection methods for performance
metrics- San Jose, March 2009.
ITU-T SG15/Q13 WD45 - xDSL Access PDV Study - San Jose, March
2009.