d2.4 synchronization service in carrier-grade ethernet...
TRANSCRIPT
D2.4
Synchronization service in Carrier-Grade
Ethernet environments
Abstract:
In the context of mobile backhaul network migration from Time Division Multiplexing (TDM)
technologies towards Packet Switched Networks (PSNs) supported by Carrier-Grade Ethernet
as one of the preferred technologies, frequency and time distribution represents a major
challenge. This paper presents the work carried and delivered by Task 2.4.
After briefly reviewing the work accomplished within the mid-term report (M24), this paper
moves to thoroughly discuss the new contributed work focusing on the study of the various
factors that impact time/frequency distribution in PSNs as well as new apparatuses to
guarantee performance compliance with the new requirements imposed by the new cellular
technologies.
Editors:
Alon Geva (RAD Data Communications Ltd)
Authors:
Dinh Thai Bui (ALBLF), Michel Le Pallec (ALBLF), Sebastien Jobert (FT), Alon Geva (RAD)
Date of publication:
October 31th 2010
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 2/76
Table of Content
1 EXECUTIVE SUMMARY ................................................................................ 9
2 INTRODUCTION .......................................................................................... 11
3 TERMINOLOGIES ....................................................................................... 13
4 MOBILE NETWORK SYNCHRONIZATION REQUIREMENTS .................... 14
5 KEY PROTOCOLS AND THEIR RELATED DEPENDENCIES ....................... 15
5.1 Protocol overview ............................................................................................................................ 15
5.2 Sync-E constraints and performance dependencies ..................................................................... 16
5.3 Timestamp protocols constraints and performance dependencies ............................................. 16
5.3.1 NTP and PTPV2 comparison ...................................................................................................... 16
5.3.2 IEEE 1588V2 constraints and performance dependencies ...................................................... 16
6 DEPLOYMENT SCENARIOS ........................................................................ 19
6.1 Mobile Operator operating all backhaul segments and Sync-E fully deployed ........................... 19
6.2 Mobile Operator operating all backhaul segments and Sync-E partially deployed ................... 20
6.3 Mobile Operator not operating all backhaul segments ................................................................. 21
6.4 Operator deployment focuses ......................................................................................................... 21
7 IMPROVING TIMING-OVER-PACKET (TOP) .............................................. 23
7.1 Summary of work accomplished in M24 ...................................................................................... 23
7.2 Study of Packet Delay Variation (PDV) accumulation over packet networks, for the purpose of
delivering frequency synchronization ....................................................................................................... 25
7.2.1 The need for developing engineering rules for packet based methods ................................... 25
7.2.2 Analysis of PDV generated by telecom networks ..................................................................... 26
7.2.3 Model of a packet slave clock .................................................................................................... 32
7.2.4 PDV accumulation study ........................................................................................................... 33
7.2.5 Conclusion (for clause 7.2) ........................................................................................................ 42
8 TIME AND PHASE DISTRIBUTION OVER PACKET NETWORKS WITH
FULL OR PARTIAL NETWORK SUPPORT .................................................................... 43
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 3/76
8.1 Summary of work accomplished in M24 ...................................................................................... 43
8.2 LTE IPSec challenging environment ............................................................................................. 45
8.2.1 Introduction ............................................................................................................................... 45
8.2.2 LTE IPSec environment description ......................................................................................... 45
8.2.3 Traditional OPS deployment discussion .................................................................................. 46
8.3 New protocol-agnostic mechanism to monitor the DA ............................................................... 47
8.3.1 Principle description with a first proposed implementation .................................................. 47
8.3.2 Alternative implementation of the Delay Asymmetry mechanism ......................................... 50
8.3.3 Conclusion .................................................................................................................................. 50
8.4 New protocol-agnostic mechanism to dampen PDV .................................................................... 51
8.4.1 Packet Delay Variation (PDV) root causes ................................................................................ 51
8.4.2 Case of single synchronization flow .......................................................................................... 52
8.4.3 Case of multiple concurrent synchronization flows ................................................................. 53
8.4.4 Discussions on implementation and conclusion ...................................................................... 55
8.5 Conclusions and outlook (Clauses 8.2 to 8.4) .............................................................................. 55
8.6 Link-to-link TOD distribution using the Sync-E’s ESMC messaging channel ........................... 56
8.6.1 Context description .................................................................................................................... 56
8.6.2 The ESMC ....................................................................................................................................57
8.6.3 Proposals .................................................................................................................................... 59
8.6.4 The Simulation Environment .................................................................................................... 63
8.6.5 Simulation Results ..................................................................................................................... 70
8.6.6 Conclusions ................................................................................................................................ 74
9 CONCLUSION ............................................................................................. 76
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 4/76
Index of Tables
Table 1 – Mobile networks: synchronization requirements at the air interface level ............. 14
Table 2 – ESMC PDU format (Table 11-3/G.8264) ................................................................. 58
Table 3 – QL TLV format (Table 11-4/G.8264) ....................................................................... 59
Table 4 – TOD enhanced ESMC PDU format (proposed changes are highlighted)............... 60
Table 5 – TOD TLV format (proposed changes are highlighted) ............................................ 61
Table 6 – Signal structure of the ESMC signal API ................................................................. 65
Table 7 – Signal structure of the Time/Frequency reference signal API ................................66
Table of Figures
Figure 1 – PTPV2 principles .................................................................................................... 17
Figure 2 – MO owning all backhaul segments with Sync-E fully deployed ............................ 19
Figure 3 – MO owning all backhaul segments with Sync-E partially deployed ..................... 20
Figure 4 – MO not owning all backhaul segments .................................................................. 21
Figure 5 – Distributed ranging mechanism............................................................................. 24
Figure 6 – SSM Tunnel across ToP domain ............................................................................ 25
Figure 7 - Load pattern and resulting delay steps generated during congestion periods (figures extracted from [16]) ........................................................................................... 28
Figure 8 - Delay steps generated with reasonably low levels of load (figure extracted from [20] and [24]) ................................................................................................................... 29
Figure 9 - Comparison of PDV histograms with lowest priority (first plot - peak-to-peak = 302µs), intermediate priority (second plot -peak-to-peak = 161µs), and highest priority (last plot - peak-to-peak = 43µs) for the timing packets in case of congestion (figures extracted from [16]) ......................................................................................................... 30
Figure 10 - PDV measurement when changing the size of the background traffic (figure extracted from [16]) .......................................................................................................... 31
Figure 11 - Comparison of PDV histograms generated by a core equipment (first plot - peak-to-peak = 302µs) and by an access equipment (second plot - peak-to-peak = 1.55ms) when timing packets are not prioritized in case of congestion (figures extracted from [16])................................................................................................................................... 32
Figure 12 - Figure A.1/G.8263 draft - Function model of a PEC-S packet-based clock (source of the data: International Telecommunication Union, ITU-T Working Party 3/15 Temporary Document 318, G.8263 latest draft, 16 May 2010) ...................................... 33
Figure 13 - Illustration of a 60µs floor delay window .............................................................. 34
Figure 14 - Network PDV accumulation model ....................................................................... 35
Figure 15 - PDV of PTP Sync messages and corresponding histograms ................................. 37
Figure 16 - Configuration with 2 nodes used during the experimentation ............................ 38
Figure 17 - Histograms and Cumulative Distribution Functions comparison at 80% of traffic load ................................................................................................................................... 39
Figure 18 - Analysis of successive convolutions ...................................................................... 41
Figure 19 - Automatic computation of the PTPV2 topology thanks to an improved BMC taking into account a local Sync-E support ......................................................................44
Figure 20 – NTP ADD/DROP concept ....................................................................................44
Figure 21 - IPSec tunnel deployment in LTE ...........................................................................46
Figure 22 - IPSec ESP tunnel mode .........................................................................................46
Figure 23 - BCs deployment scenario within a LTE IPSec environment ................................ 47
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 5/76
Figure 24 - First way to implement the new DA monitoring mechanism .............................. 48
Figure 25 - Alternate implementation of the new DA monitoring mechanism...................... 50
Figure 26 – Variation of delay with loading, as referred to G.8061 (Figure I.2) ..................... 51
Figure 27 – “Premium” packet transmission delayed by a “Best Effort” packet ..................... 52
Figure 28 - Beating effect as referred to G.8261 (Figure 20) .................................................. 52
Figure 29 - Timing flows delayed by a multiple of the elementary delay ∆ms ........................ 54
Figure 30 – Possible implementation of the new PDV dampening mechanism ..................... 55
Figure 31 – TOD ESMC messages exchange ........................................................................... 62
Figure 32 – node prime functional block inputs/outputs .......................................................64
Figure 33 – node prime functional block internal structure ................................................... 67
Figure 34 – PRTC prime functional block .............................................................................. 68
Figure 35 – Time Error Analyzer prime functional block ......................................................69
Figure 36 – Time error scope ..................................................................................................69
Figure 37 – Steady-state test setup ......................................................................................... 70
Figure 38 – Steady state test: distribution chain convergence process (node1=red, node2=cyan, node3=magenta, node4=yellow, node5=blue, node6=green, node7=red, node8=cyan, node9=magenta, node10=yellow) ............................................................ 71
Figure 39 – Steady state test: distribution chain steady-state performance (node1=red, node2=cyan, node3=magenta, node4=yellow, node5=blue, node6=green, node7=red, node8=cyan, node9=magenta, node10=yellow) ............................................................ 72
Figure 40 – Ring restoration test setup .................................................................................. 72
Figure 41 – Ring restoration process (node1=red, node2=cyan, node3=magenta, node4=yellow, node5=blue, node6=green, node7=red, node8=cyan, node9=magenta, node10=yellow) ................................................................................................................ 73
Figure 42 – Steady state test: distribution chain steady-state performance (node1=red, node2=cyan, node3=magenta, node4=yellow, node5=blue, node6=green, node7=red, node8=cyan, node9=magenta, node10=yellow) ............................................................ 74
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 6/76
Acronym list
1588v2: Std 1588TM-2008
BC: Boundary Clock
BMCA: Best Master Clock Algorithm
CAPEX: CAPital Expenditure
CES: Circuit Emulation Service
CGE: Carrier-Grade Ethernet
DNU : Do Not Use
DSLAM: Digital Subscriber Line Access
Multiplexer
EWMA: Exponentially Weighted Moving Average
E2E: End-to-End
ESMC: Ethernet Synchronization Messaging
Channel
ESSM: Ethernet Synchronization Status Messages
FDD: Frequency Division Duplexing
GbE: Gigabit Ethernet
HSDPA: High Speed Downstream Packet Access
HSUPA: High Speed Upstream Packet Access
IEEE: Institute of Electrical and Electronic
Engineers
IETF: Internet Engineering Task Force
IP: Internet Protocol
ISP: Internet Service Provider
LAN: Local Area Network
MA: Moving Average
MO: Mobile Operator
NE: Network Element
NodeB: UMTS BTS
NTP: Network Time Protocol
OPEX: OPerational Expenditure
OUI: Organization Unique Identifier
OSSP: Organization Specific Slow Protocol
P2P: Peer-to-Peer
PDF: Probability Density Function
PDH: Plesiochronous Digital Hierarchy
PDV: Packet Delay Variation
PDU: Payload Data Unit
PLL: Phase-Locked-Loop
ppb: part-per-billion
ppm: part-per-million
PPS: Pulse per Second
PRC: Primary Reference Clock
PSN: Packet Switched Network
PTP: Precised Time Protocol (other name of IEEE
1588 protocol)
QoS: Quality-of-Service
QL: Quality Level
RoI: Return-on-Investment
SDH: Synchronous Digital Hierarchy
SN: Sequence Number
SSM: Synchronization Status Message
SSF: Server Signal Fail
STM-1: Synchronous Transport Module 1
(155.44Mbit/s)
STP: Spanning Tree Protocol
Sync-E: Synchronous Ethernet
TICTOC : Timing over IP Connection and
Transfer of Clock
TC: Transparent Clock
TDD: Time Division Duplexing
TDM: Time Division Multiplex
TLV: Type Length Value
TE: Traffic Engineering
ToD: Time-of-the-Day
ToP: Timing-over-Packet
TQL: Time Quality Level
TS: Time Stamp
Tx: transmitter
UMTS: Universal Mobile Telecommunications
System
LTE: Long Term Evolution (3GPP 4G)
UMTS: Universal Mobile Telecommunication
System (3GPP 3G)
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 7/76
References
[1] K. O’Donoghue - Synchronization in Packet networks – NTPv4 and IEEE 1588 - International Telecom Sync Forum (ITSF) 2007, November 13th 2007.
[2] IETF RFC5481 - Packet Delay Variation Applicability Statement – March 2009. [3] ITU-T SG15/Q13 WD35 - Effect of routers’ architecture on timing distribution - San Jose, March
2009. [4] 3GPP TS 33.210 version 9.0.0 Release 9 – Network Domain Security (NDS); IP network layer
security. ETSI TS 133 210 V9.0.0. February 2010. [5] 3GPP TS 33.310 version 9.1.0 Release 9 – Network Domain Security (NDS); Authentication
Framework (AF). ETSI TS 133 310 V9.1.0. January 2010. [6] 3GPP TS 33.401 version 9.2.0 Release 9 – 3GPP System Architecture Evolution (SAE); Security
architecture. ETSI TS 133 401 V9.2.0. January 2010. [7] K. Correll, N. Barendt, M. Branicky - Design Considerations for Software Only Implementations
of the IEEE 1588 Precision Time Protocol – Proceedings of 2006 conference on IEEE 1588 Standard for a Precision Clock Synchronization Protocol for Networked Measurement and Control Systems.
[8] ITU-T G.8261/Y.1361. Timing and synchronization aspects in packet networks. February 2008. [9] ITU-T SG15/Q13, C685 – Huawei Technologies Co. Ltd, China Academy of Telecommunication
Research, MII - Packet Master Traffic Load Analysis based on ITU-T PTP profile message rates – Lannion, December 2009.
[10] Contribution C449 to SG15 by France Telecom, ‘Proposal of transmission of a phase information using OAM messages’, Geneva, May 2007.
[11] ITU-T G.8264 – Distribution of timing packets through packet networks – October 2008. [12] IEEE Std 1588-2008, IEEE Standard for a Precision Clock Synchronization Protocol for
Networked Measurement and Control Systems, July 2008. [13] ITU-T G.8262 - Timing characteristics of synchronous Ethernet equipment slave clock (EEC) –
August 2007. [14] ITU-T G.781 – Synchronous Layer Functions – September 2008. [15] M. Le Pallec, D.T. Bui, F. Dorgeuille and N. Le Sauze – Time and Frequency Distribution Over
Packet Switched Networks. Bell Labs Technical Journal 14(2), 131-154. [16] WD 60 - About the control of PDV using QoS mechanisms and the applicability of proposed PDV
metrics (MAFE and minTDEV), France Télécom contribution to ITU-T, interim meeting Q13/15, San Jose, March 2009
[17] WD 38 - Additional PDV measurements on Ethernet platform, France Télécom contribution to ITU-T, interim meeting Q13/15, Edinburgh, June 2009
[18] WD 39 - PDV measurements on Micro Wave platform, France Télécom contribution to ITU-T, interim meeting Q13/15, Edinburgh, June 2009
[19] C 668 - Performance aspects of a packet slave clock targeting 50ppb and based on a physical clock with a large time constant and PDV accumulation discussions, France Télécom contribution to ITU-T, plenary meeting SG15, Geneva, September 2009
[20] WD 36 - PDV measurements over different platforms – extract from FT presentation at ITSF 2009, France Télécom contribution to ITU-T, interim meeting Q13/15, Lannion, December 2009
[21] WD 37 - Continuation of the considerations on the performance aspects of a packet slave clock targeting 50ppb and based on a physical clock with a large time constant and PDV accumulation discussions, France Télécom contribution to ITU-T, interim meeting Q13/15, Lannion, December 2009
[22] WD 47 - Network PDV accumulation proposal, France Télécom contribution to ITU-T, interim meeting Q13/15, San Jose, March 2010
[23] C 780 - Test plan for PDV generation of a single equipment involved in IEEE1588-2008 transport, Cisco / France Télécom contribution to ITU-T, plenary meeting SG15, Geneva, June 2010
[24] C 1017 - Network PDV accumulation study, France Télécom contribution to ITU-T, plenary meeting SG15, Geneva, June 2010
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 8/76
[25] Study of network Packet Delay Variation for packet based methods, France Télécom article submitted to IEEE Communication Magazine, June 2010
[26] ITU-T SG15/Q13, C685 – Huawei Technologies Co. Ltd, China Academy of Telecommunication Research, MII - Packet Master Traffic Load Analysis based on ITU-T PTP profile message rates – Lannion, December 2009
[27] Taks 2.4 mid-term report on Synchronization service in Carrier-Grade Ethernet environments (M24), Feb 2010.
[28] ITU-T G.8265.1 - IEEE1588™ profile for telecom (frequency delivery without support from network nodes)– June 2010
[29] IETF RFC 5905 – Network Time Protocol Version 4: Protocol and Algorithms Specifications – June 2010.
[30] David L. Mills – Clock Disciplining Algorithm of the Network Time Protocol Version 4 - Technical Report 97-3-3, Electrical Engineering Department of Delaware, March 1997.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 9/76
1 Executive Summary
The migration from TDM to Packet based (IP, MPLS and Ethernet) cellular backhaul networks
has imposed a very difficult challenge for the network designers and operators. The inherit frequency
transparency attribute of TDM was no longer available. New advanced techniques had to be developed
in order to deliver what was already obvious with TDM: A stable, accurate frequency reference to
derive the transmission clock of the base-station. These days, the challenge is further increasing, as
new, more bandwidth efficient cellular technologies (e.g. TD-SCDMA and LTE) depend on the supply
of an accurate Time-of-Day reference signal to the base-station. As a consequence, it is commonly
understood today by the designers and service providers of cellular networks that, along with the
layout of the data plan, careful attention and planning for the synchronization issues must
simultaneously take place.
This document comprises the synchronization related work, carried by the members of Task 2.4
within the TIGERII project over the past 1.5 years. The various contents within this document present
the accomplishments of Task 2.4 based on the initial objectives definition of the Task. The work items
described in this document complement and extend previous work already published in the mid-term
report document M2.4 [27].
At the beginning of the work, Synchronous Ethernet (Sync-E), developed by SG15/Q13 of the
ITU-T in 2008, was identified as the prime technology to distribute accurate frequency information
over packet based network that rely on Ethernet as their link layer protocol (indeed, this is the case for
the majority of green-field cellular backhaul networks). The major appeal, especially for the large
Telecom carriers, is the guaranteed level of performance and the seamless interwork with the legacy
TDM networks. Based on this understanding, three backhaul network deployment scenarios have
been identified:
1. Mobile Operator operating all backhaul segments and Sync-E fully deployed
2. Mobile Operator operating all backhaul segments and Sync-E partially deployed
3. Mobile Operator not operating all backhaul segments (hence, the existence of Sync-E cannot
be guaranteed)
The work plan of Task 2.4 was based upon the above scenarios, trying to answer various open issues
concerning each and every one of them, for the benefit of the industry. The work was also focusing on
problems and issues which are currently not extensively treated by the various SDOs that work on
these topics (e.g., ITU-T SG15/Q13, IETF TICTOC and NTP WGs etc.) with the intention that this
work package could be reused by the SDOs once completed to derive standardized solutions.
The work plan of Task 2.4 was comprised of the following work items that were selected in order
to cover the above mentioned scenarios:
1. Improving ToP end-to-end performance using the following approaches:
• Studying the effect of PDV accumulation on the end-to-end frequency recovery
performance and understanding how different QoS mechanisms could be used to
improve it.
• Studying new mechanisms than can be used to better control end-to-end
performance.
2. New TOD/phase distribution methodologies that rely on network support. Two
main study courses were selected for this item:
• Understanding how interworking with an underlying Sync-E frequency distribution
can be used to improve overall performance in the case where Sync-E is partially
deployed or fully deployed.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 10/76
• Study a new link-by-link phase distribution scheme that is based on the Sync-E ESMC
SSM signalling channel.
Work item 1 is mainly addressing frequency distribution under the assumptions of network
deployment scenario (3), where physical-layer based frequency distribution does not exist. Hence,
frequency distribution must rely on high-layers ToP schemes that are inherently very susceptible to
Packet Delay Variation (PDV). Chapter 6 of M24 [27] embarked on this work, discussing a few
innovative mechanisms that could be used in order to introduce more control into the ToP
environment, aiming towards a more constrained end-to-end error budget. This work is now further
complemented by clause 7.2 of this report, comprising a comprehensive study of the way PDV is being
accumulated in legacy packet networks that does not have special capabilities to handle the ToP
traffic. The work presented in this clause constitutes an important milestone towards understanding
the PDV limits which are required in order to be able to guarantee a certain level of performance and
how standard QoS/CoS mechanisms could be efficiently employed in the network in order to increase
the level of expected performance.
The focus in work item 2 has turned to accurate time and phase1 information distribution in
packet networks, a very ‘hot’ topic that is currently being studied by many SDOs and Academic
institutes. The extremely tight end-to-end time/phase error budget (less than 1 usec in many
applications) has driven the entire industry to the understanding that network support, in one way or
another, is mandatory. This notion was also adopted by the work of task 2.4 as a principal guideline.
The first part of the work is presented in chapter 7 of M24 [27] that suggests a few interworking
concepts between various available technologies (e.g., Sync-E and PTP or PTP and NTP) to better
exploit these technologies and form a more overall optimized solution. This work in now extended
with two additional study items that occupy chapter 8 of this report. The first addresses the challenge
inherit by the use of IPSec tunnels for the synchronization distribution while the second discusses a
TOD distribution scheme that is confined to the physical layer (actually lower part of the link layer)
and which superfluous any higher-layer TOD distribution scheme (e.g. PTPv2), thus, saving valuable
network bandwidth and management resources.
1 Absolute phase synchronization among a group of devices means that a certain periodic event is
occurring in all the devices in the group at exactly the same time. In that scope, knowledge regarding
the specific Time-of-Day (TOD) instants at which the events are occurring is of non-importance. TOD
synchronization implies phase synchronization with the additional requirement that the specific
Time-of-Day (TOD) information related to the occurrence of the events is also distributed.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 11/76
2 Introduction
In the context of a strong development in data services (the majority of applications are now
over IP), especially in the mobile networks, traditional TDM infrastructure is not any more able to
meet the following equation: offering more bandwidth and guaranteeing a viable RPU (Revenue Per
User) (as explained in the deliverable D20). This gives strong arguments to operators to migrate their
transport infrastructure towards Packet Switched Networks (PSNs) with Carrier-Graded Ethernet
(CGE) being one technology of choice. However, many technical challenges are raised for this
migration and the synchronization network requirements represent a major hurdle.
Indeed, during the migration of the transport infrastructure from the TDM towards the PSN,
synchronization is lost, as the traditional Ethernet is not being capable of distributing neither a
frequency reference nor a time/phase reference. Alternatives such as the GPS (Global Positioning
System) raise many technical issues and cost drawbacks (coverage, interference, cost of the numerous
receivers, cabling and engineering costs, radio-frequency jamming, etc) as well as political concerns
(some countries avoid from being GPS dependant).
Within this framework, the ITU-T’s Synchronous Ethernet (Sync-E) is seen as the preferred
technology to distribute a frequency reference and the IEEE‘s 1588V2 (PTPV2) is considered, by
major telecom actors, as the preferred technology to distribute a time or phase reference. Concerning
this latter aspect, the IETF’s Network Time Protocol (NTP) is also to be considered, as per existing
deployments.
In this document, studies will focus on the mobile backhauling as the target application for the
network synchronization over the PSN. Such focus is motivated by the related stringent mobile
synchronization requirements driven by an optimal use of rare radio resources under high speed
mobility constraints (e.g. interferences due to Doppler effects). Through different deployment
scenarios, potential issues will be pointed out and related network requirements will be described.
This report complements mid-term report (M24) submitted in Feb 2010. The work materialized
into M24 was mainly focusing on innovative ways to mitigate Packet Delay Variation, the most
dominant component affecting time/frequency distribution performance in packet networks. Based
on the carrier (FT) input, three typical network deployment scenarios have been introduced,
representing a cellular mobile operator migration process from a case where it does not operate the
backhaul segments (thus having only little control over the synchronization support capabilities in the
network) to a case where it completely owns the backhaul segments, thus having direct control over
the various synchronization support means within the network. The work within M24 has contributed
a few innovative ideas to improve E2E performance for the various deployment scenarios.
The work in this report complements the previous work, emphasizing on the following work
items in particular:
• Improving E2E time/frequency distribution with the use of IPSec tunnels. This work item is
addressing a unique case where classic on-path support schemes like Transparent Clocks
cannot be used dues to the use of the IPSec tunnels.
• Comprehensive study of the way PDV is being accumulated across a frequency distribution
chain in packet networks. This study in extremely important in understanding the achievable
E2E frequency distribution performance given a specific network topology as well as load
profile.
• Accurate distribution of Time of Day (TOD) information using an extension to the standard
Sync-E frequency distribution scheme. Such a TOD distribution apparatus can be especially
appealing in backhaul networks having complete Sync-E distribution capabilities as it will
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 12/76
superfluous any additional higher-layer TOD distribution scheme running on top of the
Sync-E (e.g. PTPv2) and save network resources.
This report is organized in the following way: Chapter 4 presents the key motivation for the work
as is expressed by the strict synchronization requirements of new mobile technologies. Chapter 5
describes the dominant technologies used for time/frequency distribution across packet networks.
Chapter 6 discusses the key network deployment cases as they were identified by France Telecom and
from which the entire work items have been derived. Chapter 7.1 comprises a brief summary on the
accomplishment in M24 addressing networks that do not offer any on-path support (2nd and 3rd
deployment scenarios) while chapter 7.2 comprise the new study on PDV accumulation. Clauses 8.2 to
8.4 discuss the new challenges imposed by the use of IPSec tunnels across the backhaul network,
presenting a few innovative ideas to overcome them. Last, clause 8.6 extends the work already partly
presented in M24, regarding extending Sync-E to distribute TOD, with the new simulation
environment and the obtained results.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 13/76
3 Terminologies
In order to fully understand the content of this document, some terminological definitions are
provided below.
On-Path Support (OPS): On-path support refers to any intended feature of the path over which
synchronization (frequency, phase or time-of-day) is being distributed that may improve the quality of
the distribution.
On-Path Support Node: A node along the communication path between a “Master” and a “Slave”
that offers on-path support for the time/frequency distribution packets.
Syntonization: Frequency synchronization. Two oscillators are syntonized when they have the
“same” frequency within a defined error. The latter is defined as the syntonization accuracy. When
two oscillators are syntonized the difference between their respective phases remains constant with
time.
Phase Synchronization: Absolute phase synchronization among a group of devices means that a
certain periodic event is occurring in all the devices in the group at exactly the same time. In that
scope, knowledge regarding the specific Time-of-Day (TOD) instants at which the events are occurring
is of non-importance
TOD Synchronization: TOD implies phase synchronisation with the additional requirement that
the specific Time-of-Day (TOD) information related to the occurrence of the events is also distributed.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 14/76
4 Mobile network synchronization requirements
Within the mobile network infrastructure, radio interfaces impose the most stringent
synchronization constraints as explained in Chapter 1. Two technology types are to be distinguished:
- FDD (Frequency Division Duplex) technologies only require network syntonization
- TDD (Time Division Duplex) technologies require both network syntonization and
alignment in phase (or phase synchronization).
The mobile network requirements at the radio interface for different mobile technologies are
summarized in the following table:
Radio System Frequency accuracy Time accuracy
GSM 50 ppb NA
UMTS FDD 50 ppb NA
UMTS TDD 50 ppb 2.5 µs (1.5µs*)
CDMA2000 50 ppb 3 µs (10 µs worst case)
TD-SCDMA 50 ppb 3 µs
LTE FDD 50 ppb NA
LTE TDD 50 ppb 3 µs
WIMAX 802.16e 8 ppm 0.5-25µs according to system profile
(*) some standards define max base station to base station time difference and not a difference against a common reference as
here
Table 1 – Mobile networks: synchronization requirements at the air interface level
It is to be noted that radio base stations introduce noises into the synchronization signal. Thus,
requirements are typically more stringent at the network interface level (output of the backhaul
network) than at the air interface level. As an illustration, a 16 ppb value at the network level is often
mentioned while targeting a 50 ppb frequency accuracy at the radio interface level [8].
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 15/76
5 Key protocols and their related dependencies
5.1 Protocol overview
The following sections shortly describe the three main synchronization protocols that will be the
study focus within the TIGER II project.
The ITU-T Synchronous Ethernet protocol is purely dedicated to the distribution of a frequency
reference within an IEEE 802.3 Ethernet-capable network. It is based on a hierarchical
synchronization method using a synchronous physical layer. As specified by the ITU-T within a
migration perspective, Sync-E is aligned with SDH/SONET) in the sense that a Sync-E clock element
can be inserted seamlessly into a SDH/SONET clock hierarchy. This approach should be efficient in
terms of performance as Sync-E is expected to be at the same performance level as SDH/SONET.
The IEEE 1588v2 protocol, also called as the Precision Time Protocol version 2 (PTPv2),
provides a standard protocol for synchronizing clocks within a PSN. PTPv2 was approved as a revision
standard by the IEEE standards Board in March 2008 with the intention to evolve the IEEE 1588v1
from its first deployed local area network (LAN) environment. Many enhancements have been added
for this purpose, including concepts of transparent clock, boundary clock, and unicast support. Some
of these elements will be detailed later in this document. Unlike PTPv1, PTPv2 is specified for
different underlying layers — 802.3/Ethernet, User Datagram Protocol (UDP)/Internet Protocol
version 4 (IPv4), and UDP/Internet Protocol version 6 (IPv6). From this perspective, the protocol is
more suitable for meeting the needs of the telecom industry than its predecessor. The PTPv2 protocol
is designed to distribute a time reference across a communication network between a master clock
and a slave clock.
Growing out of the work conducted by Dr. David Mills at the University of Delaware, the IETF’s
Network Time Protocol (NTP) has taken advantage of lengthy deployment experience with various
standards versions:
- NTPv0 (RFC 958), 1985
- NTPv1 (RFC 1059), 1988
- NTPv2 (RFC 1119), 1989 and
- NTPv3 (RFC1305), 1992.
Today, there are millions of NTP servers and stations deployed worldwide, and NTP software is
implemented on almost every workstation and server platform [1]. The current NTPv4 version is
nearly formalized as a standard IETF RFC and comparatively to the previous version, NTPv4 targets
are:
- To accommodate Internet Protocol version 6 address space,
- To extend potential accuracy to the tens of microseconds for modern workstations and fast LANs by
including fundamental improvements in the mitigation and discipline algorithms,
- To ease NTP deployment by including a dynamic server discovery scheme, so that in many cases
specific server configuration is not required,
- To specify an autonomous authentication protocol,
and
- To correct certain errors in NTPv3 design and implementation and to include an optional extension
mechanism.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 16/76
5.2 Sync-E constraints and performance dependencies
Similarly to SDH/SONET, Sync-E has being designed for being independent from the traffic
load. Thus, the experience with SDH/SONET should be beneficial to Sync-E in that the former has
shown it has been capable of synchronizing very large networks with high accuracy. However, making
use of the physical layer to distribute a reference frequency imposes the Sync-E/SDH chain to be
contiguous and many issues can break this rule:
- A Sync-E equipment in the chain is not 100% compliant to the standard
- Interoperability issue between 2 consecutive Sync-E equipments
- Mis-configuration or mis-installation; e.g. an Ethernet port is peer by mistake to a Sync-E port
- ESMC discrepancy; e.g. Synchronization Status Messages (SSM) corrupted
- etc.
Apart from these technical aspects, regarding the full deployment of hardware support on
Ethernet nodes, cost is a major obstacle which could prevent from having a full migration of the
traditional Ethernet towards Sync-E, thus from having a contiguous chain of physical frequency
distribution.
5.3 Timestamp protocols constraints and performance dependencies
5.3.1 NTP and PTPV2 comparison
Unlike PTPv2, which has been specified for different underlying layers, IETF NTPv4 relies on
UDP/IP (version 4 or 6) and is thus a “pure layer 3” solution in terms of connectivity. All future
developments of NTPv4 have been officially shifted to the IETF Transmitting Timing over IP
Connections and Transfer of Clock (TICTOC) working group. This decision has been driven by the
desire to converge existing time protocols and thus to ensure — as a first step — 1588v2/NTPv4
coexistence within the same network. Similar to PTP, NTP has been conceived for a time distribution
purpose, but can be used for syntonization as well. Server-to-client time distribution methodology
over an NTP-capable network (over IP Packet Network) is conceptually close to the 1588v2
master/slave principle regarding the “2-way signalling” (synchronization message exchanges in both
Master-to-Slave and Slave-to-Master directions).
Thanks to many mechanisms introduced such as the hardware timestamping, the Transparent
Clock, the maximum message rate, etc, PTPV2 is often preferred to NTP for the distribution of an
accurate time (and thus frequency). Nevertheless, as final tasks within the TICTOC Transition, the
NTPV4 working group is specifying a “Carrier-Class” NTP, addressing the hardware time stamping
and maximum message rate requirements. With these additional features, the “Carrier-Class” NTP
should provide similar performance with regards to 1588V2 in an end-to-end scheme. Such a scheme
is presently investigated by the ITU-T SG15/Q13 within the context of the first IEEE1588V2
frequency-only Telecom profile (frequency distribution purpose without neither Boundary Clock nor
Transparent Clock support).
In the rest of this document, IEEE1588V2 will particularly be the investigation focus without
forgetting that the (Carrier-Class) NTP could be a relevant candidate as well for most of the scenarios.
5.3.2 IEEE 1588V2 constraints and performance dependencies
For a clear understanding of the document, the following figure briefly recalls the principle of
the PTPV2 protocol (very similar to NTP):
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 17/76
Figure 1 – PTPV2 principles
PTPV2 performance mainly depends on 2 parameters:
- The delay asymmetry: the difference between the transmission delay of the (Mater/Slave) Sync
message and the related Delay-Req message.
- The Packet Delay Variation (PDV): the difference between the transmission delay of a PTP message
and a reference transmission delay (transmission delay of the first message, theoretical minimum
delay, theoretical maximum delay, etc).
The above dependency which can be seen as the error is computed between the estimated offset
and the ideal offset. The latter offset allows the Slave to adjust its timescale exactly to the Master one:
with dMS being the transmission delay of the Sync message and dSM being the transmission delay of the
associated Delay-Req (Slave/Master). Thus,
The offset estimation error is therefore the instantaneous delay asymmetry:
idealSM
idealMS
offsetdtt
offsetdtt
−−−−++++====
++++++++====
34
12
222
)()( 3142 SMMS
estimated
SMMS
ideal
ddoffset
ddttttoffset
−−−−−−−−====
−−−−−−−−
−−−−−−−−−−−−====
2
SMMS
idealestimated
ddoffsetoffseterror
−−−−====−−−−====
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 18/76
This can be developed further as:
(Equation 1)
with,
Within Equation 1, the first term represents the impact of PDV on PTPV2 performance and
the second term the impact of delay asymmetry. Delay asymmetry, as defined above, while remaining
deterministic can be measured and injected into the Slave computation module in order to reach a
better time accuracy. PDV however has random components which are real issues (it is to be noted
that in the above equations, the random components of delay asymmetry has been arbitrary
integrated to the PDVs. These random components could be also integrated to the delay asymmetry as
another definition of this latter parameter).
From a general perspective, the Packet Delay Variation has different random components
[2][3], namely:
- The transmission Delay (variation): 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 (variation): 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 outgoing packets.
Consequently the traffic load in one direction can have a detrimental effect on the other direction.
- The buffering/queuing delay (variation): 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” 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.
It is to be noted that in a telecom environment, the buffering/queuing delay (variation) is
generally the main component of the PDV.
minMSMSMS ddPDV −−−−====
minSMSMSM ddPDV −−−−====
minmin_ SMMS ddAsymmetryDelay −−−−====
2
_
2
AsymmetryDelayPDVPDVerror SMMS ++++
−−−−====
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 19/76
6 Deployment scenarios
Considering deployment scenarios of the synchronization network, three representative use–
cases will be distinguished within Tiger II synchronization task, thanks to two high level variables:
- Backhaul infrastructure ownership: the Mobile Operator (MO) owns the backhaul
infrastructure or not,
- Sync-E availability within the MO owned infrastructure: Sync-E fully or partially deployed.
6.1 Mobile Operator operating all backhaul segments and Sync-E fully
deployed
The first considered deployment scenario is illustrated by Figure 2 below.
Figure 2 – MO owning all backhaul segments with Sync-E fully deployed
This is an ideal scenario or a targeted architecture where Sync-E is fully deployed for a
syntonization purpose. The Mobile Operator (MO) owns all the backhaul segments and thus has the
flexibility to implement on-path support nodes in order to optimize the protocol performance with
respect to end applications. Both time and frequency references are likely to be co-located and time
distribution paths are likely to follow frequency distribution paths in order to take advantage of the
syntonization supports.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 20/76
6.2 Mobile Operator operating all backhaul segments and Sync-E
partially deployed
This second considered deployment scenario is illustrated by Figure 3 hereafter.
Figure 3 – MO owning all backhaul segments with Sync-E partially deployed
This scenario is more realistic as it considers migration constraints for which Sync-E would be
partially deployed. Similarly to the previous case, the MO takes the great advantage of having control
over the entire backhaul network. In order to distribute the reference frequency towards end
applications, the MO can make use of timestamp protocols (e.g. PTPV2) in order to bound different
physical layer based frequency distribution domains together or to extend one of them for reaching
end applications. Obviously, the MO can also make use of timestamp protocols in order to distribute a
time reference.
Remark: when making use of a timestamp protocol for distributing both frequency and time,
respective time and frequency distribution topologies do not need to be congruent within the entire
network segment while considering their different respective dependencies (PDV dependency for a
syntonization purpose and both PDV and delay asymmetry for a time synchronization purpose).
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 21/76
6.3 Mobile Operator not operating all backhaul segments
The last considered deployment scenario is illustrated by Figure 4 below.
Figure 4 – MO not owning all backhaul segments
This scenario covers numerous cases where the MO does not own all backhaul segments or
does not have any backhaul network at all. Within leased parts of the backhaul network, the MO has
to make use of timestamp protocols in order to distribute both time and frequency. The MO has no
visibility on the potentially available on-path support that can be implemented along the leased
segments. End-to-end approach should be adopted in this case.
6.4 Operator deployment focuses
This chapter presents high level deployment priorities as foreseen/viewed by France Telecom.
The majority of the proposals in this document can be applied to a large number of deployment
scenarios. However, a particular focus will be put on the following choices when it comes to
performance comparison, and especially to simulations and/or test results:
1. For frequency-only distribution:
1.1 Physical-layer based methods (e.g. Sync-E): this is the preferred solution for distributing
frequency. Physical methods are standardized and largely known, thus will not be investigated
within Tiger II research project.
1.2 Packet-based methods:
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 22/76
1.2.1. Packet-based methods without any hardware support (e.g. BC/TC) from the network: it is an
interesting case especially for inter-carrier frequency distribution cases (although physical-layer
based solutions may apply also to this case, e.g. if a Sync-E service is offered as part of the
leased line). Also, this case is relevant for supporting gradual network migration towards
physical layer based technologies as a targeted architecture.
1.2.2. Packet-based methods with partial hardware support (e.g. BC/TC) from the network: regarded
by FT as less important than other approaches. Indeed, if some hardware support is required
from the network for frequency distribution, then Sync-E is seen as the preferred solution.
1.2.3. Packet-based methods with full network hardware support (e.g. BC/TC): regarded by FT as less
important than other approaches. Indeed, if some hardware support is required from the
network for frequency distribution, then Sync-E is seen as the preferred solution.
Summary regarding frequency-only distribution: regarding the deployment of new hardware
support, physical-based methods such as Sync-E are considered as the preferred approach to
guarantee the performance delivered to the end application. In other cases, the use of packet based
methods in an end-to-end approach is preferred.
2. For both frequency and TOD distribution:
2.1 PTPV2 (alternatively NTP) without Sync-E and without network hardware support (e.g.
BC/TC): this is considered by FT to be a challenging case if the purpose is to meet the stringent
LTE phase/time requirement with µs accuracy objective.
2.2 PTPV2 (alternatively NTP) without Sync-E and with partial network hardware support (e.g.
BC/TC): same remark as 2.1. This is considered as a challenging deployment.
2.3 PTPV2 (alternatively NTP) without Sync-E and with full network hardware support (e.g.
BC/TC): this is identified as possibly an interesting case where prolonged holdover is not an
issue.
2.4 PTPV2 (alternatively NTP) with Sync-E and without network hardware support (e.g. BC/TC):
same remark as 2.1. This is considered as a challenging deployment.
2.5 PTPV2 (alternatively NTP) with Sync-E and with partial network hardware support (e.g.
BC/TC): same remark as 2.1. This is considered as a challenging deployment.
2.6 PTPV2 (alternatively NTP) with Sync-E and with full network hardware support (e.g. BC/TC):
this is identified by FT as probably the best technical approach that would certainly allow to
guarantee both required TOD and frequency performance under normal operation conditions
as well as holdover.
2.7 Link-by-link TOD distribution using ESMC messages: this case is identified to be an interesting
solution for networks where Sync-E is already deployed as it will save the need to use PTP in
addition.
Summary regarding frequency and TOD distribution: according to France Telecom, physical
layer frequency distribution should be used to support time distribution (to guarantee time holdover
performance). A fully deployed hardware support is also seen as a targeted architecture in order to
meet stringent LTE phase/time requirement with µs time accuracy objective. TOD distribution can
either be accomplished end-to-end using PTPV2 (or NTP) with full network hardware support or link-
by-link using the ESMC messages (if Sync-E is already deployed).
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 23/76
7 Improving Timing-over-Packet (ToP)
7.1 Summary of work accomplished in M24
The work presented in Chapter 6 of [27] was focusing on ways, both from the time/frequency
distribution master, time/frequency distribution slave and the network element in the middle, to
mitigate and dampen the effect of Packet Delay Variation (PDV) on the overall time/frequency
distribution performance.
In particular, clause 6.1 discussed the dampening of network introduced PDV by applying
special means at the time/frequency distribution master, slave and the network elements in between.
At the master side, controlling the packet-rate and the packets transmission profile of a
time/frequency distribution master device in order to dampen end-to-end network induced PDV, was
discussed. A PDV decrease in one order of magnitude has been argued as accountable by such control.
At the network element, the benefit of accurate timestamping based on dedicated fast logic
hardware was presented. The performance degradation expected in last-mile technologies that do not
have such mechanism was presented.
Different network QoS mechanisms have also been mentioned as sometimes an effective way to
mitigate network PDV. In particular, by differentiating timing signalization flows from other packet
flows using QoS-build-in information within the transport header. This allows for prioritizing
synchronization packets with regards to data packets, and avoiding a direct competition of the latter
against the formers. QoS technique usually relies on parallel buffers whose filling depends on
scheduling policies. It is mentioned that, though being sometimes highly efficient for PDV mitigation,
such QoS techniques cannot entirely make the problem go away. A following in-depth discussion on
this issue is given in Chapter 7.2 of this report.
At the slave side, different filtering mechanisms (a.k.a. ranging mechanisms) that can be
applied in order to enhance the PDV filtering capabilities of slave were presented. In particular, two
traditional mechanisms, the Moving Average (MA) and the Exponentially Weighted Moving Average
(EWMA) mechanisms were discussed presenting their suitability for some PDV profiles as well as
their incompetence to deal with different PDV profiles. A new promising approach focusing primarily
on those packets that experience minimum network delay (sometimes referred to as ‘early packets’)
was also discussed along with a set of new related metrics, such as MinTDEV and MATIE that aim to
better characterize the relationship between a specific PDV profile and obtainable time/frequency
recovery performance.
Clause 6.2 of [27] extended clause 6.1 by suggesting a new Distributed Ranging Mechanism that
offers more PDV control compared to the legacy ranging mechanism that is used by current
time/frequency distribution protocols as was described in clause 5.3.
The new scheme, presented in Figure 5 below, suggests the use of special distributed
“selection/filtering” modules along the time/frequency network distribution path that have the ability
to recognize timing (e.g. 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 (e.g. 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 timing packet emission rate. In the opposite way, a second threshold can
be provisioned so that the filter module can request the Master to decrease the timing packet emission
rate when the number of selected packets is found to be above this second threshold. This method
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 24/76
guarantees that the Slave receives sufficient relevant packets during its filtering process in order
discipline its local oscillator.
An illustration of the benefits of the apparatus was also given using two representative PDV
examples based on different Probability Density Functions (PDF). Avoiding network congestion that
might be caused by the intended increase of synchronization traffic was discussed.
Figure 5 – Distributed ranging mechanism
Clause 6.4 of [27] discussed ToP as an extension for physical-layer frequency distribution. A
new approach to inter-connect remote isolated Sync-E islands using ToP in general and PTPv2 in
particular was presented. A new concept termed an ‘SSM tunnel’, that allow the PTPv2 protocol to
distribute native Sync-E SSM Quality Level (QL) information between the Sync-E capable islands was
presented (see Figure 6 below). It was argued that this solution will enable the network operators a
smoother migration towards Sync-E by bridging those network parts that still did not migrate towards
the new physical-layer frequency distribution technology.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 25/76
Figure 6 – SSM Tunnel across ToP domain
It should be mentioned that since the completion of [27], the ITU-T SG15/Q13 group of experts
has defined, within the PTPv2 Telecom Profile suit for frequency distribution [28], the distribution of
native Sync-E QL information using the clockClass field in the PTPv2 Announce message. The
motivation for doing that was to form frequency distribution continuity between the native Sync-E
technology (based on physical layer timing distribution) and packet based PTPv2.
7.2 Study of Packet Delay Variation (PDV) accumulation over packet
networks, for the purpose of delivering frequency synchronization
As already discussed in Chapter 5, Synchronous Ethernet, a technology standardized in ITU-T,
specifies how to synchronize Ethernet PHY in order to achieve a physical layer timing distribution.
However, some already deployed networks do not implement the relevant hardware to support
Synchronous Ethernet. Alternative methods may therefore be required in these cases, at least
temporarily.
Packet based methods, such as Precision Time Protocol (IEEE1588/PTPv2), form a family of
methods enabling to deliver frequency synchronization over a packet based network with an
asynchronous physical layer. However, as presented in Chapter 5, their end-to-end performance is
greatly affected by the PDV introduced by the packet network. It is therefore crucial for an operator to
have access to engineering rules permitting to deploy such methods in a safe way, in order to ensure
that the relevant synchronization quality will be delivered to the network equipments which need such
reference.
7.2.1 The need for developing engineering rules for packet based methods
In the scope of this document, packet based methods refers to mechanisms enabling to
transport frequency synchronization via packets over a network which handles these packets as data
traffic and does not offer any specific timing hardware support in the intermediate transport nodes.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 26/76
Those timing packets may contain timestamps, as it is for instance the case when a protocol like PTP
is used. Other types of synchronization, like phase/time, are not discussed in the scope of this section,
although some observations it contains might also be relevant in this context.
Deploying this kind of solution is not trivial and requires careful analysis of the network: the reception
of the timing packets is indeed not enough to ensure that the timing information will be recovered
with the appropriate quality.
The key parameter impacting the performance of packet based methods when there is no timing
support from the network nodes is the PDV produced by the network. Indeed, the timing packets are
not all experiencing exactly the same transit delay when transported over a network; these variations
of delays, for instance due to packet buffering, can lead to degrade the timing quality recovered by a
packet slave clock.
The reasons for the difficulties to develop engineering rules for packet based methods
deployment are mainly due to two aspects: the first one is related to the lack of knowledge on the ways
to control the PDV produced by a telecom network, and the second one is related to the fact that the
implementations of packet slave clocks are using proprietary PDV filtering algorithms.
As a consequence, engineering rules for the deployment of packet based methods are not
currently defined. Their development is related to the characterization and the control of the PDV, a
subject which is still heavily debated in standardization organisms. In particular, the maximum
number of network nodes between a packet master clock and a packet slave clock is difficult to
determine, as it depends on many parameters.
Therefore, it is essential to study the typical PDV that is seen over real telecom networks, but
also to study the mechanisms enabling to control and limit this PDV. A better understanding of the
packet slave clocks PDV tolerance and performance is also a necessary step before a safe development
of packet based methods, such as IEEE1588/PTPv2.
In addition, a methodology for building a network which meets the PDV tolerance of the
filtering algorithms implemented in the packet slave clocks is required in order to ease operational
aspects related to these technologies for the operators.
The next sub-Clauses will present some studies related to PDV aspects, with this objective of
progressing the engineering rules for deploying packet based methods. In particular, a network PDV
accumulation model will be introduced.
7.2.2 Analysis of PDV generated by telecom networks
PDV generated by telecom networks is a complex parameter to characterize. Indeed, each
network equipment is different in terms of PDV generation, and the parameters having impacts on
PDV generation are multiple and complex to control. Moreover, how the PDV accumulates over the
network still need valid models.
References [16][17][18][20][24] contain examples of PDV measurements performed over
different telecom network equipments. These results include measurements over platforms of
different technologies (fiber, micro-wave...) and different manufacturers. The main observations from
these documents will be summarized hereafter.
The first important observation from those measurements is that two equipments of the same
technology, even from the same manufacturer, can generate very different PDV. The consequence is
that testing on a case-by-case basis is necessary in order to have a precise knowledge of the PDV noise
produced by a network.
Among the parameters which may have impacts on PDV, the following can be listed:
1. Type of medium (e.g. micro-wave, copper, fiber link)
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 27/76
2. Hardware architecture of the equipment (small routers for instance do not necessarily implement the same hardware architecture as big core routers)
3. Features activated in the equipment (in general the more features activated, the higher the risk of PDV generation, because of shared resources)
4. Speed of the links carrying the timing packets 5. Transport technology (e.g. Optical Transport Network (OTN), Digital Subscriber Line (DSL),
packet network, …) 6. Characteristics of the background traffic load (e.g. level of load, size of the packets, …) 7. Prioritization of timing flows and background data traffic flows
A general idea often mentioned is that prioritizing the timing packets enables to solve the
problem of PDV generation. However, even if the way the packet timing flow is prioritized in the
network equipments has a strong impact on the PDV, it does not necessarily always fully solve the
problem.
Indeed, Quality of Service and Class of Service (QoS/CoS) mechanisms are generally designed
for improving the packet jitter of real time applications (such as VoIP for instance). The packet jitter
requirements of such applications are totally different from the PDV requirements of "timing
applications" (PDV in the order of µs can have an impact for timing applications, while only
amplitudes of PDV in the order of ms are generally of importance for real time applications).
Therefore, these mechanisms are not fit for purpose for timing flows.
For instance, QoS/CoS mechanisms may in some case improve the overall PDV amplitude, but
not necessarily stabilize the population of timing around a fixed delay window. As it will be discussed
further in section 7.2.3, the statistical distribution of the delays of the timing packets is in general a
more important parameter compared to the amplitude of the PDV.
One aspect that is generally observed however when assigning the highest (strict) priority to the
timing packets is that it helps improving the population of packets close to the minimum delay transit
of the network (also called floor delay in this document). It is an important observation which will be
reused in the PDV accumulation process presented later. However, certain technologies produce PDV
independently of the traffic load (e.g. DSL, micro-wave), and prioritization does not really help
increasing the population of packets close to the floor delay.
The following general observations have been made when measuring PDV over different types
of Ethernet platforms using fiber connections:
1. Delay steps, corresponding to an increase or a decrease of the minimum delay transit of the network, can be generated by some equipment, especially in case of congestion or high level of load, but not only (some equipments are generating delay steps even with reasonably low level of load).
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 28/76
Figure 7 - Load pattern and resulting delay steps generated during congestion periods (figures
extracted from [16])
0
10
20
30
40
50
60
70
80
90
100
0 2 4 6 8 10 12 14 16 18 20 22Time (h)
% trafic load
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 29/76
Figure 8 - Delay steps generated with reasonably low levels of load (figure extracted from [20] and
[24])
2. Intermediate priority assigned to the timing packets does not help in general to improve the PDV, and can sometimes degrade it compared to transporting the timing packet as best effort traffic. Only the highest (strict) priority significantly helps in general.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 30/76
Figure 9 - Comparison of PDV histograms with lowest priority (first plot - peak-to-peak = 302µs), intermediate priority (second plot -peak-to-peak = 161µs), and highest priority (last plot - peak-to-
peak = 43µs) for the timing packets in case of congestion (figures extracted from [16])
3. Packet size of the background data traffic has an important impact on PDV. Large packets are generally creating more PDV, since they are longer to be transmitted over the physical link.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 31/76
Figure 10 - PDV measurement when changing the size of the background traffic (figure extracted
from [16])
4. Core equipment generates very different PDV compared to access equipment. Different
hardware architectures of the equipments may explain this observation.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 32/76
Figure 11 - Comparison of PDV histograms generated by a core equipment (first plot - peak-to-peak = 302µs) and by an access equipment (second plot - peak-to-peak = 1.55ms) when timing
packets are not prioritized in case of congestion (figures extracted from [16])
5. Similar types of equipment of the same technology from different manufacturers lead to different PDV generation. Again, different hardware architectures of the equipments may explain this observation.
Other observations can be made when measuring PDV over other types of transport
technologies. For instance, OTN transport does not produce important PDV for frequency
distribution. Micro-wave systems however can create some problems (see reference [18]): in
particular, strong impacts have been observed when adaptive radio modulation feature is used,
enabling to change the radio modulation according to the weather conditions (in general, delay steps
are observed when the radio modulation changes).
7.2.3 Model of a packet slave clock
The required timing quality to be delivered by a packet slave clock depends on the end
application connected to the slave. In some cases, a packet slave clock needs to control the phase error
accumulation so that Maximum Time Interval Error (MTIE) and Time Deviation (TDEV) masks
would not be exceeded. In other cases, like for instance the case of a packet slave clock integrated in a
mobile base station, only the frequency error needs to be bounded, not the phase error accumulation.
A well-known example corresponding to this latter case is the 50ppb accuracy requirement for
the air interface of mobile base stations, which can be considered as a relaxed requirement compared
to meeting MTIE and TDEV masks. Moreover, a good oscillator needs anyway to be implemented in
the base stations, in order to meet the short term stability requirement for the radio interface.
Therefore, in the mobile network context, the integration of the packet slave clock in the base station
permits to relax the challenge related to performance. This work will therefore mainly focus on the
50ppb requirement, corresponding to this case.
As explained earlier, in order to recover the timing information, a packet slave clock receives
timing packets. Not all the timing packets received by the slave are used in general for timing
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 33/76
recovery, since the overall PDV noise would be too important for the packet slave clock, compared to
the target performance.
Instead, in almost all the packet slave clocks, some packet pre-selection is done. This is for
instance the way a packet slave clock is modelled in Figure 12, extracted from draft ITU-T
Recommendation G.8263 under discussion in ITU-T Q13/15. In this figure, a Packet Selection block is
shown before the Phase Lock Loop (PLL). Therefore, only a subset of the timing packets is applied to
the PLL. In general, the packets are selected according to their transit delay.
Low Pass filter Oscillator
Local Time
scale
Output ClockPacket Timing
SignalPacket
Selection
Local reference
Time Scale
Comparator
PEC-S
Figure 12 - Figure A.1/G.8263 draft - Function model of a PEC-S packet-based clock (source of the data: International Telecommunication Union, ITU-T Working Party 3/15 Temporary
Document 318, G.8263 latest draft, 16 May 2010)
Deriving general rules regarding the behaviour of a packet slave clock is a difficult task, because
of the proprietary nature of PDV filtering algorithms. Some assumptions are however necessary in
order to enable deriving suitable PDV criteria to be met by a network: when testing different packet
slave clock implementations, some common observations can generally be made.
For instance, a consequence of the packet pre-selection in a packet slave clock is that the
amplitude of PDV is in general not an important parameter. Instead of using all the timing packets, a
delay window where a sufficient population of packets is located may be considered by the packet
slave clocks for timing recovery. Only the packets located in this delay window may for instance be
selected and used for timing recovery by the slave in some cases.
This delay window may be located at the floor delay (i.e. the fastest packets are selected), or at
another delay window (for instance, in case of poor population of packets close to the floor delay). A
pre-selection based on the floor delay criteria is implemented in most of the packet slave clocks (some
slaves implement other criteria in addition).
Therefore, being able to build a network enabling to always offer for any condition (e.g.
independently of the level of load) a certain population of packets in a given fixed delay window is
expected to strongly help the performance of the packet slave clocks. It is the purpose of the
methodology presented in the next section.
7.2.4 PDV accumulation study
This section will introduce a network PDV accumulation model, based on statistical analysis.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 34/76
7.2.4.1 Objective and methodology
This section presents the objectives of the network PDV accumulation model. The main idea is
to define a methodology for ensuring that a network will always offer a minimum population of timing
packets located in a floor delay window of a given size, whatever are the conditions of the network
(e.g. for any load condition). More details about this methodology can be found in the references
[19][21][22].
Figure 13 shows a 60µs floor delay window; in this case, the objective is that enough timing
packets would be located in this delay window starting from the delay dmin of the fastest packet to
dmin+60µs.
Figure 13 - Illustration of a 60µs floor delay window
The key aspect leading to choose the floor delay window as the relevant window in the model
discussed here comes from the observations made when testing Ethernet platforms: prioritizing the
packet timing flow (e.g. PTPv2 flow) generally enables to reduce the amplitude of the delay steps
which occur when the network load increases, and therefore to optimize the concentration of packets
close to the floor delay during load variations.
Other models might be imagined with delay windows not located at the floor, for instance in
order to cover the technologies which natively do not experience a good population of packets at the
floor, independently of the traffic load (e.g. DSL links).
The key idea of this model is illustrated in Figure 14: at the end of a reference chain, enough
population of timing packets must remain within the floor delay window (e.g. 5% of the packets in the
100µs floor delay window) independently of the traffic load applied on the network. Each node
removes a certain percentage of the packets from this floor delay window, because of the PDV it
produces.
60µs floor
delay
window
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 35/76
Figure 14 - Network PDV accumulation model
The values provided in Figure 14 are taken as examples; it is possible to use other values than
100µs for the floor delay window size, and than the 5% objective. The appropriate values to be used
depends on the specific packet slave clock implementation, for instance on the quality of the oscillator.
In order to determine how the packets are removed from the floor delay window when
cascading nodes, a statistical analysis of the PDV produced by each node can be performed, using
convolution operations in order to anticipate the remaining population of packets in a given floor
delay window. This will be further explained in section 5.2.
This approach targets packet slave clocks embedded in base stations, where only 50ppb
accuracy is targeted. For packet slave clocks targeting bounded phase accumulation error, some more
complex approach could be needed.
7.2.4.2 Experimentation using statistical analysis and convolution operator
In order to study how the population of timing packets decreases from a floor delay window
when cascading nodes, a statistical analysis is used here, based on real PDV generation
measurements. All the details of this experimentation can be found in the reference [24].
Packet master clock
(e.g. PTPv2 master)
Packet slave clock
(e.g. PTPv2 slave)
Percentage of timing packets in the 100µs
floor delay window (worst case)
100%
N1%
N2%
N3%
N4%
5%
objective
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 36/76
The approach presented here consists in measuring the PDV produced by an equipment during
stress conditions (for instance, when loading the equipment), and to use these PDV measurements to
anticipate the PDV produced by several equipments of the same type when they are cascaded. Mixing
different types of equipments should also be possible with the same approach.
Figure 15 shows an example of the PDV produced by a single network equipment on PTP timing
packets when data traffic load of variable packet size is applied on the same output port as for the PTP
packets. The first diagram shows how the PDV generated by the network equipment is evolving when
increasing the traffic load by steps. The second diagram shows the histograms corresponding to each
step of load considered independently.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 37/76
1
10
100
1000
10000
100000
0.000E+00 1.000E-05 2.000E-05 3.000E-05 4.000E-05
Histogram10% 20% 30% 40% 50% 60% 70% 80% 90% 91%
92% 93% 94% 95% 96% 97% 98% 99% 100%
Figure 15 - PDV of PTP Sync messages and corresponding histograms
Each histogram is calculated during a stationary condition (stationary load) and represents
statistically the delay that can be experienced by a timing packet when going through the node loaded
at a given level. Figure 15 shows that when increasing the traffic load, a timing packet has higher
probability to experience higher delay.
The histograms measured over one node can then be used to determine what should be the
histograms after several nodes. Indeed, assuming that the PDV noise produced by a node is not
correlated to the noise produced by the next node (e.g. in case they are loaded with uncorrelated
traffic), it is possible to use convolution operations to anticipate the histograms produced after several
nodes. Convolution can be seen as the "addition" of histograms.
Experimentation has been performed in order to verify that this theoretical assumption is
correct. In this experimentation, two similar nodes (identical to the one used for the PDV shown in
Figure 10) have been cascaded and loaded similarly at the same instant, as illustrated in Figure 16.
The PDV is measured on PTP timing packets directly at the output of the second node.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 38/76
Figure 16 - Configuration with 2 nodes used during the experimentation
This configuration aims at measuring how the PDV accumulates when 2 nodes are loaded at the
same level at the same time, in case of a "hop-by-hop" data traffic (each data traffic flow T1 and T2
exists the network at the next node it enters).
The histograms of the PDV measured with this configuration have been analyzed and compared
with the theoretical results obtained when convolving with itself each initial histogram obtained with
a single node (e.g. histograms of Figure 15).
Figure 17 provides an example of this comparison in the case when the two nodes are loaded at
80%, with background data traffic of variable packet size. The first diagram compares directly the
histograms: the red curve corresponds to the initial histogram measured over a single node, the blue
curve corresponds to the convolution of the red histogram with itself, and the orange curve
corresponds to the histogram measured after having cascaded two nodes. The second diagram
compares the cumulative distribution functions (CDF) of the histograms. It shows how the population
of timing packets is evolving when increasing the size of the floor delay window. It can be observed
that the theoretical convolution calculation matches quite well the real measurement with 2 nodes in
this case.
0.000E+00 1.000E-05 2.000E-05 3.000E-05 4.000E-05 5.000E-05
Relative delay (s)
Packet Number
initial hist 1 node hist conv 1 node hist 2 nodes
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 39/76
0%
20%
40%
60%
80%
100%
0.00E+00 5.00E-06 1.00E-05 1.50E-05 2.00E-05 2.50E-05 3.00E-05 3.50E-05 4.00E-05
Delay Window Observation (s)
Pe
rce
nta
ge initial CDF 1 node
CDF conv 1 node
CDF 2 nodes
Figure 17 - Histograms and Cumulative Distribution Functions comparison at 80% of traffic load
Based on this experimentation, the main following points can be summarized:
- Convolution operations appear in most of the cases predicting correctly the PDV generated by 2 nodes in the configuration used (two uncorrelated data traffic with a "hop-by-hop" scheme), and seems therefore a priori a useful tool in this case. It takes into account the effect of delay steps accumulation
- In general, the predictions based on convolution operation are slightly optimistic. This
behaviour might be due to the fact that not all the components of PDV noise have been taken into account in the convolved histograms. In some other (rare) cases, the prediction with the convolution is pessimistic. An appropriate worst case histogram reflecting how the main sources of PDV in the network equipment can accumulate needs therefore to be considered.
- Identification of all the sources of PDV noise in the network equipments is critical in order to properly anticipate the PDV produced by a network. Not only the effects of the output ports need to be considered.
- The way several data traffic flows are generated seems critical when measuring the effect of cascading nodes; in case the two data traffic flows T1 and T2 applied are correlated, then the convolution operator does not provide correct estimation of the accumulated PDV.
- It is expected that all these observations can be extended to more than 2 nodes being cascaded, and to the case of nodes with different levels of load at a given instant. This has however not been confirmed at the moment with real experimentations.
- Additional testing needs to be performed to study how the PDV accumulates in the case of an "end-to-end" data traffic (i.e. the data traffic flows do not exist the network at the next node they entered). It is likely that in this case, other phenomena may be observed (e.g. delay steps), and that other tools in addition to the convolution operation might be needed to estimate how PDV accumulates.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 40/76
7.2.4.3 Theoretical maximum number of nodes which can be cascaded
Based on the observations of section 7.2.4.2, this section will present a theoretical calculation of
the maximum number of nodes which can be cascaded in case of uncorrelated hop-by-hop data traffic.
In order to compute this calculation, a given criteria in terms of population of packet in a floor delay
window has to be chosen. In addition, a theoretical worst case of PDV histogram needs to be identified
for different levels of load, reflecting the worst case of depopulation of packets in the floor delay
window. This latter point is not a trivial task, and some assumptions may have to be made, for
instance based on the PDV generated by the equipment of interest during different conditions. An
initial proposal of PDV testing procedure can be found in the reference [23].
In the computations presented in this section, a minimum of 5% of timing packets to be
maintained within a 100µs floor delay window will be used as the relevant criteria to be met. A
theoretical worst case of PDV histogram has been determined for 3 levels of load: 50%, 80%, and
100% (congestion). These histograms are then convolved together as needed in order to determine the
theoretical maximum number of nodes which can be cascaded, while meeting the previous criteria.
Figure 18 shows how PDV histograms and corresponding CDFs are evolving when convolving
the theoretical worst case histogram for 50% of load. It also shows how the population of timing
packets in a 100µs floor delay window is evolving.
0.00E+00 5.00E-05 1.00E-04 1.50E-04 2.00E-04 2.50E-04 3.00E-04
Relative delay (s)
Pa
cke
t N
um
be
r
Initial 1 Conv 2 Conv 3 Conv 4 Conv 5 Conv 6 Conv 7 Conv 8 Conv
9 Conv 10 Conv 11 Conv 12 Conv 13 Conv 14 Conv 15 Conv 16 Conv
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 41/76
0%
20%
40%
60%
80%
100%
0.00E+00 2.00E-05 4.00E-05 6.00E-05 8.00E-05 1.00E-04 1.20E-04 1.40E-04 1.60E-04 1.80E-04 2.00E-04
Delay Window Observation(s)
Pe
rce
nta
ge
Initial
1 conv
2 conv
3 conv
4 conv
5 conv
6 conv
7 conv
8 conv
9 conv
10 conv
11 conv
12 conv
13 conv
14 conv
15 conv
16 conv
100µs limit
5% limit
Figure 18 - Analysis of successive convolutions
The first diagram of Figure 18 shows that when increasing the number of nodes, the histograms
become logically more and more Gaussian, and their amplitude increases. A timing packet has
therefore higher probability to experience higher delay. The second diagram of Figure 18 shows the
CDFs of each histogram. A black dotted line has been drowned at 100µs. A red dotted line shows the
5% limit. The last diagram of Figure 18 shows how is evolving the percentage of packets within a
100µs floor delay window when simulating cascaded nodes via convolution operation. Up to 7
convolutions, corresponding to 8 cascaded nodes loaded at 50%, 100% of the timing packets are
within the 100µs floor delay window. Beyond 8 nodes loaded at 50%, the percentage is decreasing.
The maximum number of nodes loaded at 50% which can be cascaded while meeting the criteria of
minimum 5% of timing packets within a 100µs floor delay window is 16 nodes, corresponding to 15
convolutions (8.06% of packets are remaining in the 100µs floor delay window, while after 17 nodes,
only 3.46% are remaining).
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 42/76
In other words, this model shows that when more than 16 nodes loaded at 50% or less are
cascaded, there is a risk that less than 5% of timing packets would remain in a 100µs floor delay
window, because of the traffic load variations.
Similar calculation might be done with other levels of load, for instance (provided that the
criteria to be met are the same):
- 16 nodes loaded at 50% can be cascaded (8.06% remaining in a 100µs floor delay window), as shown in Figure 18
- 10 nodes loaded at 80% can be cascaded (16.16% remaining in a 100µs floor delay window) - 4 nodes at loaded at 80% + 11 nodes loaded at 50% can be cascaded (7.71% remaining in a
100µs floor delay window) - 1 node in congestion (loaded at 100%) + 4 nodes loaded at 80% + 8 nodes loaded at 50% can
be cascaded (6.06% remaining in a 100µs floor delay window)
It is important to remind that those theoretical estimations are only valid for the specific
equipment which has been tested, and for which worst case histograms have been derived. Other
equipments might produce more or less PDV, and therefore the maximum number of nodes that can
be cascaded can be different. Note that in order to simplify this point, several categories of
equipments in terms of PDV generation levels might need to be defined. Anyway, this methodology,
which needs to further analysed and completed, provides an initial model for building engineering
rules for the deployment of packet based methods.
7.2.5 Conclusion (for clause 7.2)
Clause 7.2 has presented the context and the challenges associated to packet based methods,
such as PTPv2, for frequency distribution. It has also discussed general observations which can be
made with regards to the PDV generated by telecom networks, as well as how a packet slave clock can
be modelled.
A network PDV accumulation model has been explored based on statistical analysis, with the
objective of progressing the engineering rules for deploying packet based methods. Some initial
experimentation seems to confirm that this model is valid in the case of hop-by-hop uncorrelated data
traffic. However, further analysis and models might be required in other contexts, for instance in the
case of end-to-end data traffic.
Although controlling the PDV produced by a network is a difficult task, these results show that it
might be possible to do so if the purpose is to ensure that a minimum population of packets is located
in a given fixed delay window.
Testing on a case-by-case basis of the PDV produced by the network equipments is however a
prerequisite, as it has been observed that different equipments generate different levels of PDV. It is
likely that this testing would be a task for the equipment manufacturers, who have the knowledge of
the hardware architecture of their equipments. Defining several categories of PDV generation levels
might however help simplifying this aspect.
Clarification of the input tolerance of a packet slave clock in terms of PDV is also a necessary
step, in order to specify clearly which criteria the network has to meet. It is clear that this input
tolerance specification implies that all the packet slave clocks would share at least a minimum
common criteria (e.g. utilization of the packets located at the floor delay, if the population is
sufficient), in addition to other criteria which might be implemented in smarter slaves.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 43/76
8 Time and Phase Distribution over Packet Networks
with Full or Partial Network Support
8.1 Summary of work accomplished in M24
Chapter 7 in [27] addresses time (and frequency) distribution in networks having full or partial
on-path-support.
The discussion in Chapter 7 begins with an overview of two of the most common representatives
for PTPv2 on-path support, namely, the Transparent Clocks (TCs) and the Boundary Clocks (BCs).
The differences in-terms of the distribution topology as well as the expected E2E performance are
discussed. Following, deployment considerations for the cases of full and partial network support are
presented. For the partial network support case, a discussion regarding the slave clock dominant error
contributors is given focusing on the packet filtering capabilities and the impact of the local oscillator.
Next a comprehensive comparison between BCs and TCs is given.
In clause 7.2 another form of on-path support is being discussed. An inter-work between a
PTPv2 time distribution scheme and an underlying Sync-E frequency distribution scheme is
presented. It is argued that awareness at the high-layer PTPv2 distribution node to the underlying
Sync-E physical layer frequency distribution, can be nicely exploited to improve overall time
distribution performance. Specifically it is shown that a collaboration scheme between the Sync-E and
PTPv2 management entities could be used to find the best network routes for the PTPv2 protocol that
will have the most interactions with Sync-E capable nodes (see Figure 19 below). The high quality
frequency reference, supplied by the Sync-E in such nodes, could be then used by the PTPv2 on-path
support function (e.g. TC or BC) to improve its performance, contributing to the E2E time
distribution. Sync-E failure events and thus a change in the Sync-E distribution topology, in such a
case, could trigger a PTPv2 distribution route change in order to re-adapt to the new Sync-E
distribution topology.
••
•1588V2 slave
••
••
••
••
•Grandmaster 1588V2 •SyncE
•PRC
•SEC
•G.811
•G.811
•G.811
•SSM
•����
•����
•����
•����•����
•����
•����
•����
•G.811
•Transparent nodes
•Syntonized
•Boundary clock
•Other 1588V2
•slave
•1588V2 BC
•Non •-•syntonized
•Boundary clock
Relevant Synchronization
path
•����
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 44/76
Figure 19 - Automatic computation of the PTPV2 topology thanks to an improved BMC taking into
account a local Sync-E support
Last, on clause 7.4, presents a new approach to the collaboration between two ToP schemes: the
PTPv2 and the NTP. The cooperation suggested is driven by the fact that although a solution based on
PTPv2, that is the newer time distribution scheme, usually comprises important features like HW
timestamping and on-path support (e.g. TCs) that a typical NTP solution usually lacks, NTP is much
more widely spread and has been massively deployed over the past 30 years to facilitate a very wide
range of applications. Thus, if the old NTP scheme could somehow benefit from the goodies offered by
the new PTPv2, major performance enhancement could be achieved.
The novel idea is to use the 1588V2 signaling for carrying the NTP one. This latter could be
performed by taking advantage of the flexible 1588V2 structure through a Type Length Value (TLV)
field and by ensuring a protocol inter-working at the intersection of time domains, as depicted in
Figure 20.
The proposed solution is based on 2 complementary building blocks:
- The first one relies on NTP ADD/DROP elements bordering 1588V2 transparent clocks domains
guaranteeing the coherence and continuity between NTP and 1588V2 (particularly for timestamps
and correction fields).
- The second building block is an auto-discovery/advertising mechanism dedicated to inform a
management or a control plane of the position of these ADD/DROP elements to ease the (dynamic)
establishment of the best paths between a NTP server and a NTP client.
NTP ADD/DROP element clocks have local requirements for the correction of (queuing) times
and thus should be far cheaper than ’gateway clocks’ (which stability has to be ensured during the
two-way signaling). This enables a dense deployment of these formers, yielding to a high flexibility in
terms of deployment (especially for sharing expensive NTP servers and 1588V2 masters), providing an
optimal allocation of all synchronization resources.
Figure 20 – NTP ADD/DROP concept
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 45/76
8.2 LTE IPSec challenging environment
8.2.1 Introduction
As discussed in Chapter 5, the IEEE Standard 1588TM-2008 (called hereafter as IEEE 1588V2)
suffers from Packet Delay Variation (PDV) and packet Delay Asymmetry (DA) induced by the Packet
Switched Network (PSN).
In order to maintain both of the aforementioned adverse effects within controlled boundaries
and to reach high performance in term of synchronization accuracy, the IEEE 1588V2 standard has
specified Boundary Clock (BC) and Transparent Clock (TC) concepts. Those two mechanisms are
called as On-Path Support (OPS) as referred to the terminology already defined in [27].
The respective strengths and weaknesses of BCs and TCs have also been discussed with [27].
The conclusion drawn from this discussion is that they have complementary roles and that they are
welcome for addressing network dependencies within a high-performance IEEE 1588V2 deployment.
Nevertheless, there are still several scenarios where their deployment becomes very complex or
even not applicable (e.g. TC in the middle of an IPSec tunnel). This Chapter illustrates such scenarios,
via an important study case represented by the LTE standard. Addressing security concerns, this latter
recommends the use of IPSec tunneling which strongly complicates the deployment of any BC and/or
TC.
In such a context, this Chapter proposes two innovative OPSs which better suit ciphering
environments. The first one addresses the Delay Asymmetry (DA) concern whereas the second one
addresses PDV effects. Both solutions present an advantage of being protocol-agnostic, meaning that
they could be reused to improve the performance of other packet time distribution protocols including
the well established NTP (Network Time Protocol).
8.2.2 LTE IPSec environment description
The LTE standard gives way to a full deployment of PSN for its backhaul, by adopting a flat IP
transport architecture instead of the legacy ATM-over-TDM one. Also, the use of IP as the transport
protocol potentially allows LTE traffics to share the same backhaul infrastructure with other
applications. This opens the way to an unified transport architecture.
However, this raises new concerns in term of security. The LTE standard [4][5][6] particularly
recommends to deploy ciphering and authentication means in order to protect both LTE signaling
plane and user plane traffics.
Figure 21 illustrates such a configuration where IPSec tunnels are established respectively
between the LTE Core elements and the evolved NodeBs (eNBs) and between different eNBs in order
to ensure the integrity (i.e. authentication) and the confidentiality (i.e. ciphering) of all encapsulated
packet flows. The LTE standard [4] especially recommends the use of IPSec ESP (Encapsulating
Security Payload) tunnelling mode as defined by the RFC 4303, with strong ciphering algorithm as
defined by the RFC 3602. This particularly applies when eNBs belong to untrusted locations/domains.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 46/76
Figure 21 - IPSec tunnel deployment in LTE
8.2.3 Traditional OPS deployment discussion
Figure 22 illustrates the IPSec ESP tunnel wherein the original (or encapsulated) IP packets are
entirely ciphered. In the case where IEEE 1588V2 messages are transported within these latter, it
becomes impossible for a BC or a TC to operate (e.g. modifying the Correction Field in the case of a
TC) in the middle of such an IPSec tunnel.
The deployment of BCs and TCs within the LTE secured environment is however still possible.
Figure 23 represents an example of such a possibility where a parallel IPSec tunnel architecture is
built to transport IEEE 1588V2 flows, thanks to a distributed tunneling scheme.
Figure 22 - IPSec ESP tunnel mode
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 47/76
Figure 23 - BCs deployment scenario within a LTE IPSec environment
Nevertheless, this dedicated security architecture adds complexity to the network (e.g. new
functionalities should be added onto intermediate network nodes in order to start or to terminate the
IPSec tunnels). It also implies additional management burdens (e.g. an additional certificate
hierarchy) and finally additional costs (e.g. higher OPEX).
To keep the transport infrastructure as simple as possible without sacrifying the traffic security
requirements, the natural way would be to transport IEEE 1588V2 packets using the same security
architecture as for LTE traffics. And, in this specific case, there should be new alternative OPSs with
regards to BCs and TCs concerns.
In the following Clauses of this Chapter, new agnostic OPS solutions will be presented. These
latter particularly rely on the use of a dedicated DSCP value transported within the IPSec tunnel IP
header in order to segregate IEEE 15588V2 traffics from LTE traffics. Indeed, the DSCP field within
the IPSec tunnel header is apparent to (i.e. can be read by) network nodes positioned in the middle of
the IPSec tunnel.
8.3 New protocol-agnostic mechanism to monitor the DA
8.3.1 Principle description with a first proposed implementation
Within the context of time distribution, delay asymmetry is a vector of synchronization
inaccuracy. Typically the achievable time accuracy is one-half of the delay asymmetry.
The new proposed DA monitoring mechanism, as described in this chapter, aims at minimizing
delay asymmetry induced by each network node. As such, this is an alternative to the measurement of
IEEE 1588V2 packet resident time performed by TCs, with the advantage of being protocol-agnostic
and capable to operate within an IPSec tunneling environment.
The principle of this solution consists in managing timestamp packets at the network node level,
so that Master-to-Slave packets experience similar queuing and switching effects as packets in the
opposite direction (i.e. Slave-to-Master). This solution particularly addresses the low frequency
variation of the delay asymmetry which is typically hard to filter by PTPV2 Slave algorithms [7].
A first way to implement the new concept is illustrated by Figure 24a and Figure 24b. Figure
24a illustrates the proposed path of a timing packet traversing the network node in the Master-to-
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 48/76
Slave direction. Figure 24b illustrates the proposed path of a timing packet traversing the network
node in the Slave-to-Master direction.
Figure 24 - First way to implement the new DA monitoring mechanism
As depicted in Figure 24a, an incoming timing packet (from the Master) experiences the
following sequence of queuing and switching events (the assumption here is that the input and output
ports are located in different interface cards; otherwise, switching operations would not be required):
P1 → P2 → P3 → P4 → P5 → P6 → P7
Note that paths P2 and P6 do not represent a path external to the switching matrix. Instead, they
represent a path through the switching matrix that results in the timing packet being deposited at the
input of the respective buffers.
It is important to stress here the new behavior of queuing buffers with regards to the traditional one:
- Within a common packet handling behavior, the queuing buffers are high speed memory
space dedicated to packet storage located upon a particular port card and once a packet has
been switched to a respective port card, it is at the suitable location to be placed in the
indicated buffer without further transitioning of the network node switching matrix.
- Within the new proposal, it particularly requires buffers presenting multiple inputs and/or
outputs (regarding their connectivity with other buffers). Moreover, this feature is only
applied to a selected set of packets, especially timing packets (e.g. IEEE 1588V2) identified
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 49/76
thanks to the information carried by the IPSec tunnel header (DSCP parameter) as previously
discussed.
As depicted, the transition delays related to a timing packet traversing the network node in the
Master-to-Slave direction is the sum of all elementary event delays:
dM→S = d(P1) + d(P2) + d(P3) + d(P4) + d(P5) + d(P6) + d(P7)
Similarly, Figure 24b depicts the path of a timing packet coming from the Slave side of the network.
The path of the timing packet traversing the network node through input/output buffers and
switching matrix is as follows:
P’1 → P’2 → P’3 → P’4 → P’5 → P’6 → P’7
The related network node transition delay of a timing packet traversing the network node in the Slave-
to-Master direction is thus:
dS→M = d(P’1)+d(P’2)+d(P’3)+d(P’4)+d(P’5)+d(P’6)+d(P’7)
As the result, the delay asymmetry produced by the network node is given by:
DA = dM→S - dS→M
DA = (d(P1)+d(P2)+d(P3)+d(P4)+d(P5)+d(P6)+d(P7))
- ((d(P’1)+d(P’2)+d(P’3)+d(P’4)+d(P’5)+d(P’6)+d(P’7))
Referring to Figure 24a and Figure 24b, transition delays d(P2) and d(P6) are respectively
similar to d(P’2) and d(P’6) as they are related to similar queuing events. In the same way, d(P4) and
d(P’4) refer to similar switching events.
Limiting the discussion to the low frequency variation component of the DA over time <DA>
(e.g. the result of DA after a low-pass filter or an averaging), the above transition delays have the
resulting effect which can be considered as a small constant value in the expression of <DA>
(asymmetry related to the structural hardware design of the network node). Therefore, the above
expression may be simplified by replacing the transition delays associated with traversing the
switching matrix and queuing events. This results in:
<DA> = ε+ (<d(P1)>+<d(P3) >+<d(P5) >+<d(P7) >) - (<(d(P’1) >+<d(P’3) >+<d(P’5) >+<d(P’7) >),
with ε being a constant value.
By further inspection of Figure 24a and Figure 24b it may be noted that certain of the respective
elementary delays refer to the transition delay within the same buffer, namely: P1 and P’3, P3 and P’1,
P5 and P’7, and, P7 and P’5. Even though d(P1) are different from d(P’3) with regards to their
respective instantaneous values, their average value (i.e. low frequency component) over a long
observation time should be the same as it consists of the same buffer characteristic. Thus, the low-
frequency component of transition delays for the respective queuing buffers effectively cancels out,
resulting in: <DA> ≈ ε.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 50/76
The low frequency component of the delay asymmetry is thus reduced to a constant
value (over time). Such a static asymmetry can eventually be measured and manually
provisioned at the Slave level, if it happens to be significative with regards to the
targeted synchronization accuracy.
8.3.2 Alternative implementation of the Delay Asymmetry mechanism
An alternative implementation of the new DA monitoring mechanism is illustrated in Figure
25a and Figure 25b below. Similarly to the first implementation, path P4 (resp. P’4) in Figure 25a
(resp. Figure 25b) does not represent a path external to the network node switching matrix. Instead, it
represents a path through the switching matrix (i.e. input and output ports on different cards) that
results in the timing packet being deposited at the input of the respective buffer.
The main difference with regards to the first implementation is that paths P2 and P6 (resp. P’2
and P’6) do represent a path external to the switching matrix in that these paths represent the timing
packet being deposited in the appropriate buffer on the same port obviating the need for a switching
operation. This behavior could be performed by using an external module connected to all ports.
Figure 25 - Alternate implementation of the new DA monitoring mechanism
8.3.3 Conclusion
The above new protocol-agnostic OPS allows for canceling the low frequency component of the
DA which represents a strong benefit as slave algorithms are typically not able to filter such a
component (high frequency component of the DA can be filtered out by the Slave somehow thanks to,
for instance, averaging techniques).
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 51/76
The depicted implementations demonstrate that the selection of a specific switching sequence at
the network node level can potentially allows for providing a more manageable behavior of the DA,
and thus a better end-to-end IEEE 1588V2 time synchronization performance.
In term of realization, traffic redirections from an output port towards another port on a same
network node, represents the biggest challenge related the new concept. Indeed, this requires a
forwarding table implemented at the output port, which is not the case for available network nodes.
8.4 New protocol-agnostic mechanism to dampen PDV
8.4.1 Packet Delay Variation (PDV) root causes
Packet Delay Variation (PDV) is essentially generated by buffering events within the network
nodes (e.g. switches or routers) along the communication path between the Master and the Slave. As
shown in Figure 26 (extracted from the ITU-T G.8061 Figure I.2), the network node can cope with low
traffic load so that packets (or frame) can be forwarded/routed within without additional delays
(except the minimum “propagation” delay). As the traffic load increases, while the processing capacity
of the node is not exceeded, the instantaneous packet rate may exceed the available processing
capacity, causing the need for a buffering sequence. Finally, at some traffic load threshold, the
processing capacity of the node is exceeded causing the delay to be increased further until packets are
dropped. Our study is rather focused on the network node state before such a processing overload
situation.
Figure 26 – Variation of delay with loading, as referred to G.8061 (Figure I.2)
Aiming at minimizing the PDV induced by queuing, a commonly used heuristic is to assign a
high priority Quality of Service (QoS; e.g. Premium class as compared to Best-effort one) to timing
packets. However, even in such a context, PDV experienced by timing packet is still impacted by the
following root causes:
- Root cause 1: a QoS assignment does not prevent the system from having PDV as the
transmission of a “Premium” packet can be delayed by an early departing “Best-effort” packet
which occupies the transmission link. And, as the size of the best-effort packet is not fixed and
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 52/76
predictable, this yields to an undeterministic PDV for the ‘premium’ timing packets. Such an
issue is illustrated by the Figure 27 below.
Figure 27 – “Premium” packet transmission delayed by a “Best Effort” packet
- Root cause 2: in the same way, competition between different synchronization flows (within
the Premium QoS) for the same output transmission resource induces unpredictable PDV for
each of these flows. In particular, when timing packet flows are sent with very similar
frequencies, the competition between the different flows for the access of the transmission
medium generates a “beating effect” for each of them [8]. Figure 28 below illustrates such a
phenomenon.
Figure 28 - Beating effect as referred to G.8261 (Figure 20)
8.4.2 Case of single synchronization flow
In order to avoid the aforementioned root cause 1, the new PDV dampening mechanism
consists in systematically delaying each individual “Premium” timing packet (transmitted
through a network node) by a fix delay before serializing it onto the transmission medium. The
delay is set to ∆∆∆∆MS and ∆∆∆∆SM for respectively the Master-to-Slave and the Slave-to-Master
direction.
In a first implementation, ∆∆∆∆MS can take one of the following values:
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 53/76
- )(MTUimeserializeTMS ====∆∆∆∆ : the time to serialize a packet of MTU (Maximum
Transmission Unit) size. For example, in 1Gbps Ethernet technology, the MTU size is 1518
octets, and sMTUimeserializeTMS µµµµ12)( ========∆∆∆∆
- ))(( etSizesBestEfPackMaximeserializeTMS ====∆∆∆∆ : the time to serialize a “Best effort” packet
of maximum size.
The above values allow to make sure that a timing packet will not be further delayed (by
the serialization of a best-effort packet) when it comes the moment to serialize it onto the
transmission medium.
During the “waiting” period (∆∆∆∆MS) of the timing packet, and in the absence of another
timing packet to transmit, the network node can complete the serialization of “Best effort”
packets. The network node should make sure about the size of these “Best effort” packets so
that their serialization onto the transmission medium will not further delay the next waiting
timing packet. This strategy can lead to wasted bandwidth when packet fragmentation is not
possible (e.g. IPSec tunnel case). However, this wasted bandwidth is minimal as the timing
packet bandwidth is typically small with regards to the overall consumed bandwidth. Indeed,
as reported by [9] for IEEE 1588V2 Telecom profile typical message rates, a Slave generates at
worst 0.3 Mbps for an FE link of 100 Mbps. As the monitoring point comes closer to the
Master, this consumed bandwidth increases with the aggregation of IEEE 1588V2 flows for
multiple slaves. Assuming that a Master is deployed at each LTE core element to support an
average of 1000 of LTE eNBs and each eNB represents a Slave clock, then the total IEEE
1588V2 consumed bandwidth is at worst 300 Mbps. As the core links will most likely have 10s
of Gbps bandwidth capacity, the IEEE 1588V2 traffic still remains a very small portion of the
overall bandwidth.
In a second implementation, ∆∆∆∆MS can take one of the following values in order to
optimize transmission bandwidth especially when (“Best effort”) packet fragmentation cannot
be activated:
- )%( MTUximeserializeTMS ××××====∆∆∆∆ with 100<<<<x
- ))(%( etSizesBestEfPackMaxximeserializeTMS ××××====∆∆∆∆
with 100<<<<x
In this case, there is not 100% guarantee that the waiting timing packet will not be further delayed.
Indeed, the serialization of an MTU-size “Best effort” packet during the waiting time ∆∆∆∆MS raises an
issue related to a conflicting usage of the transmission medium when the waiting timer expires.
However, the x
value can be chosen to statistically minimize such conflicting situations. In any case,
the PDV magnitude is strongly reduced in proportion to x−−−−1 value.
8.4.3 Case of multiple concurrent synchronization flows
In order to avoid root cause 2, as raised above, within the context where multiple Master
clocks transmit timing messages periodically with the same frequency mf , the new PDV
dampening mechanism consists in delaying systematically individual packet of different
“conflicting” (as per the aforementioned “beating effect”) timing streams. The detection of the
“beating effect” and the computation of the different delays require a learning period which
could be as small as mm fT /1==== .
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 54/76
The whole procedure is illustrated by Figure 29 below. The value for ∆∆∆∆MS can be for
example )(MTUimeserializeT
or a multiple of that time depending on the “conflicting”
situation and the size of different timing packets.
It is to be noted here that different timing streams can be segregated, for instance, using
their respective tunnel destination IP addresses for the Master-to-Slave direction and their
respective tunnel source IP addresses for the opposite direction.
This delaying mechanism should be performed at each network node along the communication
path. Each of these nodes requires a theoretical convergence time in order to analyze
“conflicting” situations and to compute different appropriate delays: the first node convergence
time is mm fT /1==== , the second node convergence time is
mT××××2 and the Nth node
convergence time is mTN ×××× . However, thanks to the periodical characteristic of the packet
streams, the delaying mechanism may only be required at the first node.
Figure 29 - Timing flows delayed by a multiple of the elementary delay ∆ms
The different computed delays i∆∆∆∆
applied by the network nodes to different synchronization
streams i
can be reported to the appropriate Slave clock via a centralized management system. This
strategy is possible as i∆∆∆∆
values do not vary rapidly with time which is different from the TC case
where individual residence times must be determined for each packet and then communicated to the
slave. The important difference is that i∆∆∆∆ is bound to the whole stream and not to each individual
packet within this stream.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 55/76
8.4.4 Discussions on implementation and conclusion
Implementation-wise, the mechanism can be easily realized thanks, for instance, to a
simple ordered time descriptor chain. Each descriptor would point to the associated timing
packet and indicate the time when this packet should be transmitted. Figure 30 below provides
with the illustration of such an implementation. The transmitter can simply read the chain
sequentially and transmit the associated packet at the time instant indicated by the descriptor.
Figure 30 – Possible implementation of the new PDV dampening mechanism
As a conclusion, the new methodology as described in this chapter should simply allow for
dampening PDV effects by providing more consistent network transit delays. This new OPS could
particularly be useful for the distribution of a frequency reference. A performance evaluation of such a
methodology is under progress.
8.5 Conclusions and outlook (Clauses 8.2 to 8.4)
Facing stringent LTE synchronization requirements, OPSs are largely recommended for
reaching the targeted performance with PTPV2. Within this context, BC and TC concepts introduced
by the IEEE 1588V2 standards represent a straightforward solution. However, stringent security
requirements in LTE environments raise strong concerns related to their implementation and
deployment, becoming either too complex or even not applicable. In such cases, Clauses 8.2 to 8.4
propose two protocol-agnostic OPSs allowing for addressing all security constraints within the single
and fully deployed IPSec protocol. Considering an advantageous management of the IEEE 1588V2
traffic as part of the LTE application one, the two presented mechanisms particularly make use of an
apparent DSCP code. They allows for efficiently managing timing packets by performing network
node operations which aims at dampening network dependencies.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 56/76
The first mechanism focuses on PDV dampening whereas the second proposes an innovative
way to minimize the low frequency component of DA, typically hard to manage by IEEE 1588V2 Slave
algorithms. These two OPSs could be activated altogether or separately according to deployment
scenarios. Thus, in a time distribution scenario, where IEEE 1588V2 is deployed with the frequency
support from Sync-E, DA is the only issue, and only the new DA dampening needs to be activated.
Suppressing the Sync-E support in the previous scenario makes PDV also a concern. Both new
mechanisms then are recommended.
In any case, synchronization is presently not a source of revenue for operators and the
advantage of the IEEE 1588V2 protocol is that it can be deployed using the existing or the common
transport infrastructure. This makes IEEE 1588V2 very attractive to operators.
Nevertheless, the introduction of BCs and TCs certainly raises the initial costs of the IEEE
1588V2 implementation within the network. Thus, there are needs to further investigate other OPSs
that can be shared amongst different protocols in order to amortize synchronization deployment
costs.
8.6 Link-to-link TOD distribution using the Sync-E’s ESMC messaging
channel
8.6.1 Context description
As more and more traffic is being delivered in Ethernet format, carriers are realizing the
advantages to converging on a pure Ethernet infrastructure, but are not able to do so with Ethernet as
presently defined. New ITU-T Recommendations, such as Y.1731 for Ethernet OAM and G.8031 for
Ethernet protection switching, provide features essential for operational maintenance of Ethernet
networks. Recently proposed mechanisms for configuring point-to-point connections in Ethernet
networks close another gap between Ethernet and SONET/SDH networks.
The final difference between conventional TDM-based networks and Ethernet is that the former
also transport frequency information, needed for some applications, while Ethernet does not. A
comprehensive solution for that problem has been achieved when the ITU-T has developed
Synchronous Ethernet that locks the timing of the Ethernet physical layer, much the same way that
SONET/SDH does. This enables simple recovery of the physical layer timing, which can be used
directly by locking downstream networks and applications, or indirectly by using the physical layer
timing as a common reference.
Nevertheless, today end applications (e.g. 3G and WiMAX cellular base-stations) are becoming
more and more dependent on the distribution of a precise TOD or phase reference in addition to
accurate and stable frequency reference. This ruling is even truer when it comes to future applications
such as LTE cellular base stations or within Single Frequency Networks (SFN) such as the Terrestrial
Digital Video Broadcasting (DVB-T) standard.
Existing packet based time distribution protocol such as IEEE 1588v2 (PTP) and the Network
Time Protocol (NTP) have been markedly improving over the recent years and have been shown to be
able to deliver the required level of performance in some network scenarios. Nevertheless, being
‘upper layers’ type of time distribution protocols (in the general case) they suffer from ‘upper layer’
type of network interferences such as Packet Delay Variation (PDV) and network re-route delay
changes. At the end of the day, these methods cannot be trusted to always perform as expected under
changing network scenarios.
A known fact is that in order to be able to assure a certain level of performance under (almost)
all network conditions the information has to be carried in the physical layer (or very close to it). This
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 57/76
can be accomplished either by applying precise HW timestamping very close to the physical layer to
support the upper layer synchronization distribution protocol (e.g. TCs and link-by-link BCs in the
case of PTP) or by applying the synchronization distribution directly at the physical layer (or very
close to it). An FT contribution to Q13/SG15 of the ITU-T [10] purposed exactly that: a link-by-link
phase distribution scheme based on link layer OAM messages.
The apparatus purposed in this Clause presents some modifications and improvement to the
scheme presented in [10]. Specifically, the Ethernet Synchronization Messaging Channel (ESMC),
currently used to carry Ethernet Synchronization Status Messages (ESSM) across Synchronous
Ethernet capable links, is suggested as a container to carry the time information between neighboring
nodes.
8.6.2 The ESMC
The Ethernet SSM is an ITU-T defined [11] Ethernet slow protocol that was defined in order to
distribute Quality Level (QL) information across Synchronous Ethernet capable Ethernet links.
Synchronous Ethernet SSM or ESSM distributes the same SDH/SONET 4-bits QL words. In
SDH/SONET, these 4-bits are carried in a fixed location in the frame header. In Synchronous
Ethernet, on the other hand, these 4-bit words are distributed using a dedicated unidirectional IEEE
802.3 Organization Specific Slow Protocol (OSSP) termed Ethernet Synchronization Messaging
Channel (ESMC).
In order to comply with clock selection procedures given in ITU-T Recommendation G.781 (e.g.
maximal time to generate a clock source switch over when the current selected source QL changes)
two ESMC message types have been defined. Information 1 second ‘heart beat’ messages are used to
continuously inform the downlink node of the current QL value. Additionally, in order to minimize
short-term interferences and large phase excursions during transition to holdover or clock source
switch-over, an event message is immediately send towards the downlink element to signal it on the
changed QL. To protect against possible failure, the lack of the messages is considered to be a failure
condition. The protocol behaviour is such that the SSM value is set to DNU if no SSM messages are
received after a five second period.
ESMC frames bear an ITU-T organizationally unique identifier (OUI) 00-19-A7 and a slow
protocol subtype 0x0A that distinguish it from other OSSPs. The SSM quality level is carried in a type
length value (TLV) field, which is contained within the ESMC PDU. Two types of ESMC PDU frames
are defined and are distinguished by the C flag. These are the ESMC information PDU and the ESMC
event PDU mentioned above.
The ESMC PDU format is shown in Table 2. The QL TLV is shown in Table 3. The same QL TLV
is used for both the information and event messages. To allow for potential hardware
implementations, the SSM TLV is always sent as the first TLV in the data/padding field. This means
that the QL indication always remains fixed in the PDU. Any padding must occur after the SSM TLV.
Octet number Size/bits Field
1-6 6 octets Destination Address =01-80-C2-00-00-02 (hex)
7-12 6 octets Source Address
13-14 2 octets Slow Protocol Ethertype = 88-09 (hex)
15 1 octets Slow Protocol Subtype =0A (hex)
16-18 3 octets ITU-OUI = 00-19-A7 (hex)
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 58/76
19-20 2 octets ITU Subtype
21
bits 7:4(note 1) Version
bit 3 Event flag
bits 2:0(note 2) Reserved
22-24 3 octets Reserved
25-1532 36-1490 octets Data and Padding (See point J)
Last 4 4 octets FCS
Note 1: Bit 7 is the most significant bit of Byte 21. Bit7 to bit 4 (bits 7:4) represent the four bit version
number for the ESMC.
Note 2: the three least significant bits (bits 2:0) are reserved.
Table 2 – ESMC PDU format (Table 11-3/G.8264)
ESMC PDUs have the following fields in the order specified above:
a) Destination address (DA): This is the IEEE-defined slow protocol multicast address. The format is defined in Annex 43B of [IEEE 802.3].
b) Source address (SA): The source address is the MAC address associated with the port through which the ESMC PDU is transmitted.
c) Slow protocol Ethertype: ESMC PDUs must be type encoded and carry the slow protocol type field value.
d) Slow protocol subtype: Assigned by the IEEE and fixed with a value of 0x0A.
e) ITU OUI: Organizational unique identifier assigned by the IEEE registration authority.
f) The ITU subtype is assigned by ITU-T. The value of 00-01 applies to all usage defined in this Recommendation.
g) Version: The four-bit field indicates the version of ITU-T OSSP frame format. This field shall contain the value 0x1 to claim compliance with version 1 of this protocol.
h) Event flag: This bit distinguishes the critical, time-sensitive behaviour of the ESMC event PDU from the ESMC Information PDU. A value of 1 indicates an event PDU, a value of 0 indicates an information PDU.
NOTE 1 – The behaviour of the event PDU is similar to the critical event defined for Ethernet OAM in clause 57
of [IEEE 802.3]. Event messages need to meet processing times defined in [ITU-T G.781].
i) Reserved for future standardization (27 bits). These fields are set to all zero at the transmitter and are
ignored by the receiver.
j) Data and padding: This field contains data and necessary padding to achieve the minimum frame size
of 64 bytes. The PDU must be an integral number of bytes (octets). Padding characters are not defined
and are ignored by receivers.
NOTE 2 – The recommended maximum size for the ESMC PDU is 128 bytes as per Annex 43B of [IEEE 802.3].
However, PDU sizes greater than 128 bytes may be permitted.
k) FCS: Four-byte frame check sequence as defined in clause 4 of [IEEE 802.3].
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 59/76
Octet number Size/bits Field
1 8 bits Type:0x01
2-3 16 bits Length: 0x0004
4 bits 7:4 0 (unused)
bits 3:0 SSM code
Note 1: Bit 7 of Octet 4is the most significant bit. The least
significant nibble, bit 3 to bit 0 (bits 3:0) contain the four
bit SSM code.
Table 3 – QL TLV format (Table 11-4/G.8264)
According to [11], additional TLVs comprising future extensions to the ESMC messages are
possible. Such extensions might include ‘trace-route’ like mechanisms to better mitigate timing loops.
Another possible extension to the ESMC messages is the support for phase/TOD distribution over
Ethernet links.
8.6.3 Proposals
The proposal comprises 3 major parts:
The first part of the proposal is to enhance the current ESMC messages format
with an additional TLV that will carry TOD information across an Ethernet link.
It is therefore a fact that unidirectional ESMC messages carrying QL information are distributed
across both directions of any Sync-E capable line2. Hence, Sync-E capable Ethernet networks already
possess a very bandwidth efficient communication channel between neighbouring nodes. As already
discussed in the motivation part of this Clause, extending this messaging channel with a complete
TOD distribution scheme seems, therefore, something of great value.
In order to make it happen and allow ESMC messages to convey TOD information two
additional information entities needs to be added to the current standard ESMC messages format:
(1) Timestamps (TS) that will carry the TOD information from one side of the link to the other.
As will be presented shortly, two timestamp fields will be added to each ESMC message using
the TLV format presented in Table 3 above.
(2) A Sequence number (SN) that will mark the ESMC messages according to their chronological
transmission order. As will be presented shortly, the sequence number will be used to
correlate a timestamp to the specific ESMC message it belongs to.
One possible way of incorporating these changes in the current standard ESMC message format
is given in Table 4 and Table 5 below. The changes comprise an additional 24-bit Sequence Number
introduced into currently reserved 3 octets (22 to 24) in the message header. This will be a standard
running sequence number that will reset to 0 once reaching its maximal value (223-1). The two
2 According to [11] there might be Synchronous Ethernet links with partial support. In other words, where the physical layer
clock as well as the QL distribution are being distributed in one direction only or where ESMC messages are not being
transmitted at all. Nevertheless, these cases can be regarded as exceptional and are only relevant for last mile Ethernet links.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 60/76
timestamps will be introduced into a new dedicated TOD TLV that will be placed immediately after
the currently standard QL TLV (starting at octet number 29). Each of these timestamps will have the
standard PTP 80-bit Timestamp format [12] to allow for better interworking with PTP as will be later
on explained and demonstrated (alternatively these timestamps could have been designed to bear the
standard NTP format of 64-bits. Nevertheless, as PTP seems to be getting more traction from telecom
service providers, especially in the cellular backhaul arena, and a lot of discussions and work had been
taking place concerning better PTP/Sync-E interaction, it seems only natural to adopt the PTP
Timestamp format).
The first Timestamp (octets 4-13) records the precise time (according to the local node time
counter) at which the last periodic ESMC message was received in the specific node, and is used to
feedback the other end of the link with that information. Following (octets 14-16), the sequence
number of the specific received periodic message that the timestamp refers to, is also sent to allow the
receiver at the other end to correlate between transmission and reception time events on this specific
direction of the link.
The second Timestamp (octets 17-26) records the precise time (according to the local node time
counter) at which the last ESMC message was transmitted by the specific node (alternatively this
Timestamp could have been made to comprise the precise time at which the current transmitted
ESMC message left the node. However, this would impose on-the-fly HW timestamping mechanism at
each participating node. Such a mechanism might increase the overall cost of the scheme, but more
important, will unnecessarily degrade the accuracy of the timestamps by some amount).
As will be explained in the next clause, each link end (node), independently sending and
receiving ESMC messages, will have all the information it needs to recover TOD information from the
other end. This concept is very much in the spirit of standard Sync-E frequency distribution where
(except for some specific Ethernet interfaces) the frequency information can, potentially, go both
directions and correct distribution direction is determined by the overlay clock selection procedure.
Octet number Size/bits Field
1-6 6 octets Destination Address =01-80-C2-00-00-02 (hex)
7-12 6 octets Source Address
13-14 2 octets Slow Protocol Ethertype = 88-09 (hex)
15 1 octets Slow Protocol Subtype =0A (hex)
16-18 3 octets ITU-OUI = 00-19-A7 (hex)
19-20 2 octets ITU Subtype
21
bits 7:4(note 1) Version
bit 3 Event flag
bits 2:0(note 2) Reserved
22-24 3 octets 24-bit Sequence Number (SN)
25-1532 36-1490 octets Data and Padding (See point J)
Last 4 4 octets FCS
Table 4 – TOD enhanced ESMC PDU format (proposed changes are highlighted)
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 61/76
Octet number Size/bits Field
1 8 bits Type:0x02
2-3 16 bits Length: 0x0017
4-13 80-bit Timestamp of last received
periodic ESMC message
14-16 24-bit Sequence Number of last
received periodic ESMC message
17-26 80-bit Timestamp of last transmitted
periodic ESMC message
Table 5 – TOD TLV format (proposed changes are highlighted)
It should be emphasized that the above TOD extensions need only be applied to periodic (Event
flag = ‘0’) ESMC messages with the average rate of 1 PPS. Event driven ESMC messages (Event flag =
‘1’) need not incorporate these modifications as they will not contribute anything to the TOD
distribution scheme. Hence, as far as TOD distribution is concerned, ESMC messages with Event flag
= ‘1’ should be ignored.
The second part of the proposal is to define a new apparatus exploiting the new
proposed ESMC messages TOD TLV extension to distributed TOD information across
the Ethernet link.
Lets us begin by assuming two neighboring Sync-E capable nodes X and Y that are, in the
general case, part of a Sync-E frequency distribution chain according to [11], [8] and [13]. Each of
these nodes will transmit standard QL ESMC messages to the other Naturally, depending on the
decided timing information distribution direction, one node will transmit QL=’DNU’ to the other side.
As will be discussed in the next clause, this QL information could also come very handy for the TOD
distribution process, but for now we will assume that the frequency distribution and the TOD
distribution process are completely orthogonal.
By extending the ESMC messages format as explained in the previous clause, each node will
now be transmitting the following TOD information to its neighbor:
(1) Precise local time of last received periodic ESMC message marked by: TSR(k-1).
(2) Precise local time of last transmitted periodic ESMC message marked by: TST(n-1),
where k and n are the different periodic ESMC messages indexes on both directions.
Additionally, at each node, the timestamp of the precise local time at which the last periodic
ESMC message was transmitted (TST(k-1)) as well as the timestamp of the precise local time at which
the last periodic ESMC frame was received (TSR(n-1)) are recorded and saved. A graphical
representation of the above text is given in Figure 31.
At each node, every 1 second after the completion of a successful ESMC message reception, the
exactly same following four timestamps will be available: TST(k-1), TSR(k-1), TST(n-1) and TSR(n-1).
Using these four timestamps each node could now calculate an estimation of the time error between
himself and its neighbor using the following, very known, time error estimation formula:
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 62/76
The calculation time error estimation could, at this point, be used to time-lock each node to the
other. The only left issue to be resolved is the correct direction (based on some correctness criteria as
will be discussed in the clause) of the TOD information flow. In other words, which node is going to
play the role of the Master and which one will be Slave, for this specific link (not setting an
unambiguous master-slave relationship for the specific link might results in a “time-loop” situation
where, in a very similar manner to timing-loops in frequency distribution, both ends will
simultaneously try to maintain lock to the opposite, resulting in unstable, unpredicted behavior).
Clearly, the above suggested apparatus introduces a 1-second delay (at least) in the time error
estimation procedure, and therefore, also in the TOD recovery process. At first glance, this fact might
be seen as problematic (1 second in the time error estimation procedure means that the time error
could be reduced up to that residual time error, stemming from the time error development over that
1-second calculation delay). Nevertheless, taking into account that the aforementioned residual time
error is negligibly small if the time counters on both nodes are timed by the Sync-E physical layer
recovered clock, this does not appear to be such a big sacrifice at the benefit of avoiding the
timestamping accuracy degradation that accompanies any on-the-fly timestamping mechanism.
Many time (clock) recovery apparatuses can be envisaged to perform the time lock procedure
based on the calculated time error estimations. This issue is currently left for further study.
Figure 31 – TOD ESMC messages exchange
The third part of the proposal is to define the TOD selection process that will be
used in each node to select the best TOD information source among possible multiple
sources. This TOD selection process is analogues to the frequency selection process
defined in [14].
Simplifying the matter a little bit, TOD distribution can be broke-down into the following
elements:
(1) A frequency distribution scheme that maintains the local time counters along the distribution
node in the correct pace.
(2) An additional occasional phase correction information that align the phase of the clocks along
the distribution chain, removing any developed bias.
In that view, the performance of a link-by-link TOD distribution scheme is very much impacted,
in the general case, from the same factors that impact the performance of a link-by-link frequency
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 63/76
distribution scheme. In other words, similar to the frequency distribution case, the specific free-run
quality of the local clocks along the distribution chain plays an important role. On the other hand, as
this is a link-by-link TOD distribution scheme, other attributes such as path asymmetry and path PDV
are of less importance. In other words, TOD distribution chains that comprise stable and accurate
local oscillators have better chance of delivering good TOD distribution performance.
Another reinforcement of that approach is the fact that the other capabilities of the conventional
frequency selection process such as the hold-off and wait-to-restore functions as well as the various
user configurable parameters (priority tables, force/manual switch commands etc.) seems to be also
desired functionalities for the TOD distribution case.
It therefore seems logical that the same standard clock selection process already established for
the act of frequency distribution could be largely reused for TOD distribution.
Two strategies of incorporating the fundamentals of frequency clock selection into TOD clock
selection can be envisaged:
(1) Using the exact same clock selection process for both frequency and TOD. This basically
means staying with the existing clock selection process, using it to select both frequency and
TOD. Working in such a strategy ensures that the same reference clock will always be selected
for both frequency and TOD.
(2) Using a similar but orthogonal clock selection process for TOD. The existing clock selection
process will be modified to better suit TOD source selection (for example, the specific values
of the hold-off and wait-to-restore timers might change or a different user priority table might
be used). The new TOD selection process will run in parallel to the frequency selection
process, which might result in different reference clocks selected for frequency and TOD. This
obviously means different network paths for frequency and TOD.
The first strategy is very simple and straightforward. No change (almost) to existing clock
management agents in live products will be needed. The frequency and TOD distribution will always
go together hand by hand following the same network path, a fact, that will simply overall network
management. Nevertheless, such a scheme offers little flexibility for the user to force its own TOD
distribution preferences, which might be different from his frequency distribution preferences, into
the selection process.
The second strategy, on the other hand, gives the user the freedom to design and configure its
TOD selection preferences independently from his frequency distribution preferences. Nevertheless, it
doubles the required management efforts.
8.6.4 The Simulation Environment
A complete closed simulation environment for testing the concept of TOD distribution using the
Sync-E ESMC messages was created. The simulation was created in SIMULINK of MATLAB where
participating key SIMULINK blocks where written in C. SIMULINK was selected due its nice GUI
properties as well as its inherit simulation time axis that makes it a very convenient simulation
environment to characterize time varying processes.
The simulation environment was based on the following fundamental assumptions:
(1) Either linear or ring topologies must be supported.
(2) The Ethernet physical layer’s frequency is always synchronized using the underlying standard
Sync-E frequency distribution apparatus. In other words, the amount of time error introduced by
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 64/76
the frequency instability of the physical-layer is very small and can be regarded as negligible in
practice. This issue is very important and will be addressed again later in this chapter.3
(3) The delay introduced by the Optic-fiber lines is completely symmetric (the same in both
directions). This assumption is known not to hold true in real fiber-optic installation where there
is always a slight difference of the length of the fiber between both directions. Nevertheless, as this
source of error is quite deterministic in nature and can only be compensated by applying a-priori
delay measurements on the fiber lines, there is no real point in simulating it4.
(4) Queuing delays at the output buffers of the participating network elements do not have any
impact on the time distribution performance. The reason for that is that no ‘on-the-fly’
timestamping in involved in the suggested scheme so the timestamping of the egress packets can
be directly taken at the output of the transmission queue just before the first byte of the packet is
being transmitted over the link. See Figure 31 for more information.
The simulation environment is comprised of a few prime functional blocks and is driven by the
interaction between these blocks. The connection between the prime blocks is defined by pre-defined
APIs so that any combination of blocks or network topology can be created and simulated for. By
extending these common APIs, the simulation can be rather easily extended, in the future, to cover
additional aspect and simulate additional effects. The following clauses will describe the prime
functional blocks and their relevant APIs.
8.6.4.1 Node block
8.6.4.1.1 Input/output ports
The node block comprises the fundamental prime functional block of the simulation (see Figure
32). It characterizes a basic 3-ports network element (a L2/L3 switch). All the ports are bidirectional.
Two of the ports, termed ‘L’ (Left) and ‘R’ (Right) are used to connect the block to its neighboring
node blocks, thus, forming the required network topology. The 3rd port termed ‘X’ (eXternal) is used
to input/output information (e.g. TOD information) from/to an external source/destination.
Figure 32 – node prime functional block inputs/outputs
3 The simulation was designed in such a way that it could be easily extended to take imperfections of
the underlying physical layer frequency distribution into account. 4 We could simply argue that the resultant introduced time error over a given link would exactly equal
half of the delay asymmetry.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 65/76
The L and R ports are defined by the following API, called the ESMC signal API, and that is
comprised by the following signals:
Signal number Size [octets] Content
1 1 (char) QL � Quality Level indication
carried by the ESMC message
according to [11]
2 2 (unsigned
integer)
Sequence Number of current
transmitted periodic ESMC
message
3 4 (2 x unsigned
integer)
Timestamp of last transmitted
periodic ESMC message
4 2 (unsigned
integer)
Sequence Number of last
received periodic ESMC message
5 4 (2 x unsigned
integer)
Timestamp of last received
periodic ESMC message
6 double Physical-layer instantaneous
frequency (currently taken to be
constant)
Table 6 – Signal structure of the ESMC signal API
By connecting the ‘L’ port output of one node block to the ‘R’ port input of the next node block and
vice-versa, ESMC messages with the above content can flow between the blocks, delivering the
required information back and forth. It should be noted that, for the sake of the simulation, the ESMC
message rate was selected to be fixed at 1 message per second (as defined by [11] for the ESMC
periodic messages), with the ESMC event messages (that are transmitted immediately on the line
should the QL value change) not taking any active role in the TOD distribution process.
The ‘X’ port (input/output) is defined by a different API, termed Time/Frequency reference
API, and that is comprised by the signals:
Signal number Size [octets] Content
1 1 (char) QL � Quality Level indication of
the incoming/outgoing TOD
reference
2 4 (2 x unsigned
integer)
Accurate TOD representation of
external clock source (input) or
the node itself (output)
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 66/76
3 double Physical-layer instantaneous
frequency of the external clock
source (input) or the node itself
(output) (currently taken to be
constant)
Table 7 – Signal structure of the Time/Frequency reference signal API
The external API allows a specific node block to be timed (derive time and frequency) by an external
clock source or supply its own time and frequency to an external end-application (e.g. a time error
measurement device). This interface can thought of, in practical aspects, as a frequency reference
signal (e.g. a 2.048MHz clock) accompanied by a TOD reference signal that can realized, for example,
by a 1 PPS signal (transmitted over a dedicated BNC interface) accompanied by a matching standard
TOD code, such as NMEA 0183, transmitted over an RS-442 serial link. As these dedicated reference
interfaces are usually used to carry the time/frequency information over a relatively very short
distance (not more than a few meters usually), no inherit time error is assumed for this interfaces.
8.6.4.1.2 Internal structure
In terms of internal structure, the node prime functional block resembles a standard Sync-E
capable device with the additional functionality of the time recovery. The internal structure is depicted
in Figure 33.
The ESMC L/R Tx/Rx processes are derived by the local frequency, time and QL representations
marked by the frequency synthesis, time register and QL memory blocks. The ESMC blocks exchange
ESSM messages with their neighboring nodes and derive physical-layer frequency, time error and QL
from each. Additionally, a frequency, time and their associated QL might be available from some
external reference source.
Based on the incoming QL information from all 3 sources, the G.781 based clock selection block
selects the one with the highest QL to be the reference for the frequency and time recovery blocks. The
frequency recovery block is based on the G.8262 EEC characteristics while the time recovery block
was taken to follow the clock discipline algorithm of the NTPv4 standard [29] (see more in next
clause). The resultant frequency and time recovered information is then used to ‘close the loop’ and
derive the frequency and time synthesis blocks as well as the QL memory that, on their turn, will
derive, again, both the ESMC blocks as well as any external device. It should also be highlighted that,
in between updates, the node’s local time register is advanced by the node’s recovered clock. This is
marked by the small arrow connecting the two blocks.
Note: As was already mentioned before, the current simulations carried assumed perfect frequency
recovery by the node block with the understanding that for most scenarios the contribution of this
noise source would be miniscule. Nevertheless, as can be seen in Figure 33, it should be quite easy to
introduce imperfect frequency synchronization into the simulation environment simply by replacing
the currently perfect G.8262 based frequency recovery block with a more realistic functional block
that has a finite PLL bandwidth as well as a non-zero phase noise local oscillator.
Having no introduced noise on the account of background traffic (output Queue) nor due to
frequency recovery imperfections, the only noise source contributing here is the timestamping
quantization noise. In real implementations this noise is generated by the imperfect (usually HW
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 67/76
realized) timestamping procedure embedded within the ESMC Tx/Rx processes. The timestamping
quantization noise was realized in this simulation environment as the uniformly distributed process
with distribution where σ is the noise standard deviation that can be a-
priory configured.
Figure 33 – node prime functional block internal structure
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 68/76
8.6.4.1.3 Time recovery block
The time recovery block implements the NTPv4 clock disciplining algorithm as it appears in
[29]5. The NTPv4 clock disciplining algorithm was chosen for this purpose for the following reasons:
i) It is a rather simple algorithm that is based on well-known FLL/PLL designing principals.
ii) It is the only truly standardized time recovery algorithm that exists today.
The above reasons makes NTPv4 an almost perfect candidate for the simulation as it demonstrates
that using the suggested approach, the time recovery function, that in built to be quite complex in
many other time distribution apparatuses, can be maintained quite simple and still meet very good
performance.
The NTPv4 clock disciplining algorithms was used in this simulation with the following
parameters:
STEPT . 128
WATCH 900
PANICT 1000
PLL 10
FLL 1
AVG 4
ALLAN 1500
LIMIT 30
MAXFREQ 500e-6
PGATE 4
PRECISION -18
IGNORE 0
These parameters were chosen to yield fast time recovery convergence while still maintaining good time-stamp quantization noise filtering capabilities. For more in-depth information regarding the NTPv4 clock disciplining algorithm please refer to [29] and [30].
8.6.4.2 Primary Reference Time Clock (PRTC) block
The PRTC block is a complementary prime block to the node block. Its purpose, within the
simulation environment, is to supply a time/frequency reference to a given node block, thus
originating a time/frequency distribution chain (see Figure 32). It is comprises one ‘QLin’ input port,
used by the simulation environment to control its specific QL value, as well as one ‘Refout’
Time/Frequency reference signal API (as already defined) output port.
Figure 34 – PRTC prime functional block
5 It should be noted that only the clock disciplining algorithm (time servo-loop) of [29] was
implemented and not the entire NTPv4 time recovery suit that comprises many additional processes
and algorithms dedicated for mitigating impairments introduced by the packet network (i.e. filtering,
selection, clustering and combining algorithms).
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 69/76
8.6.4.3 Time error analyzer block
The Time error analyzer block is another complementary prime block that constitutes the “end-
application” of the simulation environment. As can be seen in Figure 35 below, the block has two
input ports of type Time/Frequency reference signal API. The ‘Rec in’ input port is used to receive the
time/frequency information from the DUT (the recovered time/frequency information) whereas the
‘Ref in’ input port is used to receive the time/frequency reference from which the comparison as well
as the error calculation would be derived.
Figure 35 – Time Error Analyzer prime functional block
The measured time error through the entire simulation time is recorded by the block and can be
monitored via its inherit scope function. The time-error scope is presented in Figure 36.
Figure 36 – Time error scope
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 70/76
The upper subplot presents the QL of the incoming clock (European SSM codes) while the bottom
subplot presents the time error, in every simulation instant compared to the time reference input.
8.6.5 Simulation Results
Two important properties of the new apparatus were examined by the simulation:
1. Steady-state: The steady-state time error performance across an entire distribution path
comprised of a number of nodes as well as the time to reach a complete steady-state
conditions across the entire distribution path.
2. Restoration condition: The time-error performance while doing network time distribution
restoration (as a result of some configuration change or fault detected) as well as the time to
reach a complete network restoration.
In order to test the steady-state property, the ring network topology presented in Figure 37 was
simulated. 10 nodes are connected in a ring topology. The link between node10 and node1 is left
disconnected in order to avoid an Ethernet infinite loop6.
Figure 37 – Steady-state test setup
6 Such a property is usually achieved by employing some type of Ethernet ring protocol (e.g. G.8032)
that makes sure that at least one link within the ring is left disconnected at any given time.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 71/76
The PRTC connected to node1 with QL PRC is starting the TOD distribution chain at simulation time
0. The instantaneous QL and TOD associated with each node along the distribution chain is
monitored by the time error analyzer. In order to introduce some randomness into the time
distribution process the initial TOD of each node was randomly uniformly selected in a time window
of ±1 msec compared to some common time epoch. The timestamping quantization noise was selected
to be approximately 5.7 nsec yielding a quantization noise magnitude of ±20 msec, a customary noise
level in today’s HW implementations.
The obtained results are presented in Figure 38 and Figure 39 (each color represents a different node
according to the below given key). As can be seen from Figure 38, the TOD convergence process for
the entire distribution chain of 10 nodes take about 160 seconds to complete, while the QL
propagation time is, as was expected, 10 seconds (one second for each participating node). According
to Figure 39, steady-state performance is reached after ~300 seconds with a time error lower than ±10
nsec.
Figure 38 – Steady state test: distribution chain convergence process (node1=red, node2=cyan, node3=magenta, node4=yellow, node5=blue, node6=green, node7=red, node8=cyan,
node9=magenta, node10=yellow)
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 72/76
Figure 39 – Steady state test: distribution chain steady-state performance (node1=red, node2=cyan, node3=magenta, node4=yellow, node5=blue, node6=green, node7=red, node8=cyan,
node9=magenta, node10=yellow)
Next, ring restoration was simulated based on the setup presented in
Figure 40.
Figure 40 – Ring restoration test setup
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 73/76
This time a second PRTC, with the same TOD setting as the first PRTC, was connected to node10. The
QL of the second PRTC was configured to be SSU-A (‘0100’ or 4) while the QL of the first PRTC was
initially configured to be PRC (‘0010’ or 2). After 500 seconds of simulation runtime, the QL of the
first PRTC was degraded to SSU-B (‘1000’ or 8). The obtained results are presented in Figure 41. The
QL change group of events as a result of the ring restoration is clearly evident. Zooming-in on this
group of events, the specific nature of the QL change across the entire distribution chain can tracked.
Once the QL of PRTC1 drops from PRC (2) to SSU-B (8), the new QL starts to propagate across the
distribution chain node after node. When it reaches the last node10, the clock selection process within
node10 selects PRTC2 with QL SSU-A (4) to be its clock source (marked by the yellow line climbing
from 2 to 4 at time=509 sec). From here, the restoration process begins where the new QL SSU-A
starts to propagates towards node1. The entire QL restoration process lasts ~18 seconds.
Figure 41 – Ring restoration process (node1=red, node2=cyan, node3=magenta, node4=yellow, node5=blue, node6=green, node7=red, node8=cyan, node9=magenta, node10=yellow)
The time-error performance during the restoration process is presented in Figure 42. No time error
transient of any sort is evident. This can be explained by the fact that the two PRTC were configured
for the exact same TOD settings. Should some small mismatch between the TOD setting of the two
PRTC exists, a small transient process might have occurred. Nevertheless, more clever hitless
mechanisms (of the sort that are extensively being used for frequency distribution) would easily solve
this problem on the price of longer holdover periods while in the restoration process. Nevertheless,
having a synchronized physical-layer (thus, a very accurate holdover frequency) this should not
become a major issue.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 74/76
Figure 42 – Steady state test: distribution chain steady-state performance (node1=red, node2=cyan, node3=magenta, node4=yellow, node5=blue, node6=green, node7=red, node8=cyan,
node9=magenta, node10=yellow)
8.6.6 Conclusions
Link-by-link TOD distribution can be achieved by extending the Sync-E ESMC messages to
distribute TOD information in addition to the SSM information its distributes today. The benefits of
this approach are:
• A “light” TOD distribution scheme that demand very low network resources (one message
per second on every link) but still able to give state-of-the-art level of performance.
• Based on existing clock selection algorithm. No need to re-invent the wheel.
• Saves the need for an additional higher-layer TOD distribution protocol such as PTPv2,
imposing further resources allocation on the network and requiring an additional
management plane and increasing the operator CAPEX and OPEX costs.
• Complete inter-work with the physical layer frequency distribution (the clock selection
process taking place at the node actually applies for both). No need for special discovery
mechanisms and complicated management considerations.
• Can be easily extended using any higher-layer standard time distribution protocol (e.g.
PTPv2 or NTP). In order to do so, one only needs to implement a special “boundary clock”
comprising an ESMC TOD recovery function (a slave device) feeding any higher-layer
master/server device. The opposite case can also be considered (PTPv2/NTP to Sync-E
TOD).
The suggested TOD distribution scheme that was described in this clause will need to be re-
visited after the current work, ongoing in the ITU-T SG15/Q13, to define a new time distribution
PTPv2 telecom profile will be completed. When that happens, the proposed scheme could be re-
evaluated in order to learn it can fully comply with the requirements of the new telecom profile.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 75/76
Comprehensive comparisons to existing higher-later TOD distribution scheme could then be
conducted again to learn about its true value.
Synchronization service in CGE environment
TIGER2 – Together IP, GMPLS and Ethernet Reconsidered, Phase 2 Page 76/76
9 Conclusion
The field of frequency and time distribution over packet switched networks is rapidly moving
forward with new improved solutions to meet the stringent requirement of new end-applications
constantly appearing. The rapid, ever growing demand for new services together with the strong
competitiveness between service-provides and their desire to increase their value offering by
introducing more and more cutting-edge technologies, result in a technology migration process that is
not always smooth and bump less.
The result is quite a wide spectrum of new offered technologies where each one it tuned to solve
a specific problem in a very good way. Nevertheless, these technologies do not often interact or co-
exist in a smooth way to give the service provider a global optimized solution for his network.
This report summarizes the work carried over the past 1.5 years with the attempt to define a
portfolio of tools and solutions that will enable a network designer/service provider to make better use
and planning of the existing technologies he got. Whether these are new mechanisms to better
interwork PTPv2, NTP and Sync-E, new insights to how QoS can be efficiently employed to improve
overall end-to-end performance, new ideas to enable accurate time and frequency distribution over an
IPSec tunnel or a new apparatus to extend a widely-spread well known frequency distribution
technique to additionally support time distribution, the common denominator that binds them all
within this framework is the attempt to bridge across these multitude of new technologies perhaps
allowing for a smoother more orderly migration from one to the other.
A few years from now, networks will offer full support, allowing distributing end-to-end
frequency and time with formidable accuracies to service even the most demanding application.
Technology wise, the industry will probably converge into only a few winning technologies that will
easily interwork and bring great value to the service provides. Nevertheless, we are not there yet and
the work included in this report aims at making this journey an easier one.
* * *