CH A RTI N G TH E FUTU R E O F I N N OVATI O N | VO LU M E 103 I 2021–01

ERICSSON
TECHNOLOGY

VERSATILE
VIDEO CODING
EXPLAINED

ARTIFICIAL
INTELLIGENCE
IN RAN

APPLIED
NETWORK SLICING
SCENARIOS
 CONTENTS ✱

08 MONITORING IoT APPLICATION PERFORMANCE WITH MACHINE QoE

Awareness of the QoE that the network delivers is essential for all mission-critical
Internet of Things (IoT) systems. To reliably predict QoE in IoT applications, we have
developed a generic Machine QoE framework and tested a vertical-specific
08
version of it in a smart-grid scenario.

18 VERSATILE VIDEO CODING EXPLAINED

– THE FUTURE OF VIDEO IN A 5G WORLD
Business
Versatile Video Coding – the new video compression coding standard that Business intent
was recently approved by MPEG and the ITU-T – represents state-of-the art operations

18
Goals, SLA, behavior Visualiz
video coding and is certain to play an important role in supporting a wide
range of 5G use cases.

30 COGNITIVE PROCESSES FOR ADAPTIVE INTENT-BASED Knowledge:

NETWORKING Descriptions of objects and
relations between them

30
Extensive automation will be necessary to cope with the unprecedented flexibilityCognitive Dynamic
and dynamic adaptation that 5G networks are introducing into service delivery and knowledge
layer update
network resource utilization. Our intent-based networking approach combines
several artificial intelligence (AI) techniques within a flexible architecture and uses Data:
Efficient storage and Data and
intents to specify what the autonomous system is expected to do. retrieval of objects streams Machine learning and m

42 ARTIFICIAL INTELLIGENCE (AI) IN RAN – A SOFTWARE

FRAMEWORK FOR AI-DRIVEN RAN AUTOMATION Observations
When used correctly, AI and machine learning techniques have
Network and
tremendous potential to help overcome complex cross-domain environment
Radio Core CEM
automation challenges in radio networks. Ericsson has identified the
key software enablers for AI-based RAN automation and integrated
them into a comprehensive framework that provides a solid and flexible
technological foundation.
42 A Tprop_B1 B
Tprop_A1

52 5G SYNCHRONIZATION REQUIREMENTS AND SOLUTIONS UE1

Tprop_B2
Achieving proper network synchronization in 5G requires more planning than it did D E
Tprop_A2

in previous generations due to the fact that there is no single synchronization Tprop_UE1 = Tprop_A1 - Tprop_B1
Tprop_A3
requirement in 5G. A solid understanding of 5G timing needs and their associated Tprop_UE2
= Tprop_A2 - Tprop_B2
solutions is essential for an efficient and cost-effective implementation.

52
Tprop_D4 Tprop_E4 UE2
Tprop_UE3
= Tprop_A3 - Tprop_C3
Tprop_C3
Tprop_UE4
= Tprop_D4 - Tprop_E4
UE3
64 APPLIED NETWORK SLICING SCENARIOS IN 5G UE4 C
Campus Radio-access Campus
Many new and emerging 5G use cases require network slicing, an outcome that site is network On-premises DC WAN transport
Tprop_UE1 and
Central DC
Tprop_UE4 < Tprop_UE2 < Tprop_UE3
Hybrid private 5G networks for enterprises
achieved through the use of an evolving toolbox of enablers in five areas: cloud A. Network slice for
RAN RAN

infrastructure, RAN, core, transport and operations support systems/business critical services
capabilities/
NFs
NF 5G
NFs Appl.
Mgm
Appl.

support systems. Depending on the scenario, different combinations of RAN

B. Network slice for RAN

enablers are needed to engineer the appropriate network slice(s). capabilities/

latency-sensitive
NFs
services
NF 5G
NFs Appl.
5G
NFs

C. Network slice for

64
RAN
capabilities/ non-critical services 5GC/EPC
NFs NFs

RAN D/E. Network slices for other Transport

capabilities/ (wide-area/limited-area) services partition 5GC/EPC
NFs NFs

Wide-area RAN

#01 2021 ✱ ERICSSON TECHNOLOGY REVIEW 5

 ✱ EDITORIAL EDITORIAL ✱

Ericsson Technology Review brings you

insights into some of the key emerging
innovations that are shaping the
future of ICT. Our aim is to encourage
an open discussion about the potential,
practicalities, and benefits of a wide range
of technical developments, and provide
insight into what the future has to offer.
BUILDING BACK
address
Ericsson
BETTER ON THE
SE -164 83 Stockholm, Sweden
Phone: +46 8 719 00 00 NETWORK PLATFORM
publishing
All material and articles are published on the
Ericsson Technology Review website:
www.ericsson.com/ericsson-technology-review ■ the unprecedented events of 2020 have …SOLUTIONS THAT (IoT) systems in terms of QoE. To ensure a
provided us all with ample evidence that the consistently high level of customer satisfaction in
publisher technology and solutions Ericsson delivers in ARE SMARTER, FASTER AND these types of systems, we have created a Machine
Erik Ekudden partnership with our customers are fundamental to
the functioning and future development of both
MORE EFFICIENT THAN EVER QoE framework that is specifically designed to
measure the QoE of IoT devices.
editor
Tanis Bestland (Nordic Morning)
industry and society as a whole. BEFORE
We hope you enjoy this issue of our magazine
editorial board As governments and enterprises prepare to “build While autonomous operations are going to be of and we’d be delighted if you shared it with your
Håkan Andersson, Magnus Buhrgard, back better” post-pandemic, we at Ericsson are great value in a broad range of different applications colleagues and business partners. You can find
Dan Fahrman, John Fornehed, Kjell Gustafsson, excited about using our technology platform to and scenarios, the industrial automation use case is both PDF and HTML versions of all the articles at:
Jonas Högberg, Johan Lundsjö, help communication service providers (CSPs) www.ericsson.com/ericsson-technology-review
Mats Norin, Håkan Olofsson, Patrik Roseen,
definitely one of the most rigorous due to its
Anders Rosengren, Robert Skog, develop innovative solutions that are smarter, stringent timing requirements. The 5G
Gunnar Thrysin and Sara Kullman faster and more efficient than ever before. The synchronization article on page 52 explains the main
ability to leverage artificial intelligence (AI) categories of 5G synchronization requirements and
f e at u r e a r t i c l e techniques and automate as much as possible will presents a toolbox of solutions for all relevant
Future network trends: be key to our mutual success.
Creating intelligent digital infrastructure
scenarios, the most challenging of which can often
by Erik Ekudden be addressed within the RAN domain.
While telecom AI is still very much a work in progress,
art director our research indicates that it is already possible to reach Private 5G networks created with the help of
Liselotte Stjernberg (Nordic Morning) a high degree of practical autonomous operation in network slicing techniques are likely to become
networks by combining existing AI techniques within a very popular with both private enterprises and
project manager
flexible architecture to form what we at Ericsson call a public agencies in the years ahead. The article on
Susanna O’Grady (Nordic Morning)
cognitive layer. The article on page 30 highlights the page 64 highlights the benefits of these types of
l ay o u t critical role intents play in creating this new layer and networks and explains how CSPs can use an
Liselotte Stjernberg (Nordic Morning) enabling autonomous operations. evolving set of enablers in cloud infrastructure,
the RAN, the core and transport networks, and
i l lu s t r at i o n s Building the next generation of autonomous OSS/BSS to create appropriate slices for each
Jenny Andersén (Nordic Morning) systems will also require AI-powered automation of use case.
subeditors
the RAN. The article on page 42 presents a
Ian Nicholson (Nordic Morning) comprehensive framework of the key software The final two articles in this issue of the magazine ERIK EKUDDEN
Paul Eade (Nordic Morning) enablers of AI-based RAN automation, along with explore recent breakthroughs in video compression SENIOR VICE PRESIDENT,
two proof-of-concept examples executed in real technology and take a closer look at the CHIEF TECHNOLOGY OFFICER AND
is s n : 0 014 - 0171
RAN contexts. requirements of mission-critical Internet of Things HEAD OF GROUP FUNCTION TECHNOLOGY
Volume: 103, 2021

ERICSSON TECH NOLOGY REVIEW ✱ #01 2021 #01 2021 ✱ ERICSSON TECHNOLOGY REVIEW 7
✱ MACHINE QOE IN THE IOT MACHINE QOE IN THE IOT ✱

Monitoring IoT or similar performance requirements. The model

features can then be optimized for each CSP
according to the specific requirements of their
network environment.
URLLC provides the ultra-reliability or ultra-low
latency required for numerous uses cases in
smart grid, intelligent transportation systems
and smart manufacturing, while mMTC enables

application performance
connectivity to lots of low-power and low-cost
Classification of IOT applications devices for best-effort applications and long-range
There are several ways to classify IOT applications. coverage in smart cities or smart logistics.
One approach is to classify them according to At Ericsson, we have defined four categories
business verticals such as health care, agriculture, of IOT applications: Massive IOT, Critical IOT,
WITH MACHINE QOE automotive, transportation, surveillance, smart Broadband IOT and Industrial Automation IOT.
home, smart city, smart grid and smart metering. The first two correspond to the 3GPP mMTC
Each vertical comprises applications with various and URLLC categories respectively.
degrees of performance requirements ranging Broadband IOT provides the high-data rates
from very low to high data speed, from best and large data volume required for unmanned
New capabilities in 5G make it ideal to provide cost-effective solutions for effort to ultra-reliable, and from no latency aerial vehicles/drones and augmented reality/
mission-critical Internet of Things systems that depend on ultra-reliability to ultra-low latency. virtual reality applications. Industrial IOT
From a network perspective, applications are provides the precise indoor positioning and
and low latency. A necessary component when creating these solutions often classified into three categories based on their time-sensitive networking for real-time
is an awareness of the QoE that the network offers the IoT system. network performance requirements: massive IOT, applications that require deterministic
mission-critical control and enhanced mobile communication.
broadband (eMBB) [1]. Massive IOT requirements The M-QOE framework that we have designed
include deep coverage, ultra-low device energy generalizes on the four IOT categories by inferring
consumption, ultra-low complexity and ultra-high from specific cases. For this purpose, we first
density. Mission-critical control requirements investigated the important characteristics of IOT
include strong security, ultra-high reliability, applications in three business verticals: utilities,
CONSTANT WET TE
The Internet of Things (I OT) is a worldwide For communications service providers (CSPs), ultra-low latency and extreme user mobility. automotive and smart cities [2].
TCHOUATI, network of physical objects – buildings, cars, a typical IOT subscriber is an enterprise or a service The eMBB requirements include extreme
STEVEN ROCHEFORT, wearables, industrial machines and so on – provider from any vertical that wants to enable capacity, extreme data rates and deep awareness KPIs for machines
GEORGE SARMONIKAS that are equipped with connectivity devices connectivity in many devices spread in one region (for discovery and optimization). Our framework to assess the QOE of IOT subscribers
to build a communication network where or worldwide. Given the specificity and variability The 3GPP specifications define three main comprises a list of common IOT KPIs as shown in
connected objects can exchange data with of IOT requirements per vertical, CSPs need a new categories of applications, two of which are Table 1. These KPIs are calculated on a per device,
other objects. approach in the assessment of IOT subscribers’ QOE. dedicated to IOT, namely ultra-reliable and per subscriber and per vertical basis. The threshold
To meet this need, Ericsson has developed a generic low-latency communications (URLLC) and values are based on the application type and the
■ Almost anything can be connected to anything in Machine QOE (M-QOE) framework for accurate massive Machine-Type Communications (mMTC). IOT vertical that the device belongs to.
the IOT, creating a complex network of connected prediction of IOT subscribers’ QOE and tested a
machines, from simple wireless tags to sensors or vertical-specific version of it in a smart-grid scenario.
actuators with capabilities to sense, communicate,
process data or control their environment. Creating a framework for Machine QOE
Terms and abbreviations
The potential to apply IOT technology is endless, IOT networks support multiple technologies,
stretching across all industry verticals, resulting standards, device types and application types AI – Artificial Intelligence | CSP – Communications Service Provider | E-KPIs – Expected Value KPIs |
in a wide variety of devices and communications with various performance requirements between eMBB – Enhanced Mobile Broadband | IOT – Internet of Things | KPI – Key Performance Indicator |
M-QOE – Machine QOE | ML – Machine Learning | mMTC – Massive Machine-Type Communications |
requirements on the underlying network. verticals and between applications of the same
O-KPIs – Observed KPIs | PDC – Phasor Data Concentrator | PMU – Phasor Measurement Unit |
Large-scale deployment of cellular IOT devices vertical. A preliminary step to design an M-QOE
URLLC – Ultra-Reliable and Low-Latency Communications | WAMS – Wide Area Measurement System
is expected in the years ahead with the widespread model for this complex environment is to create
introduction of 5G technology. classes of IOT applications of the same vertical

8
2 ERICSSE
ORNI CT S
ECSO
H N OT LEO
CGHY
N ORLEO
VGI EYWR✱
E VA
I EUW
G U✱S T#101 ,1 22002201 #
AU0 1G2U0S2T1 1 1✱
, 2 E0 R ✱S EORNI CT SE C
2 0I C S SOH N OT LE O
CGHY
N ORLEO
VGI EYWR E V I E W 9
3
✱ MACHINE QOE IN THE IOT MACHINE QOE IN THE IOT ✱

KPI Definition

latency (millisecond) the time it takes to transfer a given piece of information from the
Framework architecture for Machine QOE There can be a combination of RAN, edge and
moment it is transmitted by the source to the moment it is successfully The system architecture for the M-QOE framework centralized processing of the received metrics.
received at the destination has the same common base as Ericsson’s operational Stream processing at the edge is enabled when
support system for subscriber service assurance. low latency and real-time processing are required,
packet loss (ratio) the percentage of frames that should have been forwarded It is the foundation to provide not only consistent whereas processing takes place at the central data
by a network but were not IOT service assurance but also an enhanced center in cases where there is no low-latency
bit error ratio the number of bit errors divided by the total number of transferred bits experience for the users of the connected machines requirement. A great variety of interfaces allow for
during a studied time interval (IOT devices) across several touchpoints along raw data to be preprocessed prior to the generation
the use case journey. of any insights. Processed data is stored in a
energy efficiency (joule/byte) the energy consumed for the end-to-end transport of a byte Figure 1 shows the high-level, end-to-end distributed database to make it available for
architecture for our M-QOE framework. Machines further analysis.
security level of importance for attack prevention
of different kinds are wirelessly connected over 5G A rules engine and machine learning (ML)
data (messages) rate, a time-variable function that might be important to define some (or other cellular technologies). The yellow dots models are used to provide M-QOE scores along
downlink/uplink (bit/s) parameters (such as peak, burst, average, minimum, maximum) indicate the points where the assurance metrics with other insights such as anomalies, patterns of
in order to better describe the data rate (messages) related to M-QOE are usually measured: use, mobility patterns and other use-case-specific
❭ the IOT device insights. The cognitive reasoning system, driven by
jitter (ms) the short-term variations of a digital signal’s significant instants from business intents such as the proactive service level
their ideal positions in time ❭ the RAN
assurance of smart-grid infrastructure, can use the
❭ the edge network
packet delay variation (ms) variation in latency as measured in the variability over time of the insights for autonomic decision-making. It can
❭ the core network/services network
packet latency across a network (expressed as an average of the continuously maintain autonomously the desired
❭ vertical slices and service layers (not shown).
deviation from the network mean latency) service level specification by recommending actions
The metrics are then ingested and processed at to the service/operations engineers and/or
reliability percentage of sent network layer packets successfully delivered to a different points of the network subject to the triggering closed loop actions with minimal
given node within the time constraint required by the targeted service, processing delay requirements of each use case. human intervention.
divided by the total number of sent network layer packets

availability the percentage of available time (with respect to total time) in a generic
observation period of the connection across the transport network M-QoE architecture
IoT operations center
mobility (km/h) fixed (no mobility: office, home) or maximum speed in movement 5G
(pedestrian or using a means of transportation such as a train,
road vehicle, airplane or drone)

Session processing
Data Batch Knowledge

KPI generation
Data ingestion

Rules engine
repository processing base & policies
traffic density (Mbit/s/sq km) traffic in a specific area 5G

AI & ML
Dashboards

connection density (devices/sq km) number of devices in a specific area

Stream Cognitive
processing reasoning
coverage (sq km) area of application interest 5G
Decisions & actions
battery lifetime (days, years) time of battery duration

data size (bytes) size of the atomic packet or frame (average, maximum) IoT Radio Edge/deep edge Core network/ Decision-making layer
services network
KPI measurement point
location accuracy the maximum positioning error tolerated by the application
(can be measured outdoor and indoor in 5G)
Devices KPI measurement points/analytics Rules/scores/insights Decisions/actions
cost overall cost including connection, device and cloud services

Table 1 List of common IoT KPIs Figure 1 M-QOE framework architecture

10
4 ERICSSE
ORNI CT S
ECSO
H N OT LEO
CGHY
N ORLEO
VGI EYWR✱
E VA
I EUW
G U✱S T#101 ,1 22002201 #
AU0 1G2U0S2T1 1 1✱
, 2 E0 R ✱S EORNI CT SE C
2 0I C S SOH N OT LE O
CGHY
N ORLEO
VGI EYWR E V I E W 11
5
✱ MACHINE QOE IN THE IOT MACHINE QOE IN THE IOT ✱

The computation of an IOT subscriber’s O-KPI

is done by aggregating the measurements of their
ELECTRIC POWER GRIDS
Machine learning Vertical KPI devices plus measurements at the vertical node over ARE ONE OF THE VERTICALS
different dimensions: temporal, spatial, devices and
Device traffic
Machine-type
detection
Application-type
inference
Subscriber KPI services. Our aggregation procedure consists of
THAT WILL BENEFIT MOST
computing the following statistics: average, variance, FROM M-QOE
skewness, kurtosis, percentile (5, 25, 50, 75 and 95),
minimum, maximum and range. The M-QOE_i At the same time, electric power grids and
score is derived from the dataset of the difference distribution systems are also under strain due to the
Observed KPI Gaps Expected (O-KPI_i - E-KPI_i) between the two measurements growing introduction of non-centralized renewable
values (M-QoE quantification) KPI values
and using an ML algorithm on a scale of 1 to 5 energy resources, electric storage systems and
labelled as (1) bad, (2) poor, (3) fair, (4) good or increased electrical demand from new sources such
(5) excellent. as electric vehicles. The evolving needs of power
The overall M-QOE value is M-QOE = Σ[alpha_i* grids are driving distribution systems to be more
Induced subscriber/vertical
M-QoE prediction M-QOE_i] where alpha_i represents the importance dynamic, with support for bidirectional power flows
KPI gaps
level of feature i. Mobile operators can optimize the compared with the traditional centrally managed
performance of this M-QOE model by selecting a systems [4].
subset of KPIs that are relevant to their network All of these factors introduce new challenges
Figure 2 The architecture of ML models in M-QOE
environment. in the operation, planning, protection and control
The purpose of the M-QOE prediction model is to of future power grids. The growing need for new
All generated insights, recommendations and fragmented with many device models and overcome the challenge of CSPs not being able to solutions is being met by the introduction of smart
actions are exposed through dashboards to various service providers. probe the IOT devices or the vertical’s network grids enabled by advanced communications
business and operational level management systems Our M-QOE framework also includes an infrastructure to access measurements of subscriber networks including smart devices, wireless
including IOT service operations, and usually used ML model for IOT segmentation, which was KPIs and vertical KPIs, as these are external connectivity, automation and real-time control,
for planning, management and decision-making trained with samples of traffic data of known environments. These measurements are usually edge computing and analytics.
purposes [3]. IOT applications. This model can be used in obtained from the MOS (mean opinion score) test, To improve power quality, many electrical grids
production to detect traffic patterns and predict but this is done only once, limiting the dynamic are being upgraded with synchrophasor technology.
Machine QOE models the type of IOT application and vertical. This is tracking of user satisfaction. To address these With this in mind, we decided to test our proposed
To improve the performance of our M-QOE framework, valuable, as the IOT ecosystem is fueled by a multitude limitations, we trained an ML algorithm to classify M-QOE framework by applying it in a smart-grid
we added some rule-based and ML models to detect of small and medium vertical service providers and predict the unknown E-KPI measurements scenario to predict the M-QOE of devices called
insights or anomalies in KPI measurements that whose traffic patterns have yet to be discovered. from the known O-KPI factors records. The M-QOE phasor measurement units (PMUs) in a cellular-
operators could use to improve IOT subscribers’ The M-QOE quantification in our framework prediction model uses an inductive supervised based synchrophasor.
network experience, as shown in Figure 2. is done by computing the gap between the observed learning approach.
Our M-QOE framework includes four models: KPIs (O-KPIs) and their expected values (E-KPIs) Synchrophasors
defined by the Service Level Agreement. It should Case study: Machine QOE in next-generation A synchrophasor is a time-synchronized
❭ machine-type detection
be noted that a typical IOT subscriber owns multiple power grids measurement of a quantity described by a phasor
❭ application-type inference
devices, and their overall satisfaction with a service Electric power grids are one of the verticals that will [5, 6]. Like a vector, a phasor has magnitude and
❭ M-QOE quantification
is a function of multiple features that are monitored benefit most from M-QOE in the near term. The phase information. PMUs in a traditional PMU
❭ M-QOE prediction.
by the different O-KPIs. To compute the overall business model in this vertical is undergoing a network are deployed across transmission grids
CSPs can use the machine detection model in our M-QOE of the IOT subscriber, the first step is to transformation from mega grids operated by one to measure voltage and current, and with these
M-QOE framework to detect if a device accessing aggregate the O-KPI measurements of all their large company to segmented grids and micro grids, measurements calculate parameters such as
their network is an IOT device or a mobile phone to devices and derive a feature-specific M-QOE_i giving consumers the possibility to choose between frequency and phase angle.
better manage their QOE. This is helpful because, for feature i. The second step is to add up the several offerings from different suppliers. In this PMU measurements are time-stamped to an
unlike mobile phones, there is no global database feature-specific M-QOE_i weighted by their competitive market, consumer QOE becomes the accuracy of a microsecond, synchronized using
of IOT devices, and the IOT landscape is highly importance factor to obtain the overall M-QOE. main differentiator. the timing signal available from GPS satellites

12
6 ERICSSE
ORNI CT S
ECSO
H N OT LEO
CGHY
N ORLEO
VGI EYWR✱
E VA
I EUW
G U✱S T#101 ,1 22002201 #
AU0 1G2U0S2T1 1 1✱
, 2 E0 R ✱S EORNI CT SE C
2 0I C S SOH N OT LE O
CGHY
N ORLEO
VGI EYWR E V I E W 13
7
✱ MACHINE QOE IN THE IOT MACHINE QOE IN THE IOT ✱

or other equivalent time sources. Measurements

taken by PMUs in different locations are therefore
A µPMU IS A SMALLER, Application Accuracy
Availability/
Low latency Message rate
reliability
accurately synchronized with each other and can be LOWER-COST AND MORE
Automation
time-aligned, allowing the relative phase angles
between different points in the system to be
ACCURATE VERSION 4 4 4 4

determined as directly-measured quantities. OF A PMU Reliability 2 2 3 2

Synchrophasor measurements can thus be
combined to provide a precise and comprehensive IOT Connectivity as a Service (CaaS) is a Planning 4 3 1 4
view of an entire interconnection. subscription-based solution offered by operators
A typical synchrophasor network is composed of that utilities can purchase to manage the Operation
PMUs and phasor data concentrators (PDCs) that connectivity of their PMUs and μPMUs. 1 1 2 2
are the control centers of the wide area measurement Private LTE is another cost-effective alternative
Table 2 Classification of PMU applications in power systems
system (WAMS). The communication systems play for grid operators to build synchrophasors over a
a crucial role in the performance of the WAMS. dedicated communication system managed by the
Requirements for data reporting rates are typically mobile operator to meet the desired requirements Operations is associated with the monitoring and in monitoring and predicting the M-QOE of an
30 to 60 records per second, and may be higher; in of latency, reliability and coverage. visualization of the network performance. enterprise with widespread devices over their
contrast, current SCADA (supervisory control and A CSP supplying network connectivity services to infrastructure without the need for expensive
data acquisition) systems often report data every four Synchrophasors – Machine QOE assessment a smart-grid operator can use the proposed framework large-scale measurements.
to six seconds – over 100 times slower than PMUs. According to a recent report from the North for the assessment of customer M-QOE by In another experiment, we used the M-QOE
Multiple communications technologies are used American SynchroPhasor Initiative (NASPI) [4], calculating M-QOE = Σ[alpha_i* M-QOE_i] where: framework for the measurement of packets lost and
in synchrophasors, from power-line communications PMU applications in a power system can be ❭ i represents the relevant features of delay gaps in the transmission of synchrophasor data
to microwave, but in order to meet the latency classified into four main categories based on due to communication link failures. The results
synchrophasor applications (accuracy,
requirements, optical fiber is widely used. Traditional their reliability ranges: showed early prediction of failures and M-QOE
reliability, latency, message rate and security).
optical fiber solutions may become prohibitively automation (latency of ≤ 10-100ms) as well as accurate detection of failure types [7].
❭ ❭ alpha_i represents the importance level
expensive, though, as smart grids become more
❭ reliability (latency of ≤ 1,000ms) coefficients of the KPIs derived from the Conclusion
complex and the need for additional measurements
❭ planning (latency of ≤ 100-1,000ms) number in Table 2 between 1 and 4.
in distribution grids grows. In light of this, work is The deployment of 5G networks will accelerate
❭ operation (latency of ≤ 1,000ms).
underway to develop a new, micro-PMU (μPMU). We tested this M-QOE framework in an experimental the growth of the Internet of Things (IOT) and
A μPMU is a smaller, lower-cost and more accurate The report also lists the relevant KPIs – accuracy, 5G lab environment, where we built a synchrophasor offer a wealth of new business opportunities for
version of a PMU that can be deployed more reliability, latency, and message rate – and ranks them network and connected between one and ten μPMUs communications service providers (CSPs).
extensively in a distribution grid and provide per category on an importance-level scale of 1-4 in to a PDC, depending on the test scenario. Probes It is important to recognize, however, that the
accurate information from more places, allowing which 4 means critical, 3 means important, 2 means were designed and installed in the communication churn of a single IOT enterprise due to network
for better dynamic distribution grid management. somewhat important and 1 means less important. network, in the μPMUs and in the PDC to collect and issues may represent many device connections
Table 2 presents the key details of the NASPI’s stream network and device O-KPI_i measurements with significant revenue impact. To ensure
Cellular-based synchrophasors classification of PMU applications in power systems. associated with the transmission of payload data consistent QOE in IOT applications, CSPs need a
The cost of extending the WAMS to connect more Automation refers to the automated protection between μPMUs and the PDC. smart solution to consistently monitor the diversity
devices can become an issue when deploying and control applications of PMUs in distribution To obtain enough O-KPI measurements, of performance requirements that are requested by
μPMUs at scale in distributed grids. 5G with systems. The aim of these applications is to improve we identified the locations of potential μPMU IOT subscribers. Ericsson’s Machine QOE
network slicing can meet the connectivity the reliability and security, automated remedial devices in the grid network and used the M-QOE framework is designed with this in mind, utilizing
requirements of a WAMS and reduce the cost action schemes, and asset utilization. Reliability is a prediction model to generate the measurement data. the power of artificial intelligence and machine
of deploying μPMUs. Using low-band and mid-band class of PMU applications that can be divided into This data was aggregated in the knowledge-based learning to automatically discover and predict
radio frequencies, the data bandwidth requirements topology and disturbance detection and situational system running the rules-based and ML models for events in the CSP network, so that they can be
of PMU applications can be achieved, with latency awareness. Planning is a class of PMU applications continuous evaluation and monitoring of the overall addressed before they have a negative impact
values in access networks (radio access to network focused on better system understanding and im- M-QOE of the grid operator. The experiment on QOE.
edge) of 10ms. proved system modeling of the distribution network. confirmed that the framework will assist CSPs

14
8 ERICSSE
ORNI CT S
ECSO
H N OT LEO
CGHY
N ORLEO
VGI EYWR✱
E VA
I EUW
G U✱S T#101 ,1 22002201 #
AU0 1G2U0S2T1 1 1✱
, 2 E0 R ✱S EORNI CT SE C
2 0I C S SOH N OT LE O
CGHY
N ORLEO
VGI EYWR E V I E W 15
9
✱ MACHINE QOE IN THE IOT MACHINE QOE IN THE IOT ✱

Steven Rochefort
◆ joined Ericsson in 1994

the authOrs
with a background in
References software development
for command and control
1. Qualcomm, Making 5G NR a reality, December 2016, available at: https://www.qualcomm.com/media/
systems. Rochefort has
documents/files/whitepaper-making-5g-nr-a-reality.pdf
been involved in almost
2. Elsevier, Pervasive and Mobile Computing, vol. 48, pp. 59–68, 2018, Traffic characterization and every aspect of mobile
LTE performance analysis for M2M communications in smart cities, Malandra, F; Chiquette, L.O; telephone development
Lafontaine-Bédard, L.-P; Sansò, B, available at: https://www.sciencedirect.com/science/article/abs/pii/ Constant Wette at Ericsson, from software George Sarmonikas
S1574119217306089 Tchouati development to system ◆ joined Ericsson in 2013
3. Ericsson Technology Review, Generating actionable insights from customer experience awareness, ◆ joined Ericsson in 2000 design, with a focus on after several years of
September 30, 2016, Niemöller, J; Washington, N; Sarmonikas, G, available at: https://www.ericsson.com/ and has since worked as a IoT solutions. In 2016, working for mobile
en/reports-and-papers/ericsson-technology-review/articles/generating-actionable-insights-from-customer- developer and project leader he returned to his academic operators. He currently
experience-awareness in product development, passion for mathematics, heads AI & IoT Solutions
4. Energies 2019 vol 12, A Comprehensive Survey on Phasor Measurement Unit Applications in Distribution research projects, innovation becoming a data scientist within Business Unit Digital
Systems, November 29, 2019, Hojabri, M; Dersch, U; Papaemmanouil, A; Bosshart, P, available at: https:// and new business for Ericsson’s OSS Services at Ericsson,
doi.org/10.3390/en12234552 development involving (Operations Support developing novel products.
5. North American SynchroPhasor Initiative (NAPSI), Synchrophasor Technology Fact Sheet, October various technologies, Systems) product area. Prior to this, he was
2014, available at: https://www.naspi.org/sites/default/files/reference_documents/33.pdf?fileID=1326 data science, IoT/machine- He holds a B.Sc. responsible for the product
The authors would
to-machine, IMS, and telco in mathematics and management of Ericsson’s
6. IEEE, Micro-synchrophasors for distribution systems, ISGT 2014, May 19, 2014, von Meier, A; Culler, D; like to thank
cloud architectures. a graduate certificate customer experience
McEachern, A; Arghandeh, R, available at: https://ieeexplore.ieee.org/document/6816509 Brunilde Sansò,
He holds an M.Eng. in marketing research management and analytics
7. IEEE, A New Approach to Reliability Assessment and Improvement of Synchrophasor Communications in Younes Seyedi,
in electrical engineering from Concordia University portfolio, including assets for
Smart Grids, May 12, 2020, Seyedi Y; Karimi H; Wetté C; Sansò B, available at: https://ieeexplore.ieee.org/ Orestes Gonzalo
from the National Advanced in Montreal, Canada. subjective experience
document/9091616 Manzanilla-Salazar,
School of Engineering in scoring. Sarmonikas holds
Hakim Mellah,
Filippo Malandra Yaoundé, Cameroon, an both an M.Sc. in
and Houshang M.Sc. in computer communication systems and
Karimi – all based engineering from the École an M.Eng. in electronic
Further reading at École Polytechnique de Montréal, engineering and computer
❭ Ericsson, Internet of Things, available at: https://www.ericsson.com/en/internet-of-things Polytechnique de Canada, an M.B.A. from HEC science from the University
Montréal – for their Montréal and a graduate of Bristol in the UK, as well as
❭ Ericsson, 5G, available at: https://www.ericsson.com/en/5g
contributions to certificate in data science a graduate degree in artificial
❭ Ericsson, Network slicing, available at: https://www.ericsson.com/en/digital-services/trending/network-slicing from Harvard University intelligence from Stanford
this article.
in the US. University in the US.

16
10 ERICSSE
ORNI CT S
ECSO
H N OT LEO
CGHY
N ORLEO
VGI EYWR✱
E VA
I EUW
G U✱S T#101 ,1 22002201 #
AU0 1G2U0S2T1 1 1✱
, 2 E0 R ✱S EORNI CT SE C
2 0I C S SOH N OT LE O
CGHY
N ORLEO
VGI EYWR E V I E W 17
11
✱ VERSATILE VIDEO CODING VERSATILE VIDEO CODING ✱

Versatile Video
These new services are expected to rely on HEVC has not been economically beneficial,
continued video evolution toward 8K resolutions since the reduction in bit rate of 40 percent has not
and beyond, and on new strict requirements such as compensated for the cost of replacing a large number
low end-to-end latency for video delivery. Since the of set top boxes. Even in some of these cases, the
data rate of such video-enabled services is extremely introduction of VVC may change this calculation

Coding explained
high, the cost would be prohibitive – even in the most since the bit rate reduction compared to the older
modern networks – unless the video was carried codec is so much greater. Figure 1 provides an
in a compressed format, with the help of a next overview of potential application areas.
generation video codec.
Versatile Video Coding (VVC), the new video Video compression options available today
compression coding standard that will be published A video codec, which can be implemented in
– THE FUTURE OF VIDEO IN A 5G WORLD as ISO/IEC 23090-3 and ITU-T recommendation hardware or software, encodes and/or decodes
H.266, offers the highest compression efficiency digital video. Modern video codecs can reduce the
available today and is therefore the codec that is bitrate of uncompressed video to less than one
Many new and emerging 5G use cases will soon require video best suited to offer a suitable performance level for percent of the original rate without any noticeable
new media services over 5G networks. It can also visual quality degradations. The majority of these
compression efficiency and functionality that are beyond the capabilities enhance the user experience for existing video video codecs have been standardized and published
of today’s leading video codecs. Versatile Video Coding – the new video services by delivering substantially higher quality
at the same bitrate.
by a standards development organization (SDO).
The most popular and widely deployed codecs
compression coding standard recently approved by the Moving Picture Alternatively, it can be used to reduce the bitrate include the MPEG (Moving Picture Experts Group)
Experts Group and the International Telecommunication Union – includes (roughly by half) while maintaining the same quality.
For some legacy applications, the cost of replacing an
series from ISO/IEC (International Organization
for Standardization/International Electrotechnical
both improved compression efficiency and new features to enhance older codec such as H.264 with a newer one such as Commission) and the corresponding H.26x series
support for immersive video and low-delay video coding.

8K HDR video
streaming
High-quality
RICKARD SJÖBERG, The latest Ericsson Mobility Report estimates ■ While the quality of video content depends on HD and 4K
JACOB STRÖM, that video accounted for 63 percent of the several characteristics such as high pixel bit depth, video streaming Enhancing Enabling
ŁUKASZ LITWIC, traffic in mobile networks during 2019 and high frame rate, wide color gamut (WCG) and high existing new Low-latency
that this share will increase to 76 percent by dynamic range (HDR), it is the resolution (the services services cloud gaming
KENNETH ANDERSSON Versatile
2025 [1]. Enabled by continuously improving number of pixels in a video picture) that is most 360-video
network performance, this growth is driven directly associated with the bandwidth required VR streaming Video
by the ever-increasing availability of video for transmission. Other key factors determining Coding
Immersive XR
content in streaming services and online the required bandwidth are related to the type and telepresence
apps, and by changing consumer behavior of the video content and the latency with which Telepresence
toward consumption of the content on mobile the content is delivered to the end user. and screen
devices. The evolution toward large screens At the same time, innovations in 5G networks sharing
with high resolutions beyond HD on mobile offer new opportunities for video-enabled services IoT and
devices further raises consumer expectations for both consumers (remotely rendered virtual/ automative
regarding the quality of delivered video extended reality and cloud gaming, for example)
content. Video codecs play a critical role and industries, particularly with respect to the
in coping with these expectations. Internet of Things (IoT) and the automotive sector. Figure 1 VVC application areas

18
2 E RTI ECCSH
ERICSSON SNOO
N LTOEG
CYH N
ROE VL IOEG
WY ✱R EOVCI ETW
O B✱
E R #1 40 ,1 22002201 #
O0C1T2O0B2E1 R ✱
1 4 ,E2R0I2C0S S✱O N
E RTI ECCSH
SNOO
N LTOEG
CYH N
ROE VL IOEG
WY R E V I E W 19
3
✱ VERSATILE VIDEO CODING VERSATILE VIDEO CODING ✱

from the ITU-T (ITU Telecommunication in 2010, released the VP8 video codec in 2008, operators, broadcasters, chipset vendors and for higher resolution video streams due to its
Standardization Sector), which were developed which then became a proposed royalty-free option academics. Throughout the process, all coding tools, which can operate on block sizes
jointly. The naming convention for these codecs is in 2013. It has mostly been used as an alternative documentation including technical contributions, of up to 128x128 pixels and with 64x64 sample size
MPEG-2/H.262, AVC/H.264 (Advanced Video to H.264 in WebRTC, a framework for real-time draft specification text, reference software and transforms. VVC can achieve a reduction in bitrate
Coding) and HEVC/H.265 (High Efficiency Video web communication. VP9, a successor to VP8, conformance bitstreams were made available of around 40 percent for existing HD and 4K
Coding). Other video codec options include was developed internally at Google and released as publicly. services deployed with HEVC, at the same
Google’s VP8 and VP9, and the Alliance for Open open source in late 2012. The use of VP9 for 4K Ericsson has been an active participant in video visual quality.
Media’s AV1. YouTube videos has led to TV manufacturers standardization for more than 20 years and was Figure 2 provides a performance comparison
incorporating VP9 decoding into virtually all 4K closely involved in the development of the VVC between VVC and four other video codecs.
VVC IS THE LATEST TV sets, thereby spreading support of VP9 standard. Throughout the process, we led several The comparison was done by Ericsson Research
to a substantial number of devices. of the core experiments, chaired ad-hoc working using various sources of information including
MEMBER OF A SUCCESSFUL groups and made significant contributions to the in-house testing [3, 4]. The figure shows the
FAMILY OF VIDEO CODECS AV1 development of the technology behind the video approximate relative bitrate requirement for each
In 2015, a group of companies including Amazon, codec, most notably in the areas of deblocking video codec in order to reach the same video picture
MPEG-2/H.262 Facebook, Google, Intel, Microsoft and Netflix filtering, reference picture management, low-delay quality for HD and 4K content, benchmarked
The MPEG-2 video codec was published in 1994 founded the Alliance for Open Media (AOMedia) video coding and optimized encoder configurations. against the HEVC video codec, which is normalized
and is still in wide use today in standard definition with the aim of developing a new royalty-free video We also participated in efforts that made an impact to 100 percent. VVC performs significantly better,
digital TV services. Its compression efficiency is low codec. The specification for that codec, AV1, was in other areas of the video codec such as intra requiring only 60 percent of the relative bitrate
when compared with more recent video codecs. published on January 8, 2019. Although patent and inter prediction and adaptive loop filtering. compared with both HEVC and AV1.
The main reason for using MPEG-2 today is to reviews were conducted during the development to On top of its superior compression performance,
support existing set-top-boxes, as the cost of avoid infringing on third-party intellectual property Key benefits of versatile video coding the versatility of VVC also makes it an attractive
replacing these may be higher than any potential rights, the patent licensing organization Sisvel As the latest and most sophisticated video codec to choice beyond mainstream 2D video services.
savings from a new video codec. announced in March 2019 that it would form a date, VVC offers the highest compression efficiency VVC has been designed to handle both traditional
patent pool for AV1. As a result, the royalty-free of all video codecs, and it is particularly appropriate camera-captured content as well as the increasingly
Advanced Video Coding status of AV1 is uncertain.
AVC/H.264 was published in 2003 and is currently
the most widely used video codec. It is supported by Versatile Video Coding – the development
practically all mobile devices, is heavily used for process Relative bitrates for the same video quality (lower is better)
video carried over the internet, and is the preferred Versatile Video Coding (H.266) was standardized
video codec for HDTV. in a joint effort by the Video Coding Experts
Group of the ITU-T and MPEG of the ISO/IEC and 180% 170%
High Efficiency Video Coding is therefore the latest member of a successful family 160%
HEVC/H.265 [2] is the successor to H.264 and was of video codecs that includes MPEG-2, 140%
published in 2013. Compared with H.264, it delivers H.264 and HEVC. In contrast to proprietary 120%
120%
the same visual quality at roughly 40 percent lower alternatives, all four of these video coding standards 100% 100%
bitrate. There is widespread support for HEVC were developed in an open and collaborative
100%
across TVs and mobile devices. HEVC has been fashion with agreed requirements and timelines. 80%
60%
selected by the DVB (Digital Video Broadcasting) The development of VVC followed a 60%
and ATSC (Advanced Television Systems well-established standardization process, 40%
Committee) standards organizations for 4K which started with a technology exploration activity 20%
broadcast services and is recommended by 3GPP in 2015, included a formal call for proposals in 2018
for HD (HDR) and 4K mobile streaming. and concluded in July 2020 as a technically frozen
0%
standard. The development process involved H.264 VP9 HEVC AV1 VVC
VP8 and VP9 international stakeholders across the entire media
On2 Technologies, which was acquired by Google ecosystem: content producers, manufacturers, Figure 2 Relative bitrates for the same video quality

20
4 E RTI ECCSH
ERICSSON SNOO
N LTOEG
CYH N
ROE VL IOEG
WY ✱R EOVCI ETW
O B✱
E R #1 40 ,1 22002201 #
O0C1T2O0B2E1 R ✱
1 4 ,E2R0I2C0S S✱O N
E RTI ECCSH
SNOO
N LTOEG
CYH N
ROE VL IOEG
WY R E V I E W 21
5
✱ VERSATILE VIDEO CODING VERSATILE VIDEO CODING ✱

prevalent computer-generated imagery used in and designed to be easy to extract and merge with of tuning in at GDR positions. However, during such an architecture does not code pixel values
applications such as online gaming, e-sports video other subpictures. Previous video codecs provided standardization of VVC, an Ericsson proposal to directly but instead predicts and codes only error
streaming and screen-sharing applications. similar functionality using motion-constrained tile add a GDR picture-type indicator to the standard information to compensate for the inaccurate
sets, but subpictures are both much more efficient was adopted. This makes GDR support mandatory prediction. The process runs on blocks of pixels
Immersive video and easier to manage. in VVC, which ensures decoder interoperability rather than the entire picture, since this makes it
In contrast to traditional two-dimensional video, The efficiency of the subpicture design comes for such low-delay applications. possible to adjust to local video picture
where one particular view is captured by a camera, from built-in adjustments of low-level coding tools, Besides the GDR feature, VVC also inherited characteristics, minimizing the prediction error.
immersive video is video in which every angle is where previous designs require the encoders to be decoding unit (DU) based decoder operation from For video codecs based on such an architecture,
recorded. During playback, which at present very restrictive. Previous designs also require its predecessor, HEVC. DUs enable an encoder to an efficient picture partitioning scheme is pivotal
typically occurs on a virtual reality (VR) headset, rewriting of substantial amounts of coded data – output a first part of a picture, in a specified to achieving high compression efficiency.
the user is not constrained to a particular view but a burden that is largely alleviated by the use of controlled manner, without requiring the entire In VVC, each picture is split into non-overlapping
can look around freely. When the viewer does not subpictures in VVC. The use of subpictures in VVC picture to first be encoded. DUs and mandatory squares called coding-tree units (CTUs). The largest
change position, this is called three degrees of significantly reduces the complexity of application GDR support make VVC the prime video codec CTU size allowed in VVC is 128x128 pixels, larger
freedom (3DoF) immersive video, or more systems. This capability, together with its superior choice for low-delay video applications. than the 64x64 maximum size allowed in HEVC.
commonly 360-degree video. compression efficiency, makes VVC the best video Large blocks improve the efficiency of coding
360-degree video is typically stored in a codec choice for immersive video applications. How VVC works flat areas such as backgrounds, especially for
projection format. One such format is It is important to note that the superior coding high-resolution videos such as HD and 4K.
equirectangular projection, which is similar
to how the earth is pictured on a world map.
THE USE OF SUBPICTURES efficiency of VVC is not due to any single
compression tool. Rather, it is the result of combining
In order to efficiently represent highly detailed
areas such as textures and edges, VVC employs
A more popular format, however, is cube map IN VVC SIGNIFICANTLY many tools, each contributing with a small a flexible partitioning scheme that can partition
projection. Here, each face of the cube represents
one-sixth of the sphere area, and the six faces
REDUCES THE COMPLEXITY compression improvement. The most significant
of these tools are block partitioning, advanced inter
128x128-sized CTUs down to coding units (CUs)
as small as 4x4 pixels.
are arranged in one rectangular video picture. OF APPLICATION SYSTEMS prediction, dependent quantization, adaptive The scheme is based on two parts. The first is the
This enables the use of two-dimensional video loop filtering and improved deblocking filtering. quaternary tree (quad tree) split that is also available
codecs to handle the immersive content. Low-delay video coding The addition of sophisticated new coding tools in HEVC, which can recursively split CTU into
Due to the nature of the human visual system, Low-delay video coding is a key technology for has also had an impact on computational complexity. squared CUs down to 4x4 pixels, smaller than the
a high-quality representation of the full 360-degree certain time-critical video applications such as As a result of these developments, it is expected that 8x8 minimum CU size in HEVC. The second part
video sphere is only needed for the user’s current videoconferencing, cloud gaming and remote VVC decoder complexity will be around twice that consists of binary-tree and ternary-tree splits that
gaze direction. A well-known technique for control of road vehicles and drones. To reduce the of HEVC. partition a block into two and three rectangles
exploiting this property is to split the full 360-degree latency caused by excessive data buffering, the video respectively. Both binary and ternary tree splits
video content into multiple small rectangular regions encoder typically aims to generate as smooth a data Block partitioning can operate in either horizontal or vertical
and only convey and render regions in high quality rate as possible. One well-known video codec Similar to its predecessors, VVC uses a block-based directions, be recursively applied and mixed
that cover the area the user is currently looking at, feature is gradual decoding refresh (GDR), hybrid coding architecture. A video codec based on together in a nested multi-type tree.
called the viewport. To accommodate fast head which provides smooth tune-in points in the
motion, video is also transmitted for the bitstream without the need to encode single
non-viewport area, but the quality of that refresh pictures (intra pictures) that result in
video is much lower. bitrate spikes and high latencies. Instead, smooth Terms and abbreviations
To that end, and to help with other immersive tune-in points are generated by spreading the 3DOF – Three Degrees of Freedom | ALF – Adaptive Loop Filter | CTU – Coding-Tree Unit | CU – Coding
video applications such as remotely rendered VR, refresh across multiple pictures. Unit | DU – Decoding Unit | DVB – Digital Video Broadcasting | GDR – Gradual Decoding Refresh |
VVC introduces the concept of subpictures, which GDR has been supported in many older video HDR – High Dynamic Range | HEVC – High Efficiency Video Coding | IEC – International Electrotechnical
allows for the efficient extraction and merging coding standards such as H.264 and HEVC, but in Commission | IETF – Internet Engineering Task Force | ISO – International Organization for Standardization
of pieces of video of different quality that is a those cases the feature is conveyed in a supplemental | ITU– International Telecommunication Union | ITU-T – ITU Telecommunication Standardization Sector |
requirement for larger resolution immersive formats. enhancement information (SEI) message, which MPEG – Moving Picture Experts Group | SDO – Standards Development Organization | UHD – Ultra High
A subpicture is a rectangle within the full video makes GDR optional. This means that an encoder Definition | VR – Virtual Reality | VVC – Versatile Video Coding | XR – Extended Reality
picture that is fully independent of other subpictures cannot be certain that the decoders are capable

22
6 E RTI ECCSH
ERICSSON SNOO
N LTOEG
CYH N
ROE VL IOEG
WY ✱R EOVCI ETW
O B✱
E R #1 40 ,1 22002201 #
O0C1T2O0B2E1 R ✱
1 4 ,E2R0I2C0S S✱O N
E RTI ECCSH
SNOO
N LTOEG
CYH N
ROE VL IOEG
WY R E V I E W 23
7
✱ VERSATILE VIDEO CODING VERSATILE VIDEO CODING ✱

but also rotation and zoom, which can save a lot of To optimize the determination of quantization
bits for sequences containing such motion. levels within a block, sophisticated encoders
Another bit-saving technique in VVC is the use an algorithm called trellis quantization.
ability to vary the precision of the motion vectors. In VVC, dependent quantization has been
Relative bitrates for the same video quality (lower is better) For example, the encoder can signal to the decoder introduced, where the codec can switch between
that the incoming motion vectors are in integer or two shifted quantizers and thereby reduce the
four-times-integer resolution. This can save bits quantization error. In the VVC design, the
180% 170% when predicting smooth areas where the exact quantization levels for a given transform coefficient
sub-pixel precision does not give much image depend on the values of the preceding quantized
160% improvement over integer precision and also coefficients. To use this tool effectively, an encoder
140% when representing large motion vectors. therefore needs to evaluate how the determined
120%
120% The advanced inter prediction in VVC further quantization level for each coefficient impacts
100% 100%
100% exploits the fact that in some cases motion both the bit count and the total reconstruction
80% information does not have to be explicitly error for the whole block. In this way, the dependent
60% transmitted (as motion vectors) but can instead be quantization removes the inefficiency of quantizing
60% inferred from similarly moving parts of the picture. coefficients independently and thereby provides
40% For example, the bi-directional optical flow tool a substantial bit-reduction over HEVC.
20% in VVC can infer motion vectors by measuring
0% the optical flow in reference pictures and the DEPENDENT QUANTIZATION
H.264 VP9 HEVC AV1 VVC decoder side-motion vector refinement can infer
motion vectors by minimizing differences between REMOVES THE INEFFICIENCY
Figure 3 Comparison of the partitioning of a close-up section of a video picture in HEVC and VVC reference pictures. OF QUANTIZING COEFFICIENTS
Finally, the advanced inter prediction in VVC
also exploits motions of non-rectangular shapes INDEPENDENTLY
Figure 3 provides an example of the partitioning copy from – the horizontal and vertical displacement to better align with the shape of moving objects.
of a close-up section of a picture. HEVC block for a block – is stored in what is called a motion This means that one half of the block can have Adaptive loop filtering
partitioning (at top right) uses quaternary-tree split vector. Techniques that improve inter prediction are one motion vector and the other half of the block The adaptive loop filter (ALF) is a new in-loop filter
with coding blocks up to 64x64 pixels. VVC block responsible for a substantial part of the bitrate can have another with the halves separated by a in VVC. ALF scans the picture after it has been
partitioning (at bottom right) uses quaternary-tree reduction between VVC and HEVC. single geometrical line that is determined by an decoded and selectively applies (on a CTU basis)
split with coding blocks up to 128x128 pixels and While many techniques have been used to achieve angle and an offset. one of several two-dimensional finite impulse
nested multi-type tree split employing binary and the advanced inter prediction in VVC, five of them response filters before the picture becomes output
ternary tree splits. are particularly noteworthy: the affine motion Dependent quantization or is used for prediction. The video encoder
The block partitioning in VVC is highly flexible model, adaptive motion vector resolution, The quantizer is a core part of a video codec that calculates the sets of filter coefficients that will
and provides about 8 percent bitrate reduction over bi-directional optical flow, decoder side-motion is directly linked to operational control of video lead to the smallest error and transmits those to
HEVC [5]. However, this flexibility comes at a vector refinement and geometric partitioning mode. bitrate and visual quality. By adjusting the the video decoder. ALF has the ability to clean up
computational cost, especially on the encoder side, The precision of the motion vectors has also been quantization step, the encoder controls the fidelity artifacts in the picture and also contributes to a
where many more permutations need to be increased to 1/16 pixel compared to the quarter pixel of the error signal (transform coefficients), which is substantial bit-reduction over HEVC.
evaluated to select the optimal partition. resolution of HEVC. then coded with an entropy coder and sent in the
In previous video coding standards, it has been video bitstream. Improved deblocking filtering
Advanced inter prediction possible to compensate for translational motion; Previous standards such as HEVC have used VVC and its predecessors HEVC and H.264
One of the most efficient bit-saving techniques in that is, the decoder can be instructed to fetch sample a scalar quantizer but since the entropy coder are all block-based video codecs. The downside
video compression is inter-picture prediction (more values not from the same place in a previous image processes each transform coefficient in a coding of the block-based approach is that it can give rise
commonly referred to as inter prediction), which but rather from another position (1.75 pixels to the block in sequential order, there was an inherent to “block artifacts” – visible edges at some block
simply means copying sample values from previously left, for example). In VVC, the affine motion model inefficiency if the quantization level of each borders. Deblocking filtering is an approach to
coded pictures. The information about where to tool makes it possible to specify not only distance coefficient were to be determined independently. reduce these artifacts by selectively smoothing

24
8 E RTI ECCSH
ERICSSON SNOO
N LTOEG
CYH N
ROE VL IOEG
WY ✱R EOVCI ETW
O B✱
E R #1 40 ,1 22002201 #
O0C1T2O0B2E1 R ✱
1 4 ,E2R0I2C0S S✱O N
E RTI ECCSH
SNOO
N LTOEG
CYH N
ROE VL IOEG
WY R E V I E W 25
9
✱ VERSATILE VIDEO CODING VERSATILE VIDEO CODING ✱

across the block boundaries. The deblocking media transport specifications. Organizations
filtering in VVC is based on the HEVC deblocking such as 3GPP and DVB are investigating VVC
filtering, for which Ericsson was the leading in the context of next-generation services including
contributor. On top of this already strong base, 5G-enabled ones such as 8K (7680x4320 video)
VVC is capable of using longer deblocking filters, and 360-degree video VR streaming. Unlike its
where major parts were designed by Ericsson. predecessors, the first version of the VVC standard
The long deblocking filters allow for stronger includes support for a broad range of applications
deblocking that can be more effective in hiding across the media ecosystem, which is likely to have
block artifacts, especially for larger blocks a positive impact on the cost of deployment and
(128x128 sized blocks, for example) in relatively interoperability of VVC-based solutions and
smooth areas. The long deblocking filters contribute services.
significantly to the improved subjective quality In order to facilitate cross-industry discussion
of VVC compared to HEVC. around non-technical aspects of VVC deployment
such as licensing, marketing and interoperability
THE LONG DEBLOCKING activities, Ericsson and other industry leaders
launched the Media Coding Industry Forum
FILTERS CONTRIBUTE (MC-IF) in 2018. Since then, MC-IF has hosted a
SIGNIFICANTLY TO THE series of workshops and events to gather wider References
industry input on commercial aspects that may 1. Ericsson Mobility Report, June 2020, available at: https://www.ericsson.com/en/mobility-report/reports/
IMPROVED SUBJECTIVE further early VVC adoption. In particular, timely june-2020
availability of licensing terms for VVC was
QUALITY OF VVC COMPARED established as one of the key factors for VVC
2. Ericsson Technology Review, Next generation video compression, April 24, 2013, Fröjdh, P; Norkin,

TO HEVC deployment. To this end, shortly after the finalization

A; Sjöberg, R, available at: https://www.ericsson.com/en/reports-and-papers/ericsson-technology-review/
articles/next-generation-video-compression
of VVC development, MC-IF initiated fostering
What’s next for VVC? efforts for a patent pool program essential to VVC. 3. 2019 Picture Coding Symposium (PCS), Ningbo, China, 2019, pp. 1-5, Compression Performance of the
In today’s competitive video codec landscape, video Versatile Video Coding: HD and UHD Visual Quality Monitoring, Sidaty, N. et al., available at:
compression performance is key to successful Conclusion https://ieeexplore.ieee.org/document/8954562/authors#authors
market adoption, but it is not the sole determining Versatile Video Coding represents state-of-the 4. ResearchGate, SMPTE Motion Imaging Journal, vol. 128, no. 10, pp. 14-24, Nov.-Dec. 2019, Analysis
factor. Availability of the technology is critical, art video coding and is certain to play an important of Emerging Video Codecs: Coding Tools, Compression Efficiency and Complexity, J. L. Tanou and M.
especially the availability of hardware-accelerated role in supporting a wide range of 5G use cases. Blestel, available at: https://www.researchgate.net/publication/332073033_Analysis_of_Emerging_Video_
decoders in chipsets and processors. A recent Designed to manage the high demand that Codecs_Coding_Tools_Compression_Efficiency_and_Complexity
prediction from a major chipset vendor stated that increasing amounts of video poses on networks, 5. IEEE Transactions on Circuits and Systems for Video Technology, vol. 30, issue 5, pp. 1311-1325, May
first commercial VVC shipments could start as soon our research reveals that VVC achieves the best 2020, doi: 10.1109/TCSVT.2019.2945048, A VVC Proposal With Quaternary Tree Plus Binary-Ternary
as 2021. Software-based solutions are typically faster available compression performance at a Tree Coding Block Structure and Advanced Coding Techniques, Y. Huang et al., available at:
to roll out and are essential especially in the early computational complexity suitable for https://ieeexplore.ieee.org/document/8859290
phases of deployment across the ecosystem. In the implementation in both software and hardware.
case of VVC, the first demonstrations of real-time As one of the main contributors to VVC, we believe
software decoders took place shortly before the its deployment will significantly reduce the
standard was completed. data rates of existing video services as well as Further reading
Since video codecs do not operate in a silo, serving as a primary enabler for next-generation
❭ Ericsson blog, Cutting the bitrate with Versatile Video Coding, available at: https://www.ericsson.com/en/
support and interoperability across the ecosystem media services.
blog/2019/7/mpeg-cut-the-bitrate-versatile-video-coding
in terms of media delivery protocols and application
specifications are required. Some of the SDOs ❭ Ericsson blog, Industry Forum to promote Versatile Video Coding, available at: https://www.ericsson.com/
such as MPEG and IETF are looking into providing en/blog/2018/9/industry-forum-to-promote-versatile-video-coding
support for carriage of VVC in their respective

26
10 E RTI ECCSH
ERICSSON SNOO
N LTOEG
CYH N
ROE VL IOEG
WY ✱R EOVCI ETW
O B✱
E R #1 40 ,1 22002201 #
O0C1T2O0B2E1 R ✱
1 4 ,E2R0I2C0S S✱O N
E RTI ECCSH
SNOO
N LTOEG
CYH N
ROE VL IOEG
WY R E V I E W 27
11
 ✱ VERSATILE VIDEO CODING

Jacob Ström Kenneth Andersson

◆ is a principal researcher ◆ is a senior specialist in
the authors

at Ericsson Research with a video coding at Ericsson

focus on video compression. Research. He joined Ericsson
He joined Ericsson in 2001 in 1994 to work on speech
and has contributed to coding and since 2005 has
standardization in the area of been active in video coding
high dynamic range video as standardization in ITU-T
well as to the standardization and ISO/IEC for
Rickard Sjöberg of both HEVC and VVC. Łukasz Litwic development of HEVC
◆ is an expert in video He is coauthor of more than ◆ is a research leader at and VVC. He holds an M.Sc.
compression at Ericsson 120 granted patents and has Ericsson Research. He joined in computer science and
Research where he currently a similar number of patents Ericsson Television in 2007, engineering from Luleå
works as a technical lead in pending. Ström holds a Ph.D. where he worked on various University in Sweden
video coding research. in image coding from aspects of image processing and a Ph.D. from Linköping
He joined Ericsson in 1996 Linköping University, and video compression University.
and has contributed several Sweden, and has been a research, which formed
hundred proposals for visiting Ph.D. student at the the foundation of Ericsson
the ITU-T and MPEG University of California real-time broadcast
video-coding standards. San Diego and the encoding products. In 2017,
In addition, he has worked Massachusetts Institute of he joined Ericsson Research
in product development Technology (MIT) in the US. in Stockholm, Sweden,
related to video coding where he leads the
at Ericsson, including six Visual Technology team.
months at Ericsson He holds an M.S. from
Television in Southampton Gdansk University of
in the UK. Sjöberg holds an Technology, Poland, and a
M.S. in computer science Ph.D. from the University of
from KTH Royal Institute Surrey, in Guildford in the UK.
of Technology in Stockholm,
Sweden.

12 ERICSSON TECH NOLOGY REVIEW ✱ OCTO B ER 14, 2020

28 ERICSSON TECH NOLOGY REVIEW ✱ #01 2021
✱ INTENT-BASED NETWORKING INTENT-BASED NETWORKING ✱

performance metrics. Complexity arises as a result An intent in an autonomous system is ideally

of changes to contracts, products, customer expressed declaratively – that is, as a utility-level goal
preferences, the business strategy or the that describes the properties of a satisfactory
environment in which the service is being offered. outcome rather than prescribing a specific solution.

Cognitive processes
Users may start exhibiting new behavior, leading This gives the system the flexibility to explore
to varying service usage patterns and network loads, various solution options and find the optimal one.
or the network may change due to upgrades, It also allows the system to optimize by choosing
reconfiguration or outages. Some changes may its own goals that maximize utility.

for adaptive intent-based

be regular and predictable, whereas others are Unlike traditional software systems, where
sudden and surprising. requirements are analyzed offline to detect and
To manage all of these concerns autonomously resolve conflicts prior to implementation, intents
and adapt its behavior appropriately, a zero-touch are added to an autonomous system during runtime.
system must understand every aspect of what is Adaptation to changed intent as well as conflict

networking
expected of it. Each requirement and goal must be detection and resolution are therefore essential
carefully defined in order for technical processes to capabilities of an autonomous system.
derive suitable and optimized actions to manage it. One of the benefits of expressing intents as utility-
These definitions are known as intents. level goals is that it helps the system cope with the
From the perspective of a human operator, an conflicting objectives of multiple intents. This is vital,
intent expresses the expectation of what the because an autonomous system often has to take
Autonomously operated and self-adapting networks will make it possible to operational system is supposed to deliver and how multiple intents into account before making a decision.
utilize the capabilities of 5G networks in new business models and achieve it behaves. In light of this, we define intent as For example, an autonomous system may have
“formal specification of all expectations including one intent to deliver a service with high QoE,
an unprecedented level of efficiency in service delivery. Intents will play a requirements, goals and constraints given to a while another may be to minimize resource
critical role in achieving this zero-touch vision, serving as the mechanism technical system.” spending. It can resolve such conflicts either
explicitly from weights that introduce relative
that formally specifies what the autonomous system is expected to do. The role of intents in cognitive networks importance or implicitly from properties of
Everything an autonomous system needs to know preferential outcomes as defined in utility-level goals.
about its goals and expected behavior must be Expectations originate from contracts or business
defined with intents. The system will not perform strategy and remain constant when the underlying
any operation unless it relates to the fulfillment system is replaced or modified. Consequently,
and assurance of an intent, which means that all when setting up the intents, it is important that
JÖRG NIEMÖLLER,
5G networks introduce unprecedented ■ The zero-touch paradigm implies that the goals – including those that may have been they are formulated in an infrastructure-agnostic
LEONID MOKRUSHIN, flexibility and dynamic adaptation into operation of services and the underlying networks considered “common sense” in human-operated way, so that they can be transferred across system
SWARUP KUMAR service delivery and network resource is autonomous and does not require human systems – must be expressed as intents. generations and implementations.
MOHALIK, utilization. In the business layer, this is intervention. To achieve this, the zero-touch system
MARTHA VLACHOU- reflected in the ability to offer customizable must be able to handle the complexity caused by
KONCHYLAKI, service products with detailed agreements continuous changes to the system at the same time
GEORGE SARMONIKAS on functional and non-functional that it delivers services to users and manages issues
Terms and abbreviations
characteristics as well as fast delivery. such as the cost of resources versus the budget
Dynamic adaptation to changes within available, the legal compliance of the service and AI – Artificial Intelligence | BSS – Business Support Systems | CEM – Customer Experience Management |
the constraints of stringent requirements the security of the setup. This is challenging for IOT – Internet of Things | KPI – Key Performance Indicator | OSS – Operations Support Systems |
RDF – Resource Description Framework | SLA – Service Level Agreement | SLO – Service Level Objective |
on lead and reaction times is beyond the technical systems in real-world scenarios.
SLS – Service Level Specification | SON – Self-Organizing Networks | TOSCA – Topology and
capacity of a human workforce. Extensive For example, successful service operation
Orchestration Specification for Cloud Applications | VNF – Virtual Network Function
automation will be necessary to overcome requires each service to be properly provisioned and
this challenge. assured to deliver the promised function with agreed

30
2 E R I C S S O NETREI C S
HSNOONL O
TGEC
YHRNEO
V LI EOW
G Y✱ R N
EOV IVEEW
M B✱E R#101 ,1 22002201 #
NO ✱1 1E, R2I0C2S0S O
0 1V2E0M2B1 E R ✱NETREI C SHSNOONL OT G
ECYHRNEO
V LI EOW
GY REVIEW 31
3
✱ INTENT-BASED NETWORKING INTENT-BASED NETWORKING ✱

In short, the intent establishes a universal model is the instrumental goal of OSS and becomes
mechanism for defining expectations for different an intent and terminal goal for the orchestrator.
layers of network operation. It expresses goals, This pattern of making decisions based on a given Intent Intent Intent Report

utility, requirements and constraints. It defines intent and taking action by sending lower-level
expectations on service delivery as well as the intents to subsystems is the key interwork
behavior of the autonomous operational system mechanism of intent-based operation, according to
and the underlying network. which the entire operations stack of autonomous Agents
Intent-handling
networks is built.
function
[THE INTENT] EXPRESSES Knowledge Decision Actuation Policies Services Models
Strategic and behavioral intents
GOALS, UTILITY, REQUIREMENTS Beyond all the service-specific intents that an
AND CONSTRAINTS autonomous system must have, it also requires
guidance on how to handle strategic and
Service-specific intents behavioral concerns. Traditionally implemented Report Intent Intent Action

One essential type of intent relates to the in the form of manually coded policies, this type
specification of services. Service-specific intents of guidance steers general system behavior and
state expected functional and performance supports the type of decision-making that has Network
characteristics. Service Level Agreements (SLAs), traditionally been based on human intuition and
Service Level Specifications (SLSs), Service Level experience, along with knowledge about context
Objectives (SLOs) and TOSCA (Topology and and operator strategy. Intent-based operation Figure 1 The intent-handling function
Orchestration Specification for Cloud makes it possible for operators that want to handle
Applications) [1] models are all examples of these concerns in a more dynamic fashion to Formally expressing an intent actions must be taken to optimally fulfill all given
service-specific intents that are used on replace manually coded policies with strategic An autonomous system requires intents to be intents and implements its decisions.
different levels in the operations stack. and behavioral intents. formally defined in a machine-readable and Intent-handling functions have a knowledge base
SLAs are business support systems (BSS) objects. This is useful in cases where the operator processable way, but the broad range of that contains the intent ontology. They also have
Service-specific intents based on SLAs specify the chooses to require a default minimum security considerations involved and their abstract machine-reasoning capabilities to realize knowledge-
promised service and include expected performance level that differs from that which is implemented semantics are often difficult to structure. driven decision-making processes.
details and business consequences such as payment into the service, for example. In these cases, Techniques from knowledge management and Machine reasoning plays a key role in intent
for delivery and penalties when failing. dedicated intent can be used to set the security semantic modeling enable the creation of an handling, with its capability to understand abstract
SLSs/SLOs define the service delivery details level for all services that do not specify it directly. ontology of intent, based on an extensible concepts from diverse domains and provide precise,
at operations support systems (OSS) level. With regard to legal compliance, services may metamodel. Resource Description Framework specialized conclusions based on precedent and
Based on this input, autonomous OSS would be delivered in multiple markets where different (RDF) [2] and RDF Schema [3] standards can be observation. Probabilistic modeling contributes
plan detailed tasks to realize the service delivery. rules apply. A legal-compliance intent requires used for knowledge modeling. quantification of risk and uncertainty, which is essential
TOSCA models would be used to express compliance and potentially specifies the details. Technical functions such as contract and order to make informed decisions when facing conflicting
further technical details the OSS generate Since there is always a risk of service degradation management would directly use RDF objects to goals and new situations.
(expectations from orchestration and assurance). when changes are initiated and risky actions may communicate intent. Intent specified directly by Figure 1 shows how the intent-handling function
In multiple stages, the autonomous operation sometimes lead to a higher margin, risk-management human operators would require an intuitive works. While its implementation is domain-specific,
makes decisions about further details. Higher-level intents can be used to convey how the operator frontend, potentially using natural language. its interface is generic. It receives intents that express
intent is the input leading to the lower-level intent wants the autonomous system to balance risks all types of expectations. It is equipped with policies
that is used to distribute specific goals to subsystems. versus potential gain. Intent handling and artificial intelligence (AI) models that implement
For example, SLAs/SLSs are the intents that Reporting/escalation intents steer how the The operation of services within an intent-based the capabilities needed for analyzing the system state
express a terminal goal of the OSS. The OSS then autonomous system interacts with the human network also requires the introduction of intent- and finding optimized operational actions based
decide which TOSCA model would be the best workforce by reporting progress status on handling functions in the operations stack and on observations from the operated environment.
option to deliver the promised performance with intent fulfillment and seeking manual decision functional architecture. An intent- handling The intent handler also reports the fulfillment and
minimal resource usage. The selected TOSCA in escalations. function receives the intents, decides which assurance status of its intents.

32
4 E R I C S S O NETREI C S
HSNOONL O
TGEC
YHRNEO
V LI EOW
G Y✱ R N
EOV IVEEW
M B✱E R#101 ,1 22002201 #
NO ✱1 1E, R2I0C2S0S O
0 1V2E0M2B1 E R ✱NETREI C SHSNOONL OT G
ECYHRNEO
V LI EOW
GY REVIEW 33
5
✱ INTENT-BASED NETWORKING INTENT-BASED NETWORKING ✱

Business
Business intent video quality (lower is better)
operationsRelative bitrates for the same
Goals, SLA, behavior Visualization, approval, escalation, monitoring
Customer portal Customer
engagement
180% 170%
160% Knowledge: Reasoning:
140% Descriptions of objects and Access to knowledge

Cognitive
relations between them
120% Co
mp
via inference
120% Dynamic
knowledge 100%
Re
qu
e100%
os
itio
n
BSS layer st
Business 100% update fo
r in
fer

Business portal intent Contact and order level 80% Data:

en
ce Agents:
handler management
Efficient storage and Data and 60%
Modules for machine
60% retrieval of objects streams Machine learning and modeling
ML learning, policies,
actuation, etc.
40%
20% Observations Actions
Network
0% and
environment
Operational H.264 VP9 Radio HEVC
Core CEM BSS AV1 IoT VVC
OSS intent handler level
Operations portal Figure 3 Functional architecture of the cognitive layer

Operational and understanding through experience and contains the ontology of intents along with domain-
the senses.” [4] As it is designed to perform the specific knowledge such as the current state of the
domains
equivalent operational tasks of understanding system. The domain-independent reasoning engine
SON Orchestration Network manager through experiencing and sensing, an autonomous uses the knowledge graph and serves as the central
system is, therefore, a technical implementation coordinator function for finding actions, evaluating
of cognition. their impact and ordering their execution. Finally,
Creating an intent-handling function that the agent architecture allows any number of models
Network and
understands complex and abstract intent semantics, and services to be used. Agents can contain
resource level derives the optimal target state and plans actions machine-learned models or rule-based policies,
Network Cloud IoT
for transitioning the system into this state is a or implement services needed in the cognitive
challenging task. The function must be able to reasoning process.
Figure 2 Intent-driven multi-layer operations system explore options, learn from precedents and assess To be usable, an agent needs to be registered and
the feasibility of actions based on their expected described in the knowledge base. Its description can
The API (application programming interface) of a complete intent-driven operations system. consequences. be added and modified at any time, allowing life
the intent-handling function is domain-independent. Every major system layer and subsystem domain, By combining well-understood AI techniques cycles of the models, policies and supplementary
Its main objective is to manage the life cycle of including BSS, OSS, orchestration and network within a flexible architecture, we have designed services to be decoupled from the overall life cycle
intents. It implements methods to set, modify and management, contains an intent-handling function. a cognitive system that specializes in autonomous of the cognitive layer.
remove intents and send reports. Intent is Intent originates from functions such as contract service and network operation. We refer to it as the The agent metadata contains a description of the
constructed based on a common intent meta-model and order management. Additional intents can be cognitive layer, and its role is to serve as an interface agent interface, along with its function, role and
and its details are specified according to domain- entered directly through portals. between business operations and the network/ capabilities. For example, we have implemented a
specific information models. Intent management is environment, as shown in Figure 3. machine-learned model that can propose radio base
therefore primarily knowledge management. Introducing the cognitive layer The cognitive layer consists of three essential station configurations that optimize the service
Figure 2 provides an example of how intent- Lexico defines cognition as: “… the mental components: a knowledge base, a reasoning engine experience. This model is registered as an agent
handling functions can be combined to realize action or process of acquiring knowledge and an agent architecture. The knowledge base in the role of a “proposer” for configuration actions.

34
6 E R I C S S O NETREI C S
HSNOONL O
TGEC
YHRNEO
V LI EOW
G Y✱ R N
EOV IVEEW
M B✱E R#101 ,1 22002201 #
NO ✱1 1E, R2I0C2S0S O
0 1V2E0M2B1 E R ✱NETREI C SHSNOONL OT G
ECYHRNEO
V LI EOW
GY REVIEW 35
7
✱ INTENT-BASED NETWORKING INTENT-BASED NETWORKING ✱

The separate life cycle makes it possible for the solution strategies, resulting in a diverse set of parameters of the underlying TOSCA model. When a proposed action is selected after verifying
model to be replaced with an improved version options for the prediction and evaluation steps Combined with measured KPIs, these variations that the impact on all intents is advantageous and
when available, independent of the cognitive layer that follow. This coexistence of agents makes it create training data sets for learning a model that is risk is reasonably low, it is then handed over for
release cycles. possible to combine rule- and policy-based capable of connecting deployment options to execution to the orchestrator behind the actuation
We have also demonstrated agents in the role of implementations with machine-learned alternatives expected performance. This exercise created a agent. This concludes the first intent-handling loop
“predictor” with the ability to estimate the effect in the same system, enabling the system to acquire proposal agent that is able to recommend a TOSCA after the new intent for the new SLA was introduced
of actions on key performance indicators (KPIs). new advanced abilities without losing current ones. model configuration optimized for the latency and to the cognitive layer. The gap between the new
An agent in an “observer” role would monitor data The cognitive process is a perpetual loop that throughput figures required by the SLA. intent and the current state was particularly wide
sources, keeping knowledge about the state up-to- starts again directly after the previous iteration has In our experiment, online fulfillment – from since the new service was not yet provisioned.
date. An agent in the “actuator” role can implement finished. Any degradation of the network or issues receipt of the SLA to optimized deployment – This gap narrowed through a provisioning action.
actions in the network by utilizing, for example, with services would be visible in the observed state. is fully autonomous and controlled by the reasoning Further iterations continuously monitor the service
established network management functions. By trying to close the gap to the wanted state set by process of the cognitive layer. When the intent deployment, optimize it when possible or heal when
intent, the cognitive layer implicitly addresses derived from the SLA arrives in the knowledge base, needed. In this way, the continuous cognitive
SPECIALIST AGENTS incidents. Even without explicit issues, the the reasoning loop starts processing it. The reasoner operation process provides fulfillment and
continuous cognitive process would still seek actions finds that the agent learned in the test environment assurance of expectations formulated as intents.
ARE USED INTENSIVELY for further optimization. It could, for example, try to matches the needs and requests a proposal.
IN EVERY STEP OF deliver the same services with reduced resources. The agent delivers a TOSCA model fully configured ONLINE FULFILLMENT
Through its reasoning-based core process, the and optimized for the requested KPI target and
THE PROCESS cognitive layer reaches a high degree of dynamic proposes to deploy accordingly. IS FULLY AUTONOMOUS
How the cognitive layer works
adaptability to new situations. This is in stark
contrast to systems that have been realized through
The cognitive layer then uses prediction and
evaluation agents to assess the proposal. In this case,
AND CONTROLLED BY THE
The successful operation of the cognitive layer rule-based policies and fixed workflows, where every our prediction agent contains a state-action model – REASONING PROCESS
depends on smooth interaction between the supported situation needs consideration at design a probabilistic graph based on Markov decision
reasoning engine and the knowledge base. time through suitable branches in the decision tree process modeling. The model is continuously Conclusion
The reasoning engine continuously executes and diversifying rules. Existing rule-based policies learned from observing states and the results of Zero-touch autonomy is an ideal beyond reach
a process that tries to find actions to close the can, however, still be used in the cognitive layer observed actions. It has the ability to estimate the as long as artificial intelligence (AI) cannot
gap between the current observed state and the through integration as agents. This opens an upgrade probability of expected result states for proposed match human capability to reason and decide
wanted state, according to the intent. It collects path from legacy automation, with AI-based models actions, enabling informed decisions about actions within complex dependencies and broad domains.
proposals, obtains predictions on the effect of added gradually. considering their risk. However, by using a combination of currently
each proposal, evaluates gain versus risk and In the evaluation step, the autonomous system available and well-understood AI techniques
certainty, prioritizes actions and executes its Example use case detects and resolves all remaining conflicts between within a flexible architecture, it is possible to reach
decisions. Specialist agents are used intensively To demonstrate how the cognitive layer works, intents. It also decides on escalations if the risk of the a high degree of practical autonomous operation.
in every step of the process. we have experimented with a use case that optimizes proposed deployment or uncertainty of the models is Our use case example demonstrates how the
The reasoning engine is an adaptive knowledge- the provisioning decisions for the deployment of a too high. If escalation is required, the cognitive cognitive layer that we have developed can enable
driven composer that can instantiate the cognitive virtual network function (VNF). The source of the process requests the support of a human technician, autonomous operation with the use of intents.
process following changes in intent, state and intent is an SLA that specifies the service to be presents the situation including proposals and It is built with an agent architecture that decouples
context. It can dynamically compose specialized delivered along with the KPI targets that have predictions, and asks for approval. the life cycles of the agents. It is coordinated by
agents and add them into the process if their been promised to the customer. Our example While the cognitive layer can operate fully knowledge-driven reasoning that composes
capabilities and roles match the needs according use case requires that the VNF be deployed autonomously, it knows when the human workforce machine-learned models and legacy implementations.
to intent and context. This is, for example, how the with strict targets on latency and throughput. wants to be involved. The exact threshold for It evaluates action impact on intent fulfillment and
cognitive layer obtains action proposals from We started by developing the required proposer escalation is at the discretion of the operator and is makes decisions based on expected gains and risks,
agents that are implementing suitable models. and predictor agents in an offline data science determined by behavioral intent. If the system gains while resolving conflicting goals and maximizing
In cases where the capabilities of multiple agents process. We then deployed the service in a test human trust as a result of presenting many good utility. It can adapt to new situations through
match the requirements of a role, each of them environment and exposed it to a range of usage loads. actions, the threshold for fully autonomous decisions learning. The resulting system can execute many
can be used simultaneously to generate alternative We also explored deployment options by variating can be lowered. of the cognitive considerations human technicians

36
8 E R I C S S O NETREI C S
HSNOONL O
TGEC
YHRNEO
V LI EOW
G Y✱ R N
EOV IVEEW
M B✱E R#101 ,1 22002201 #
NO ✱1 1E, R2I0C2S0S O
0 1V2E0M2B1 E R ✱NETREI C SHSNOONL OT G
ECYHRNEO
V LI EOW
GY REVIEW 37
9
✱ INTENT-BASED NETWORKING INTENT-BASED NETWORKING ✱

make when they operate services and networks the need for humans to manually modify
manually. However, the cognitive layer does policies or participate in online decision-making.
this with continuous attention and near- This frees up the human workforce to
immediate reaction. concentrate on offline tasks such as executing
The adaptive reasoning capabilities of the data science processes, optimizing the available
cognitive layer enable effective management of models and defining business strategies that
growing network complexity, dramatically reducing are implemented by intent setting. in the telecom domain. focuses on applying them to of Technology in Stockholm,
Mokrushin joined Ericsson in service automatization and Sweden, and a B.Eng.

the authOrs
2007 after postgraduate the Internet of Things (IoT). in electrical and computer
studies at Uppsala University He has research experience engineering from the
in Sweden, where he in the areas of formal University of Patras, Greece.
specialized in formal specification and verification
verification of real-time of real-time embedded George Sarmonikas
software and AI planning ◆ joined Ericsson in 2013
techniques. Mohalik holds after several years of
Jörg Niemöller a Ph.D. in computer science working for mobile
◆ is an expert in analytics from the Institute of operators. He currently leads
and customer experience. Mathematical Sciences, AI business innovation within
He joined Ericsson in 1998 Chennai, India, and a post- Digital Services. Prior to this,
and has since held multiple doctoral fellowship at LaBRI, he was responsible for
positions in research as well University of Bordeaux, product management
as in system management France. of Ericsson’s customer
for core network and digital systems. He holds an M.S. experience management
services. His current focus in software engineering (CEM) and analytics
is innovation in OSS through from Peter the Great St. portfolio, including assets
architecture and solutions Petersburg Polytechnic for subjective experience
References for autonomous network and University, Russia. scoring. Sarmonikas holds
1. OASIS, TOSCA Version 2.0, Committee Specification Draft 02, June 25, 2020, available at: http://docs. service operation. Niemöller both an M.Sc. in
oasis-open.org/tosca/TOSCA/v2.0/TOSCA-v2.0.html holds a Ph.D. in computer Swarup Kumar communication systems
science from Tilburg Mohalik and an M.Eng. in electronic
2. W3C, RDF 1.1 Concepts and Abstract Syntax, W3C Recommendation, February 25, 2014, available at:
University, the Netherlands, ◆ is a principal researcher at engineering and computer
https://www.w3.org/TR/2014/REC-rdf11-concepts-20140225/
and a diploma degree in Ericsson Research who Martha Vlachou- science from the University
3. W3C, RDF Schema 1.1, W3C Recommendation, February 25, 2014, available at: https://www.w3.org/TR/ electrical engineering from joined the company in 2015. Konchylaki of Bristol in the UK, as well
rdf-schema/ the TU Dortmund University, His expertise is in the areas ◆ is a director of technology as a graduate diploma in
4. Cognition, Lexico.com, available at: https://www.lexico.com/definition/cognition Germany. of AI and formal methods, strategy specializing in AI artificial intelligence from
and his work primarily and data strategy. She joined Stanford University, US.
Leonid Mokrushin Ericsson in 2015 and has
◆ is a principal researcher at been working within different
Further reading Ericsson Research. With a groups within Ericsson,
background in computer from machine learning
❭ Ericsson Technology Review, Cognitive Technologies in Network and Business Automation, 2018,
science and formal methods, prototyping and business
Niemöller J; Mokrushin, L, available at: https://www.ericsson.com/en/reports-and-papers/ericsson-technology-
he is currently focusing on development to AI strategy.
review/articles/cognitive-technologies-in-network-and-business-automation
knowledge-intensive Vlachou-Konchylaki holds an
❭ Ericsson, AI by Design, available at: https://www.ericsson.com/en/ai-and-automation symbolic AI systems and M.S. in machine learning
their practical applications from the KTH Royal Institute

38
10 E R I C S S O NETREI C S
HSNOONL O
TGEC
YHRNEO
V LI EOW
G Y✱ R N
EOV IVEEW
M B✱E R#101 ,1 22002201 #
NO ✱1 1E, R2I0C2S0S O
0 1V2E0M2B1 E R ✱NETREI C SHSNOONL OT G
ECYHRNEO
V LI EOW
GY REVIEW 39
11
✱ AI IN RAN AI IN RAN ✱

Artificial intelligence Network applications (RAN / Core)

Dynamic infrastructure optimization

• Digital sibling (E9)

RAN-algorithm- and user-centric

Network design / O&M
• Rules & policies (E8)

DevOps design flow

V/CNF infrastructure PNF infrastructure

in RAN
management management

optimization
• Agent layer (E7)
• Simulation (E6)
Transport network
• Open source (E5)
• Model life-cycle
management (E4) Site
A SOFTWARE FRAMEWORK FOR AI-DRIVEN RAN AUTOMATION • Training & inference (E3)
• Message bus (E2) Application and infrastructure zones Automation domains
• Data layer (E1)
Software enabler framework for AI-based RAN automation

Artificial intelligence and its subfield machine learning offer well-established Figure 1 Software-enabler framework for AI-based RAN automation
techniques for solving historically difficult multi-parameterization problems.
Used correctly, these techniques have tremendous potential to overcome
efficiency by drastically simplifying certain types of parameters are chosen and tuned is partially
complex cross-domain automation challenges in radio networks. automated management and control. It is important dependent on the selected distribution over
Our ongoing research reveals that an integrated framework of software to realize, however, that this development will not compute, transport and site resources. The right
occur without effort; on the contrary, the need for side of Figure 1 shows the four automation domains
enablers will be essential to success. AI and ML techniques will have a profound effect that play a role in optimal network tuning across
on how the next generation of RAN should be applications and infrastructure:
designed and realized. ❭ DevOps (development and operations)
As part of this work, we have created a set of design flow
DIARMUID CORCORAN,
Modern telecommunications and mobile concurrent Internet of Things device access and software enablers that target efficient and effective ❭ Network design/operations and maintenance
ANDREAS ERMEDAHL, networks are becoming increasingly complex enhanced mobile broadband. use of AI technologies in RANs. Together, they form
(O&M)
CATRIN GRANBOM from a resource management perspective, It is well accepted in the communications an important framework of building blocks in
❭ Dynamic infrastructure optimization
with diverse combinations of software and community that appropriately dimensioned, efficient realizing the next level of autonomous, AI-based,
❭ RAN-algorithm and user-centric optimization.
infrastructure elements that need to be and reliable configurations of systems like 5G are a resource management in 5G and future-generation
configured and tuned for efficient operation complex technical challenge. Increased real-time, networks. Sophisticated use of AI techniques involves deep
with high QoS. closed-loop, automation at all levels and time frames and disruptive changes to many existing software
is a critical tool in controlling this complexity and Landscape of automation in RAN engineering, algorithm design and workflow stages.
■ The latest 5G mobile system is a good example ultimately reducing the capex and opex of radio The wide variety of architectural and transport The DevOps (Development and Operations) flow,
of a sophisticated radio network that allows many network operations. options that are possible in 5G results in several for example, needs to be extended with new data
deployment variations – such as centralized, There is currently a significant trend toward the different configuration permutations [1]. RAN and ML model life-cycle management (MLCM)
distributed or various hybrids of both – while use of artificial intelligence (AI) technology as an application(s) can be distributed in a flexible way procedures. Network design/O&M, dynamic
simultaneously supporting diverse categories enabler for automating both repetitive and complex over both physical (PNF) and virtualized/container infrastructure optimization and RAN-algorithm/
of applications such as mission-critical control tasks. Applied correctly, AI and machine learning (V/CNF) infrastructure, as shown in the middle user optimization also need to be reengineered
with ultra-reliability and low latency, massively (ML) techniques have the potential to boost system section of Figure 1. How RAN application in a radically but carefully staged approach [2].

42
2 E R I C S S O NE RTIECCSHSNOONLTOEGCYH RNEOVLIO
EWG Y✱R EDVEICEEWM ✱
B E R# 80 ,1 22002201 #
D0 M2B1 E R✱8 E
1 2E 0
EC 2 0S S✱
, 2R0I C O NE RTIECCSHSNOONLTOEGCYH RNEOVLIO
EWGY REVIEW 43
3
✱ AI IN RAN AI IN RAN ✱

Software-enabler framework for AI-based Depending on the problem targeted, the best
RAN automation ML model and supporting life cycle (E4) can vary Ericsson scope CSP/customer scope
Our research and supporting proof-of-concept considerably in complexity. For simpler problems,
systems have shown that a set of software enablers linear regression, smaller decision-tree and simple Simulation
RAN application
is required to best facilitate new AI approaches to neural networks (NNs) with a few nodes and layers Data Explore (distributed)

sophisticated closed-loop automated network can be sufficient. For more complex problems, large 1
Retrain
design. These enablers must be tightly integrated decision trees or deep neural networks (DNNs) with
Ericsson
Model scope
staging zone
into a software framework that allows flexible and many layers and nodes and several convolutional layers 2 3
open information exchange. As shown on the left could be necessary to achieve the required accuracy. Application
design
side of Figure 1, our software-enabler framework RL approaches and supporting agents (E7)
for controlling optimization goals can be especially 4
for AI-based RAN automation consists of nine key
enablers: effective at learning novel RAN management Model harvesting zone
❭ A data layer (E1) strategies. Training RL models depends on active 5

❭ A message bus (E2) exploration, through a software agent’s trial and Centralized/distributed/federated
Central training infrastructure
❭ Training and inference (E3) error experiments, which will not always be possible training infrastructure
or appropriate in a live RAN system. To help solve
❭ Model life-cycle management (E4)
this problem, and to generate the required quantities
❭ Open-source AI software (E5)
of data to train models, we have included simulation Main pipelines: Data/KPIs Explore Deploy Harvest
❭ Integrated simulation (E6)
(E6) in our set of software enablers.
❭ A decision-making agent layer (E7)
Once trained, an ML model can be used in the
❭ Rules and policies for intent-based Figure 2 Model life cycle management
inference phase (part of E3), where a selection of
management (E8) data is used as input into the model that will then
❭ A digital sibling for knowledge harvesting (E9). produce a set of predictions, actions or rules, with When moving toward AI-based automation, faults. With potentially many instances of a model
Perhaps the most essential enabler of all is the the exact details depending on the ML algorithm operators will no longer spend time tweaking RAN deployed at cell, cell cluster or indeed user level,
production of industry-standard, machine-readable type. In RAN, the hardware and software performance by adjusting individual parameters this challenge is compounded. Figure 2 captures the
data (E1) at a high temporal rate by RAN network requirements on the training and inference phases exposed by the RAN system. Instead, well-defined main conceptual stages in model management from
nodes (eNodeB or gNodeB). This data can be can be vastly different. Training typically requires high-level rules and policies (E8) for expressing a software process perspective.
aggregated, stored, filtered and distributed by powerful central processing unit or specialized overall operational intent are used. The AI-based After an initial model design and training phase
advanced real-time, hierarchical, message bus (E2) graphics processing unit (GPU) hardware with automation machinery is then responsible for (marked as 1 in Figure 2), possibly supported by
solutions. A distributed training and inference (E3) large memory and data storage. realizing these goals. simulation components, the model enters a staging
architecture can efficiently consume this data AI software platforms, such as TensorFlow, Keras Long term, to allow the most sophisticated forms zone (phase 2). This is where customer- or hardware-
and use it to facilitate the training of ML models and PyTorch, together with other extensive, open of automation, we see a need to build partial digital specific model adaptions [6] occur and are packaged
to create new AI algorithms and/or parameter source (E5), often Python-based, ML software representations of deployed systems. Our unique into an industry-deployment format. For example,
selection for classic algorithm design [2]. ecosystems need to be integrated into the software digital sibling (E9) concept makes it possible to a DNN can be pruned to minimize nodes that
There are many alternatives [3] to creating/ engineering flow. During the inference phase, harvest ML models to build context-specific are not significant with additional quantization
training models for use with ML algorithms, a trained model (or models) is made available knowledge. transformation to reduce the number of bits required
with three main classifications: to the RAN application(s) through model life-cycle to represent a number. This can significantly reduce
management (E4). For latency-critical RAN edge Model life-cycle management size, producing an embedded DNN device
1. Supervised learning, in which algorithms
applications, the inference needs to be efficiently Machine-learning models are valuable, evolving representation with just a minor impact on accuracy.
are trained using labeled examples
realized, with low latency, power consumption and assets. As such, a sophisticated and controlled In phase 3, the model(s) is deployed into a RAN
2. Unsupervised learning, using a cluster
memory footprint, taking the characteristics of the software engineering chain (E4) is required for application for use. Initially, these models will be
or grouping technique
target hardware and software architecture into MLCM. When a model is updated through tightly integrated with the RAN application. In the
3. Reinforcement learning (RL) [4], using a target
account. retraining outside its initial context, for example, long term, however, it will be necessary to separate
reward strategy that allows guidance toward Our software enablers are fully compatible many significant challenges are introduced in terms model and application life cycles to allow for rapid
an optimal set of actions. with intent-based management solutions [5]. of traceability, stability and the possibility to identify model training/retraining and deployment cycles.

44
4 E R I C S S O NE RTIECCSHSNOONLTOEGCYH RNEOVLIO
EWG Y✱R EDVEICEEWM ✱
B E R# 80 ,1 22002201 #
D0 M2B1 E R✱8 E
1 2E 0
EC 2 0S S✱
, 2R0I C O NE RTIECCSHSNOONLTOEGCYH RNEOVLIO
EWGY REVIEW 45
5
✱ AI IN RAN AI IN RAN ✱

WE ENVISION THE USE when moving toward RAN systems that utilize
ML and RL technologies, this will no longer hold.
management allows different data sources to be
controlled and activated in a very fine-granular
Looking ahead, more sophisticated AI-based
automation in the RAN will require partial digital
OF OUR NOVEL MODEL- For example, ML models may develop manner. Information on the data and associated representations of deployed systems to support
HARVESTING TECHNIQUE TO independently and not be the same for all RAN
nodes, and they may not remain constant between
schemas is stored and organized in distributed
databases, called data- and schema inventories.
complex decision-making. In our framework,
we call this enabler a digital sibling (E9).
CREATE DIGITAL SIBLINGS RAN software releases. As a result, new ways to Our data distribution is based on message bus
continuously test and validate RAN software technologies, allowing the produced data to be Digital sibling
Retraining and model drift utilizing ML models must be developed to fit the automatically tagged with relevant keys and to be The digital sibling is a way to capture and transfer
Certain types of models (those dependent on traffic MLCM phases. automatically routed to different subscribers based data, and especially ML model data, from a real-world
and variation in load, for example) may not initially Moreover, new fine-grained KPIs and metrics on their tag subscription requests. Our research- context. These digital representations can then be
perform well in their deployed context. Over time, must be collected during system run time and fed developed message bus is hierarchical and distributed, used for increasingly accurate data generation and
there may also be drift between the model training back to the MLCM. This includes new KPIs for meaning that the produced data will not be routed in-loop simulation. A related concept, the digital
context and the current operating state. As such, it is deployed ML models, such as their perceived longer than needed, making it particularly suitable twin, endeavors to duplicate, more completely, a full
necessary to allow for periodic, active retraining accuracy, as well as performance and resource for the requirements of a virtualized and digital copy of a physical resource or system. The
with context-specific data and features in phase 3. consumption KPIs for the RAN applications disaggregated RAN. scope of our digital sibling is limited to the
utilizing the ML models. Similar to other type of transferable digital knowledge, as captured in
Model harvesting RAN data relevant for machine learning, the KPIs Enhancing machine learning in the RAN models that have been trained, interchangeably, in
Model harvesting is a way to improve simulation should be distributed by the data layer. with simulation both a simulated and real-world RAN environment.
and ML training capabilities by enabling a model Despite ongoing work to improve high-fidelity data After an initial training phase, an ML model
feedback loop from deployment to design contexts. A data layer targeting machine learning collection in the RAN, its distributed nature means it captures a considerable amount of transferable,
Phase 4 includes the anonymized collection of model in the RAN is likely that data will continue to be sparse and that problem-specific knowledge, including knowledge
data from a CSP deployment back into our model A cornerstone for producing high-quality ML and data collection will continue to be a resource- that is compressed and coded in the weights of an
harvest zone. This facilitates sophisticated data RL models is good access to data. Currently, there consuming task. At the same time, learning through NN or a set of rules in a decision tree. Sets of models
mining from these models, including valuable are many mechanisms for the generation, collection exploration can be difficult in live systems, where can be used to capture the different decision-making
clustering information on RAN resource usage. and analysis of data from the RAN. One example is exploration may degrade or crash the system. needs of a system. For example, a set of ML models
The harvested data can then be used as input for the performance management and configuration Our approach to mitigating these challenges is to can be trained in a local CSP context to predict the best
model training and retuning when the simulator management data reported in a standardized format integrate in-loop simulation (E6) into our framework, link adaption or power control parameters for a user.
configurations are updated, for example. In the from RAN nodes. Another example is debug and providing the means for much richer data generation Other examples include the choice of
longer term we envision the use of our novel model- trace data, which includes detailed information as well as the use of explorative RL techniques. geographical baseband workload placement
harvesting technique to create digital siblings about the internal RAN system state. The latter data Allowing in-loop simulation creates new possibilities required for a given service category and anomaly
(phase 5) that partially reflect different real-world is often generated on demand – that is, associated for complex, multistep, scenario exploration without detection like rogue users or a base station with
RAN deployments. with a trouble report or customer service requests. compromising system integrity. The first phase of malicious intent. This aggregated model information
When targeting generation and use of ML models our in-loop simulation is restricted to the Ericsson (transferred through our MLCM in phase 4),
Test, verification and run-time monitoring of for time-critical RAN control loops, none of the development scope shown in Figure 2, but it will together with actions by goal-oriented software
RAN applications utilizing machine learning existing mechanisms is a perfect match. At Ericsson, evolve to include customer deployment in the future. agents, forms a considerable body of knowledge that
A data-driven and continuously learning RAN will we are actively investigating and introducing new
introduce new requirements on how software is tested, mechanisms for the generation, management and
verified and monitored during system run time. distribution of data suitable for ML in the RAN.
Today, exhaustive test and verification is performed, Our lightweight data generation mechanisms allow Terms and abbreviations
both on individual software modules and at system high quantities of data to be generated without
AI – Artificial Intelligence | CSP – Communication Service Provider | DevOps – Development and Operations |
level, automated over various continuous integration adversely affecting core RAN system performance.
DNN – Deep Neural Network | DSP – Digital Signal Processor | GPU – Graphics Processing Unit |
loops before deployment. The same software is The data is produced using a few standardized
ML – Machine Learning | MLCM – Model Life-Cycle Management | NN – Neural Network | PNF – Physical
deployed on most RAN nodes, perhaps somewhat machine-interpretable formats. Each data format has
Network Function | RL – Reinforcement Learning | V/CNF – Virtualized/Container Network Function
differentiated depending on the hardware, surrounding an associated schema allowing for easy serializing,
equipment and customer preferences. However, deserializing and content interpretation. The data

46
6 E R I C S S O NE RTIECCSHSNOONLTOEGCYH RNEOVLIO
EWG Y✱R EDVEICEEWM ✱
B E R# 80 ,1 22002201 #
D0 M2B1 E R✱8 E
1 2E 0
EC 2 0S S✱
, 2R0I C O NE RTIECCSHSNOONLTOEGCYH RNEOVLIO
EWGY REVIEW 47
7
✱ AI IN RAN AI IN RAN ✱

Trained DNN Optimized DNN

RL framework RL models RL framework

Store Read
Generic DNN optimizations RL agents RL agents

Quantization, distillation,
pruning, factorization DNN inference Data layer Observables Actions Observables Actions
(message bus) & rewards & rewards
Open source
software RAN edge hardware Trace Actuation
Target-specific optimizations aggregation control

GPU + x86 Instruction selection, DSP - Core network

parallelization, memory handling, Log data Actuation Log data Actuation
aggregation realization aggregation realization - Radio units Control
hardware accelerators ... Traces
- Channel emulators signals
Ericsson RAN simulator Ericsson RAN simulator - Mobility simulators
- Traffic generators
342 0.643662 -5.72 -2.3 3 Latency, memory and
- Mobile devices Trace Actuation
accuracy measurements 1. RL model derivation generation realization
533 0.733400 -3.13 3.4 4 Execution constraints:
390 1.210092 0.22 3.4 3 latency, number of DSPs, memory use using exhaustive RAN simulations RAN testbed
Ericsson eNodeB
environment
472 0.351166 -4.21 2.5 1

Data set Software 2. RL model evaluation in RAN testbed

developer

Figure 3 DNN inference optimization for RAN edge hardware Figure 4 AI-accelerated RAN platform

can continuously evolve to form a learning system. The optimized DNN was realized in terms supports the use of RL techniques and model transfer isolation from the RAN application, using RL domain
The digital sibling is a unique part of our framework, of digital signal processor-C (DSP-C) code – to control parts of the RAN functionality. Our second experts, and with required compute and memory
which will ultimately enable advanced automation. that is, C extended with some specific DSP proof of concept explores phases 1,2 and 3 in Figure 2. resources, including the use of powerful GPUs to train
keywords and instructions, taking into account We have used sophisticated RAN simulators to set the DNNs that are used by the RL agents.
Proof of concept #1 – deep neural network constraints given by the target RAN hardware up and run many different deployment scenarios to After achieving high performance in the different
inference optimization and the RAN software designers, such as the train the RL models. The RAN test bed we used simulator scenarios, the simulator-trained RL models can
In the proof of concept shown in Figure 3, we have number of DSPs and the amount of memory consists of several cells and multiple UEs be used for automation use cases in real-life RAN contexts.
investigated how to tune, optimize and realize available for the DNN inference. Latency and (user equipment). The data layer uses a message bus,
pretrained DNNs for inference on RAN edge memory measurements were made on the RAN which organizes, aggregates and filters data. Conclusion
hardware. We selected this example because radio- edge hardware platform for the generated DNN To obtain a high degree of scalability and improve Growing complexity and the need to solve repetitive
and baseband data processing are time-critical and DSP-C code. We also tuned how to best parallelize model generalization and RL exploration possibilities, tasks in 5G and future radio systems necessitate new
computationally intensive tasks. These steps would the DNN inference over several DSPs. the RAN simulators have been containerized, automation solutions that take advantage of state-
be part of phase 2 in Figure 2. Based on performed evaluations, we concluded allowing many concurrent instances to execute of-the-art artificial intelligence and machine learning
The DNN training was done in a container-based that DNN inference for the target hardware can with different configurations. techniques that boost system efficiency. At Ericsson,
environment (Kubernetes) using GPUs and state-of- benefit significantly from the generic- and target- The RAN simulator can extract observables and we have identified the key software enablers for
the-art open source DNN training frameworks. specific DNN optimizations implemented. However, rewards based on internal simulator data, as well as AI- based RAN automation and integrated them into
Several generic DNN optimization techniques were there is normally a trade-off between ML model change its internal state and resource allocation a comprehensive framework that provides a solid and
investigated and implemented, such as quantization, accuracy, execution time and memory consumption. strategies based on the actions provided by the RL flexible technological foundation for the development
distillation and pruning, with the goal to reduce the agents. The RL agent framework is separated from of AI-based RAN. Our two proof-of-concept examples,
footprint and execution time of the DNN inference, Proof of concept #2 – AI-accelerated both simulated and actual RAN application by the executed in real RAN contexts, clearly illustrate the
still keeping the accuracy close to the accuracy of the RAN platform same well-defined data and action interfaces. RL benefits of our framework of software enablers in
originally trained DNN. The proof-of-concept platform shown in Figure 4 algorithms can be developed and evaluated in building the next generation of automated RAN systems.

48
8 E R I C S S O NE RTIECCSHSNOONLTOEGCYH RNEOVLIO
EWG Y✱R EDVEICEEWM ✱
B E R# 80 ,1 22002201 #
D0 M2B1 E R✱8 E
1 2E 0
EC 2 0S S✱
, 2R0I C O NE RTIECCSHSNOONLTOEGCYH RNEOVLIO
EWGY REVIEW 49
9
✱ AI IN RAN AI IN RAN ✱

Andreas Ermedahl
References ◆ joined Ericsson in 2010.

the authors
1. Ericsson Technology Review, 5G New Radio RAN and transport choices that minimize TCO, November 7, He works as a software
2019, Eriksson, A-C; Forsman, M; Ronkainen, H; Willars, P; Östberg, C., available at: researcher within the
https://www.ericsson.com/en/reports-and-papers/ericsson-technology-review/articles/5g-nr-ran-and- Networks division, where he
transport-choices-that-minimize-tco is responsible for driving a
2. Ericsson Technology Review, Enhancing RAN performance with AI, January 20, 2020, Calabrese, F.D.; software program research
Frank, P; Ghadimi, E; Challita, U; Soldati, P., available at: https://www.ericsson.com/en/reports-and-papers/ area targeting AI and ML in
ericsson-technology-review/articles/enhancing-ran-performance-with-ai RAN. His research interests
3. MIT Press, Deep Learning, 2016, Goodfellow, I; Bengio, Y; Courville, A., available at: Diarmuid Corcoran include static program Catrin Granbom
https://www.deeplearningbook.org/ ◆ joined Ericsson in 1992. analysis, optimization ◆ has worked within
He currently works within methods and software Ericsson’s telecom and
4. MIT Press, Reinforcement Learning: An Introduction (second edition), 2018, Sutton, S; Barto, A.,
the Networks division of techniques for facilitating information technology area
available at: http://incompleteideas.net/book/RLbook2020.pdf
Ericsson as an expert in ML and AI in RAN. Ermedahl since 1987, most recently
5. Ericsson Technology Review, Cognitive technologies in network and business automation, June 28, software architecture, received a Ph.D. in computer with 4G and 5G. In recent
2018, Niemöller, J; Mokrushin, L., available at: https://www.ericsson.com/en/reports-and-papers/ericsson-
actively driving and systems from Uppsala years, her focus has been
technology-review/articles/cognitive-technologies-in-network-and-business-automation
participating in software- University in Sweden and on research close to
6. KTH Royal Institute of Technology, Squeezing and Accelerating Neural Networks on Resource related activities across earned the title of docent implementation, including
Constrained Hardware for Real Time Inference, 2018, Presutto, M., available at: the company. Corcoran in computer science at collaboration ties to
http://www.diva-portal.org/smash/get/diva2:1299467/FULLTEXT02.pdf has contributed important Mälardalen University in academia supporting this.
software and system Sweden. He currently holds She is currently responsible
solutions to five generations an adjunct professorship in for software technology
of mobile technology, computer science at research activities within
Further reading including 5G. His current KTH Royal Institute of the Networks division,
❭ Ericsson, How will AI enable the switch to 5G?, available at: https://www.ericsson.com/en/networks/offerings/ research interests include Technology in Stockholm. including software
network-services/ai-report methods, architectures implementation technology
❭ Ericsson, Autonomous networks, available at: https://www.ericsson.com/en/future-technologies/autonomous- and algorithms for enabling in areas such as cloud RAN,
networks self-learning resource AI software in RAN, test and
management solutions quality, high-performance
❭ Ericsson, Ericsson launches unique AI functionality to boost radio access networks, available at:
with a focus on advanced software and software for
https://www.ericsson.com/en/news/2019/10/ericsson-ai-to-boost-ran
automation. Corcoran cyber-security. Granbom
❭ Ericsson, AI by Design, available at: https://www.ericsson.com/en/ai-and-automation
holds a B.Eng. in computer holds an M.S. in computer
❭ DeepMind blog, AlphaGo Zero: Starting from scratch, available at: https://deepmind.com/blog/article/ engineering from the science from Uppsala
alphago-zero-starting-scratch University of Limerick, University.
Ireland.

50
10 E R I C S S O NE RTIECCSHSNOONLTOEGCYH RNEOVLIO
EWG Y✱R EDVEICEEWM ✱
B E R# 80 ,1 22002201 #
D0 M2B1 E R✱8 E
1 2E 0
EC 2 0S S✱
, 2R0I C O NE RTIECCSHSNOONLTOEGCYH RNEOVLIO
EWGY REVIEW 51
11
✱ 5G SYNCHRONIZATION 5G SYNCHRONIZATION ✱

5G synchronization
requirements
Synchronization for new 5G applications RAN network synchronization

Smart industry Positioning Smart power grids

and solutions
Time-sensitive networking

Solution toolbox for network synchronization

The need for synchronization in the RAN has grown as new radio GNSS OAS Frequency over transport Time/phase over transport Clocks

technologies and network architectures emerge to boost efficiency

and support demanding 5G use cases. Although the fundamental Figure 1 Synchronization in new network scenarios
synchronization requirements have not become more stringent in 5G,
the need for time synchronization has become much more critical. the distributed unit (DU) and the radio unit (RU). The two main types of synchronization
The baseband function in a base station is split into requirements that are relevant for 5G networks
two logical units: a CU hosting the higher-layer are those that depend on the radio network operation
protocols and a DU handling the lower layers and those that depend on the supported services
to the user equipment (UE). (application-driven requirements).
5G makes it possible to serve several applications
Accurate and reliable synchronization has Many of the commercial 5G networks going live that have previously been supported either by Radio network-driven synchronization
STEFANO RUFFINI,
long been a fundamental prerequisite for the around the world today use TDD. TDD radio frames wireline or by non-standard radio technologies, requirements
MIKAEL JOHANSSON,
BJÖRN POHLMAN, correct operation of telecommunications inherently require time and phase alignment such as Critical Internet of Things (IoT) and Synchronization requirements related to
MAGNUS SANDGREN networks. Its importance has grown in 4G, between radio base stations, to prevent interferences Industrial Automation IoT services [1] for which communication over the radio interface can be
and it will be more important than ever in and related loss of traffic. Time synchronization is time synchronization is fundamental. While divided into two categories: TDD cell phase
5G and future networks. also required in FDD networks when different radio many applications benefit from accurate time synchronization (Tsync) and communication features
coordination features are used. synchronization, it is important to realize that high based on coordinated transmission or reception from
■ RANs are designed to optimize service New network architectures imply new approaches time accuracy over large areas can be very costly. multiple Transmission Reception Points (TRxPs).
performance and reliability, and synchronization to addressing the various synchronization needs. Services with special accuracy or availability
is a vital enabler. The RANs use different sources, This includes using Precision Time Protocol (PTP) demands should be carefully analyzed to TDD cell phase synchronization
often in combination, to provide synchronization and radio interface-based methods to synchronize minimize costs. TDD cells operating at the same frequency (or adjacent
characteristics to match the network use case distributed radio units in the evolved RAN Figure 1 provides some examples of frequencies) in overlapping coverage areas require
and services delivered. Finding the right balance architecture, where the upper and lower parts of synchronization in new network scenarios. time domain isolation to prevent base-station-to-
between timing accuracy, availability and cost the 5G New Radio (NR) RAN are separated in the Time and phase synchronization is particularly base- station and UE-to-UE radio frequency (RF)
is key to making services successful. different logical units: the centralized unit (CU), important in 5G networks. interference.

52
2 E R I C S S OENR ITCESCSHONNOTLEOCGHYNROELVOI G
EWY R✱E VJ A
I ENWU A✱R Y#1031, 2 0 2 1 #
J A0N1 U2A
0R2 1Y 1✱
3, E
20 1 S✱
R2I C S OENR ITCESCSHONNOTLEOCGHYNROELVOI G
EWY REVIEW 53
3
✱ 5G SYNCHRONIZATION 5G SYNCHRONIZATION ✱

can be used. To decrease overheads, the emphasis is Communication features based on coordinat-
on striving to keep the guard periods short, while ed transmission or reception from multiple
Guard period Transmission Reception Points
still catering for the desired effects. Allocation of a
budget to the different components of the guard A variety of features that benefit from coordinated
Tprop_UE period is the result of a trade-off between cost transmission or receptions from multiple TRxPs have
(product and deployment), availability, TDD been standardized over the years, all with different
@ BS A DL DL UL DL periodicity and overheads. purposes and characteristics. Some relate to
A
In 3GPP New Radio (NR), cell phase combining spectrum assets, thereby allowing total
A TOff On synchronization is specified as 3µs [2] – that is, the higher aggregated bandwidth and throughput
TOn Off same as for LTE. This is because the reduced (carrier aggregation, dual connectivity and so on),
Tprop_BS2BS TSync
TSync transient times in NR made it possible to keep the while others relate to improving link performance
B
same synchronization requirement with low overhead. at the cell edge (variants of coordinated multi-point
B The requirement on cell phase synchronization is operation, for example). Still others relate to specific
@ BS B DL DL UL DL
ultimately specified in terms of maximum deviation services like Multimedia Broadcast Multicast Service
in relation to a common absolute timing requirement over a Single Frequency Network. Coordination may
Tprop_UE and dividing the requirement by half (±1.5µs, for also occur between NR and LTE.
RF propagation delay example). This allows a design with independent These features and related timing requirements are
300m~1us Timing advance synchronization references. applicable within a single operator network, and as a
In cases where the TDD-unicast area is not result, control of relative time error between antennas
isolated, the 3GPP has specified that traceability to a used by the feature is sufficient. What matters is the
@ UE DL UL standard timing reference (UTC) is required [3]. relative frame timing alignment at the receiver. The
This prevents interference between different 3GPP has therefore defined a maximum receive timing
B TSync TOff On networks using adjacent frequency bands or difference (MRTD) as a maximum relative receive
between national borders, because it enables the timing difference the UE must be capable of handling.
phase alignment of the radio frames generated by The MRTD consists of a base-station relative time
different networks with overlapping areas. alignment error (TAE) and an RF propagation delay
UTC traceability does not imply that UTC time difference (ΔTprop). That is, MRTD = TAE + ΔTprop.
TRX TX (total transition time is recovered or that leap seconds are used; in fact, Based on the level of required MRTD, three main
between RX and TX mode) the 3GPP also requires the use of continuous time categories could be identified:
without leap seconds. In practice, a base station
❭ MRTD as a fraction of a cyclic prefix (CP)
with a Global Navigation Satellite System (GNSS)
❭ MRTD without CP relation
Figure 2 Synchronization for TDD receiver fulfils the requirement on UTC traceability
❭ no timing requirement.
as specified by the 3GPP. The ITU-T (International
There are two requirements for time domain isolation: ❭ air propagation time (Tprop) Telecommunication Union Telecommunication In the case of MRTD as a fraction of the CP, the
❭ the cells must use the same TDD configuration ❭ sufficient transient time when the transmitter Standardization Sector) specification of the primary remaining duration of the CP allows for channel
❭ the deviation in frame start timing between cells changes between defined ON/OFF power levels reference time clock (PRTC) includes the possibility delay spread. In 3GPP TS 38.104 [10], the TAE ranges
must be below a maximum value specified as cell (TOn Off , TOff On ) for the PRTC to be synchronized with GNSS. between 65ns to 260ns depending on the feature and
phase synchronization accuracy in the 3GPP. ❭ sufficient time for the UE and base station to This means that a PTP-based solution can also on the CP duration but is only valid for colocated/
change between transmit and receive modes fulfill this requirement. intra-site deployments where ΔTprop~0. The strict
For TDD synchronicity and interference, the critical (TRX TX, TX TRX) MRTD, and thereby TAE, relates to features using
points are when switching between transmission and ❭ allocate margin for cell phase synchronization THIS ALLOWS A DESIGN the same frequencies (like MIMO (multiple-input,
reception, as shown in Figure 2. Guard periods are errors (Tsync). WITH INDEPENDENT multiple-output)) or at adjacent spectrum (like contiguous
used for isolation with a configurable total guard time carrier aggregation) where common or shared
expressed as an integer number of symbols. The duration Since guard periods are not used for communication, SYNCHRONIZATION functions in the RF chain could cause strict timing
of the guard period needs to cater for four effects: they reduce the time ratio when spectrum resources REFERENCES dependencies.

54
4 E R I C S S OENR ITCESCSHONNOTLEOCGHYNROELVOI G
EWY R✱E VJ A
I ENWU A✱R Y#1031, 2 0 2 1 #
J A0N1 U2A
0R2 1Y 1✱
3, E
20 1 S✱
R2I C S OENR ITCESCSHONNOTLEOCGHYNROELVOI G
EWY REVIEW 55
5
✱ 5G SYNCHRONIZATION 5G SYNCHRONIZATION ✱

cost-efficient solutions. For example, 1m corresponds

OAS METHODS to a synchronization accuracy of around 3ns,
A
Tprop_A1 Tprop_B1 B PROVIDE AN ATTRACTIVE which is orders of magnitude more stringent than
other radio network synchronization requirements.
SYNCHRONIZATION A method that depends on such a high level of
UE1 ALTERNATIVE synchronization accuracy would significantly
Tprop_B2 increase the cost of the solution, especially when
Time-sensitive networks synchronization to the base stations is distributed
Tprop_A2 TSNs in industrial applications generally require the over the network. Over-the-air synchronization
D E
distribution of time synchronization to the industrial (OAS) methods provide an attractive
Tprop_UE1 = Tprop_A1 - Tprop_B1 subnetworks used for functionality such as robot synchronization alternative in this case.
Tprop_A3 control or autonomous vehicles. This results in
= Tprop_A2 - Tprop_B2 synchronization requirements in the microsecond Evolving network synchronization
Tprop_UE2
range between nodes in the TSN. standards and technologies
Tprop_D4 Tprop_E4 UE2 The 3GPP technical specification TS 23.501 [4] Most of the relevant synchronization requirements
Tprop_UE3
= Tprop_A3 - Tprop_C3
presents an architecture and methods that enable are defined by the 3GPP. The fundamental
Tprop_C3 the successful integration of a 5G network into a technologies to address these requirements come
Tprop_UE4
= Tprop_D4 - Tprop_E4 TSN synchronization network. Industrial from the GNSS, which provides the basic technology
UE3 automation scenarios often involve multiple timing for a time synchronization master, enabling global
domains. The recommended solution is to relay distribution of a UTC-traceable reference. GNSS
UE4 C
timing transparently across the 5G network, where services include GPS, BDS (the BeiDou Navigation
Tprop_UE4 <
the delay of the TSN-time messages through the 5G Satellite System) and Galileo [6], and are offered by
Tprop_UE1 and Tprop_UE2 < Tprop_UE3
network is measured and compensated. The same several space agencies. GNSS is based on satellites
architecture can also be used to support the having a known time and orbit position, transmitting
Figure 3 Synchronization for coordinated transmission/reception industrial application by distributing the timing messages whose arrival are measured by the GNSS
of the 5G network to the TSN [5]. receiver. GNSS satellites visibility is important for
proper operation.
Even if features in this category can be served from For example, synchronous dual connectivity Smart grid applications Specified by IEEE 1588, PTP has been selected
distributed TRxPs (where generally ΔTprop≠ 0) allows an MRTD of up to 33µs, of which 30µs When 5G is used to support smart grid applications, by the telecom industry to distribute accurate time
– that is, not being restricted to a colocated/inter-site can be allocated for ΔTprop and the remaining the 3GPP synchronization architecture [4] can be used synchronization from an accurate master, such as
deployment scenario – it would be difficult to specify 3µs for TAE. to support the distribution of timing to the synchro- GNSS. The basic concept is to distribute time
and mandate a single fixed TAE. Rather, the 3GPP-defined asynchronous dual connectivity phasors that in this way can measure the phase synchronization from the PTP “grandmaster” to the
required TAE would depend on actual deployment, without TAE requirements and ΔTprop restrictions relationships in the AC power distribution network. leaf PTP clocks, using PTP messages. The protocol
its RF characteristics and the UEs’ relative position is an example of a feature with no timing requirement. These applications require synchronization is based on a two-way exchange of timing messages
between the TRxPs, as shown in Figure 3. accuracy in the same range as the TSN. [7]. This message exchange enables the distribution
As noted in the bottom of the figure, ΔTprop_UE1 Application-driven synchronization In this case, the distribution of a single time of the time from the PTP grandmaster and estimates
and ΔTprop_UE4 are less than ΔTprop_UE2, which requirements reference is sufficient. the path delay. The path delay estimation assumes
is less than ΔTprop_UE3. For the same delay spread, While the introduction of 5G did not cause that the delay of the two transmission directions
UE1 and UE4 can tolerate larger TAE than UE2 any fundamental change to radio network Positioning of UE devices is symmetric. This is not always the case in real
and UE3. For colocated D and E, the TAE of D-E synchronization requirements, some applications Some approaches to solving the UE device- networks due to factors such as different fiber length
is generally less than the TAE of A-B. may put more stringent local accuracy requirements positioning use case rely on accurate or the use of different wavelengths in the optical
In contrast, features in the “MRTD without on the synchronization of the 5G nodes. synchronization of the radio nodes with the help transmission. Since any source of asymmetry results
CP relation” category allow a relatively large Examples include time-sensitive networks of methods such as OTDOA (Observed Time in error in the time synchronization distribution, the
part for ΔTprop and thereby more flexible (TSNs), smart grid applications and the Difference of Arrival). Depending on the required asymmetry must either be known and compensated
deployments, such as heterogeneous networks. UE device-positioning use case. positioning accuracy, these may not always result in for or sufficiently small to be negligible.

56
6 E R I C S S OENR ITCESCSHONNOTLEOCGHYNROELVOI G
EWY R✱E VJ A
I ENWU A✱R Y#1031, 2 0 2 1 #
J A0N1 U2A
0R2 1Y 1✱
3, E
20 1 S✱
R2I C S OENR ITCESCSHONNOTLEOCGHYNROELVOI G
EWY REVIEW 57
7
✱ 5G SYNCHRONIZATION 5G SYNCHRONIZATION ✱

To limit the effects of asymmetry and packet delay station must always be better than ±1.1µs when network without synchronization support from the One promising method for time alignment
variation that could be created by the network nodes measured against a standard time reference. transport network. The two main RAN-based of the radio network is to use information
(switches and routers, for example), IEEE 1588 has A margin of 400ns is left to the end application to solutions are GNSS-based synchronization and over the NR air interface. OAS uses timing
specified the “boundary clock” and “transparent fulfill ± 1.5µs on the radio interface. Synchronization over-the-air synchronization (OAS). information sent over the radio interface between
clock” functions with event message timestamping methodologies and architectures applicable to a A GNSS-based solution installed directly at base neighboring base stations. Features requiring
in the hardware layer. switched fronthaul have also been covered in station sites can provide cost-efficient, accurate and stricter synchronization are generally implemented
The distribution of time-phase and frequency G.8271.1 [8]. predictable time synchronization of the radio between neighboring base stations. OAS is based
synchronization has been addressed in several The synchronization aspects of fronthaul networks network without any support from the transport on round-trip time measurement similar to PTP,
ITU-T recommendations [8]. In particular, these are also addressed by the O-RAN Alliance. In network. The higher accuracy of GNSS compared with the advantage that the asymmetries are
recommendations define reference synchronization particular, one of the groups in O-RAN addresses with other synchronization sources allows for generally small over the air, making it possible
networks, where the synchronization is generated by the split between the DU and RU – O-DU and O-RU comparatively longer holdover periods. Moreover, to achieve good time alignment.
time synchronization masters or PRTCs, which are in the O-RAN terminology. The split option is the time error can be better estimated during
typically based on GNSS technology and where the denominated as 7-2x and it is an intra-PHY split – holdover, making it possible to take consequent Transport-based synchronization solutions
reference timing signal is carried across a network that is, some parts of the 3GPP PHY layer are actions per service (TDD, for example). Transport-based solutions, in which synchronization
of clocks. Frequency synchronization in these implemented in the DU and others in the RU [9]. is distributed over the transport network, rely on two
key technologies: frequency synchronization over
reference networks is carried over the physical layer
(typically using synchronous Ethernet). The time
To address the synchronization needs in these
scenarios, the O-RAN specification defines different
TRANSPORT-BASED the physical layer (synchronous Ethernet) and
synchronization reference is carried through a PTP, synchronization topologies. The DU may be part SOLUTIONS ... RELY frequency and time synchronization over the
where the PRTC is the source of time for the PTP of the synchronization chain or just a consumer of packet layer (PTP).
Grandmaster, also known as the Telecom the synchronization. The RU is the main user of
ON TWO KEY Frequency synchronization based on packet
Grandmaster (T-GM). synchronization and it may also be synchronized TECHNOLOGIES protocols like Network Time Protocol (NTP) or PTP
by a timing reference generated directly at the radio using the G.8265.1 profile (in a PTP-unaware
IN MANY CASES, A site. The data carried between a DU and an RU uses Short disruptions due to interference and network) has proven to be the most cost-efficient
Ethernet-based transport. The time and frequency blocking are common, but the negative impact can solution in LTE-FDD networks, where frequency
COMBINATION OF TECHNIQUES synchronization distribution can be achieved using be avoided or limited by the base station’s holdover synchronization alone was often sufficient. In 5G
IN BOTH DOMAINS WILL BE a combination of PTP and synchronous Ethernet. using the local oscillator – up to a few hours for a networks the situation is different because time and
There are no specific synchronization requirements TDD base station, for example. Longer holdover phase synchronization is more often needed
REQUIRED for the CU. periods can be achieved when GNSS is combined compared with previous generations of radio
with other technologies that provide a stable timing networks, due to factors such as wider use of TDD or
Two PTP profiles have been defined for the use Recommended synchronization solutions reference (traceable to PRTC). the need for coordination between NR and LTE.
of PTP in telecom – G.8275.1 (PTP with full timing for evolving mobile networks
support from the network) and G.8275.2 (PTP We have compiled a comprehensive toolbox
with partial timing support from the network) [8]. of synchronization techniques that can be
The purpose of the G.8275.1 profile is to meet the used to provide appropriate frequency Terms and abbreviations
highest accuracy, which requires the implementation synchronization and time alignment of the radio APTS – Assisted Partial Timing Support | CP – Cyclic Prefix | CPRI – Common Public Radio Interface |
of an IEEE 1588 telecom boundary clock or telecom network with minimal disruption. Some of them CU – Centralized Unit | DL – Downlink | DU – Distributed Unit | eCPRI – Enhanced CPRI | ePRTC –
transparent clock in every node in the timing are implemented in the RAN domain, while others Enhanced Primary Reference Time Clock | FTS – Full Timing Support | GNSS – Global Navigation
distribution network. Assisted partial timing are implemented in the transport domain. In many Satellite System | IOT – Internet of Things | ITU-T – International Telecommunication Union
support (APTS), which concerns the use of PTP cases, a combination of techniques in both domains Telecommunication Standardization Sector | MRTD – Maximum Receive Timing Difference | NR – New
as a backup for GNSS, is an important will be required to create a robust and reliable Radio | NTP – Network Time Protocol | OAS – Over-the-Air Synchronization | PRTC – Primary
implementation of the G.8275.2 profile. solution. Reference Time Clock | PTP – Precision Time Protocol | RF – Radio Frequency | RU – Radio Unit |
To meet the TDD “relative” synchronization RX – Reception | T-GM – Telecom Grandmaster | TAE – Time Alignment Error | TSN – Time-Sensitive
requirement that cell phase synchronization RAN-based synchronization solutions Network/Networking | TX – Transmission | UE – User Equipment | UL – Uplink | UTC – Coordinated
must be better than 3µs, the accuracy of the time A solution is considered to be RAN based if it can Universal Time
synchronization reference at the input to the base fulfill the synchronization requirements of the RAN

58
8 E R I C S S OENR ITCESCSHONNOTLEOCGHYNROELVOI G
EWY R✱E VJ A
I ENWU A✱R Y#1031, 2 0 2 1 #
J A0N1 U2A
0R2 1Y 1✱
3, E
20 1 S✱
R2I C S OENR ITCESCSHONNOTLEOCGHYNROELVOI G
EWY REVIEW 59
9
✱ 5G SYNCHRONIZATION 5G SYNCHRONIZATION ✱

specific network may depend on existing and PRTC/enhanced PRTC traceable synchronous
synchronization feature support of the network and Ethernet. The time period a clock can remain within
network elements, transport network modernization the required accuracy depends on the stability of the
plans that improve synchronization support, as well assisting source.
GNSS Router Router
ePRTC
as operator preferences. Possible future regulations
Primary (primary) of national authorities may also be relevant. THE TIGHTEST
RU Transport domain In many cases, synchronization in a mobile
OAS
Backup without PTP support GNSS
network will benefit from the implementation of a REQUIREMENTS SHOULD
RU
DU Router Router ePRTC
combination of methods. As shown in Figure 4, BE ADDRESSED LOCALLY
a fairly typical 5G mobile network can combine
(backup)
methods like GNSS and OAS at base station sites, ONLY WHEN AND WHERE
time and frequency distribution via the transport REQUIRED
OAS
RU GNSS network, and redundancy and holdover functions
Router Router in the nodes. This is particularly relevant for cases Conclusion
RU
that require GNSS redundancy (to protect against Proper network synchronization is a prerequisite
CPRI/eCPRI DU jamming or spoofing events, for example), where the to excellent radio network performance. Some of
OAS
distribution of timing over the transport network the most compelling use cases for 5G, including
ePRTC
RU (primary) as a complement to a local GNSS receiver should industrial automation, depend on more accurate
Transport domain with
be considered. timing and will likely generate additional
RU GNSS
PTP support PTP with “full timing support from the network” synchronization requirements in the near future.
DU
should be used to meet the relevant time While the level of the required synchronization
ePRTC
synchronization requirements. Redundancy is accuracy depends on several factors, it would be
TSN Router Router Router (backup) achieved through a PTP network that is fed from a mistake to apply the tightest synchronization
RAN domain Transport domain T-GM geographically redundant T-GMs and distributes requirement as a general 5G requirement,
timing over the same physically redundant as doing so would make the cost for 5G and the
topologies that are used for user traffic. future evolution of the mobile technologies
The best control over the synchronization unsustainable. The tightest requirements
PTP - G.8275.1 session (FTS) PTP - G.8275.2 session (APTS) GNSS input OAS
network architecture will be achieved by rolling should be addressed locally only when and
out PTP from the lower part of the transport where required.
Figure 4 Recommended synchronization solutions network, with GNSS receivers and T-GM An efficient synchronization solution that
functionality in base stations or routers, and be addresses all relevant scenarios requires the support
In these cases, the delay asymmetries and packet has the benefits of providing the same level of extended to include T-GMs that are higher up of a toolbox of synchronization methods that are
delay variation present in PTP-unaware networks robustness and redundancy for timing as for the user in the network when feasible. implemented in the RAN domain (GNSS and OAS),
would make it impossible to meet the stringent traffic itself. Another important aspect of the PTP In parts of the network where transport does not the transport domain (such as PTP), or both.
accuracy and stability requirements. To mitigate network planning is time error budget planning, support PTP, PTP should not be used as the primary For many features, what matters more than
such effects, the transport network requires PTP- which depends on the dimension of the network and time source for RAN nodes due to the limitations on anything else is the relative time error between
aware network elements such as boundary clocks or the accuracy of the PTP-aware network elements. the achievable performance. In these cases, it should neighboring base stations. In this context,
transparent clocks as well as support of PTP across There are guidelines for this in G.8271.1 [8]. only be used as an assisting reference in APTS mode over-the-air synchronization (OAS) is a
the entire time distribution chain (“full timing for GNSS. In these parts of the network, OAS powerful tool.
support from the network” as described in G.8275.1). Combining techniques for best results provides an alternative solution for synchronizing
An important aspect of this is that a base station may Several aspects need to be considered when the radio nodes without requiring the support of a
use PTP for the delivery of both time selecting the most appropriate synchronization transport network.
synchronization and frequency synchronization. solution(s), including installation and operation When the GNSS reference is lost, the holdover
Distributing time synchronization over the same costs, synchronization accuracy, robustness and time can be prolonged by using assisting
transport network infrastructure used for user data availability targets. The optimal solution for a synchronization sources such as APTS, OAS

60
10 E R I C S S OENR ITCESCSHONNOTLEOCGHYNROELVOI G
EWY R✱E VJ A
I ENWU A✱R Y#1031, 2 0 2 1 #
J A0N1 U2A
0R2 1Y 1✱
3, E
20 1 S✱
R2I C S OENR ITCESCSHONNOTLEOCGHYNROELVOI G
EWY REVIEW 61
11
✱ 5G SYNCHRONIZATION 5G SYNCHRONIZATION ✱

References
1. Ericsson Technology Review, Critical IoT connectivity: Ideal for time-critical communications, June 2,
2020, Alriksson, F; Boström, L; Sachs, J; Wang, Y.-P. E; Zaidi, Ali, available at: https://www.ericsson.com/en/
reports-and-papers/ericsson-technology-review/articles/critical-iot-connectivity
2. 3GPP technical specification TS 38.133, NR; Requirements for support of radio resource
management, available at: https://portal.3gpp.org/desktopmodules/Specifications/SpecificationDetails.
aspx?specificationId=3204
3. 3GPP technical specification TS 38.401, NG-RAN; Architecture Description, available at:
https://portal.3gpp.org/desktopmodules/Specifications/SpecificationDetails.aspx?specificationId=3219 transmission accuracy for

the authOrs
4. 3GPP technical specification TS 23.501, System architecture for the 5G System (5GS), available at: 20 years. He joined Ericsson
https://portal.3gpp.org/desktopmodules/Specifications/SpecificationDetails.aspx?specificationId=3144 in 2005 and has served
as a member of the
5. Ericsson Technology Review, 5G-TSN integration meets networking requirements for industrial
ORAN Working Group 4
automation, August 27, 2019, Farkas, J; Varga, B; Miklós, G; Sachs, J, available at: https://www.ericsson.
(Open Fronthaul) with
com/en/reports-and-papers/ericsson-technology-review/articles/5g-tsn-integration-for-industrial-automation
standardization since
6. ITU Technical Report TP-GSTR-GNSS – Considerations on the use of GNSS as a primary time reference
Ericsson joined the ORAN
in telecommunications, 2020, available at: https://www.itu.int/pub/T-TUT-HOME-2020
Alliance in 2019. He holds
7. IEEE 1588-2019 – IEEE Standard for a Precision Clock Synchronization Protocol for Networked Stefano Ruffini area for 12 years, with a an M. Sc. in electrical
Measurement and Control Systems, June 16, 2020, available at: https://standards.ieee.org/ ◆ joined Ericsson in 1993. special focus on packet- engineering from KTH Royal
standard/1588-2019.html He is an expert at Ericsson based methods (PTP and Institute of Technology,
8. ITU-T Recommendations G.826x and G.827x series (G.8200-G.8299: Synchronization, quality and Research working on NTP). He also contributes Sweden.
availability targets, available at: https://www.itu.int/ITU-T/recommendations/index_sg.aspx?sg=15 synchronization and actively to ITU-T SG15 Q13
9. 3GPP technical specification TR 38.801, Technical Specification Group Radio Access Network; Study on transport solutions. He is and IEEE 1588. Johansson
new radio access technology: Radio access architecture and interfaces, available at: https://portal.3gpp. currently an active holds an M.Sc. in applied
org/desktopmodules/Specifications/SpecificationDetails.aspx?specificationId=3056 contributor to ITU-T SG15 physics and electrical
Q13 (serving as rapporteur) engineering from the
10. 3GPP technical specification TS 38.104, NR; Base Station (BS) radio transmission and
and other relevant Linköping Institute of
reception, available at: https://portal.3gpp.org/desktopmodules/Specifications/SpecificationDetails.
synchronization Technology, Sweden.
aspx?specificationId=3202
standardization bodies, as
well as serving as the chair
of the International Timing Magnus Sandgren
and Sync Forum (ITSF). ◆ joined Ericsson in 2002
Ruffini holds a M.Sc. and currently works as a
Further reading in telecommunication researcher at Business Area
The authors would
❭ Ericsson white paper, 5G wireless access: an overview, available at: https://www.ericsson.com/en/reports- like to thank Jonas engineering from Sapienza Networks, where much
and-papers/white-papers/5g-wireless-access-an-overview Edstam, Anders University of Rome, Italy. of his time is spent on
❭ Ericsson Technology Review, Integrated access and backhaul – a new type of wireless backhaul in 5G, June Höglund, Mats analysis and solutions
23, 2020, available at: https://www.ericsson.com/en/reports-and-papers/ericsson-technology-review/articles/ Höglund, Gábor Mikael Johansson related to synchronization,
introducing-integrated-access-and-backhaul Kovács and Mårten ◆ joined Ericsson in 1997 Björn Pohlman including contributions
Wahlström for their and currently serves as a ◆ has worked with timing to 3GPP standardization.
❭ Ericsson white paper, Cellular IoT in the 5G era, available at: https://www.ericsson.com/en/reports-and-
contributions to system developer. He has and synchronization He holds an M.Sc. in
papers/white-papers/cellular-iot-in-the-5g-era
this article. worked with timing and distribution within the RAN electrical engineering from
synchronization in the RAN base station and with air Lund University in Sweden.

62
12 E R I C S S OENR ITCESCSHONNOTLEOCGHYNROELVOI G
EWY R✱E VJ A
I ENWU A✱R Y#1031, 2 0 2 1 #
J A0N1 U2A
0R2 1Y 1✱
3, E
20 1 S✱
R2I C S OENR ITCESCSHONNOTLEOCGHYNROELVOI G
EWY REVIEW 63
13
✱ APPLIED NETWORK SLICING APPLIED NETWORK SLICING ✱

Applied network Access Aggregation Edge DC Backbone Central DC

slicing scenarios in 5G
Shared cloud partition
Dedicated cloud partition
Dedicated Dedicated
Dedicated Dedicated Dedicated
RAN NFs
transport transport NFs
capabilities/
capabilities capabilities
NFs
Dedicated cloud partition
Shared cloud partition
Shared Dedicated
Shared Shared
RAN NFs Shared
transport transport
Areas capabilities/
of capabilities capabilities NFs Network
Network slicing enables new business opportunities across a wide range coverage
NFs
slices

of use cases and sectors by making it possible to create fit-for-purpose

virtual networks with varying degrees of independence. However, the Shared Shared Shared
Dedicated
RAN transport
diversity of new commercial and technical requirements has significant capabilities/
NFs
transport
capabilities capabilities
NFs

implications on how networks are built and managed. Access Aggregation

Figure 1 Deployment overview

these networks, network slicing is used to customize The enablers of network slicing
Private 5G networks – the new business a customer or provided as a managed service. and isolate slices per enterprise customer and use Figure 1 provides a deployment overview that
HENRIK BASILIER,
offerings that network slicing enables – Network slicing is used to customize the behavior for case/traffic type. The hybrid approach enables a illustrates the different ways of composing private
JAN LEMARK, ANGELO
CENTONZA, THOMAS
deliver functionality that extends well beyond different use cases/traffic types and to provide more flexible distribution of functionality, more 5G networks (network slices). To provide the best
ÅSBERG that of the current offerings that are typically isolation between them. A good example of a efficient use of infrastructure and improved mobility level of isolation, resources assigned to a network
based on existing public network services. standalone private 5G network is a dedicated in and out of the customer premises. A good example slice are ideally dedicated. Assuming that it is
For example, network slicing makes it solution deployed at a customer premises such as of a hybrid private 5G network is a manufacturing acceptable, some slices may share resources to
possible to create a private 5G network with a manufacturing plant or airport. Although these plant or airport where dedicated on-premises reduce cost. Distribution and coverage are
specific service characteristics as well as networks use 5G technologies, they are fully hardware is integrated with and reuses the public considered per slice. Some slices are local, while
varying degrees of security/isolation, independent and isolated from the public 5G infrastructure to improve service and cost-efficiency. others may be wider in reach. Some slices require
exposure, self-management, and so on. infrastructure.
Virtual private 5G networks, on the other hand,
■ There are three main approaches to offering and are provided on top of an infrastructure layer that is
delivering private 5G networks: the standalone shared with public services. Network slicing is used Terms and abbreviations
approach, the virtual approach and the hybrid to meet customization and isolation needs per use 5GC – 5G Core | AMF – Access and Mobility Management Function | BSS – Business Support Systems |
approach. Network slicing enables the customization case/traffic type as well as per enterprise customer. DC – Data Center | DNN – Data Network Name | DRB – Dedicated Radio Bearer | E2E – End-to-End |
of system behavior and the isolation of resources/ Public-safety and connected-car services that make EPC – Evolved Packet Core | MBB – Mobile Broadband | NF – Network Function | OSS – Operations
functions for specific services in all three of them. use of the public 5G infrastructure are examples of Support Systems | PDU – Protocol Data Unit | RRM – Radio Resource Management | RRP – Radio
Standalone private 5G networks are independent, virtual private 5G networks. Resource Partitioning | SLA – Service Level Agreement | SMF – Session Management Function |
on-premises deployments that have limited Hybrid private 5G networks are provided by S-NSSAI – Single Network Slice Selection Assistance Information | UP – User Plane | UPF – UP Function
interoperability with public networks. They may be combining infrastructure adopted for public services | WAN – Wide Area Network
sold through a mobile network operator, managed by with infrastructure at a customer’s premises. In

64
2 E RTIECCSHS N
ERICSSON OON LTOEG
CYH N
ROE VL IOEG
WY✱R EFVEIBERWU A✱R Y#1011, 2 0 2 1 F#E0B1 R2U
0A2 1R Y✱1 1 E 2 1S S✱O N
, 2R0I C E RTI ECCSH
SONO
N LTOEG
CY RO
HN E VL IOEG
WY R E V I E W 65
3
✱ APPLIED NETWORK SLICING APPLIED NETWORK SLICING ✱

RAN
DIFFERENT COMBINATIONS The 3GPP has defined enablers that can be used
keep sufficient RRPs to guarantee resource isolation
per slice where needed, while not impacting radio
to distribute data-plane NFs such as unified data
management (UDM) and unified data repository
OF ENABLERS WILL BE within a RAN to select appropriate functions and performance due to excessive partitioning. (UDR) to achieve full independence.
REQUIRED TO ENGINEER capabilities (such as policies) for network slices.
However, the selection of such functions and
Network slice mobility is also enabled by means of
signaling between RAN nodes of slice support per Transport
THE APPROPRIATE definition of capabilities relies on implementation. tracking area. A mobility function can also take Traffic flows from one network slice (or a group of
NETWORK SLICE(S) The most important enabler defined in the 3GPP is
that of associating each protocol data unit (PDU)
handover decisions on the basis of slice support at
the target RAN to achieve radio efficiency while
them) should be mapped into transport resources that
match the required Service Level Agreement (SLA)
session to a slice identifier known as a single network maintaining service continuity. for the slice or group of slices. It is important to take the
local network functions (NFs) – for latency reasons, slice selection assistance information (S-NSSAI) as capabilities and capacity of the transport
for example – while others do not. soon as a UE (user equipment) context is created. Core infrastructure into consideration when selecting which
There is no one-size-fits-all multi-tool for network The RAN reduces a PDU session into dedicated Core has several enablers, mainly defined by the enablers to use.
slicing. In fact, the ability to engineer network slices radio bearers (DRBs), which allows the RAN to 3GPP. These enablers make it possible to define There are multiple enablers in the transport
depends on an evolving toolbox of versatile enablers associate an S-NSSAI to each DRB and to select dedicated (or shared) user-plane, control-plane or domains to support network slicing use cases. For
in five areas: cloud infrastructure, RAN, core, NFs and capabilities to serve DRB traffic. As an even data-plane NFs at design time that steer the many use cases, it is sufficient to rely on QoS
transport, and operations support systems/business example of capability, a specific next-generation orchestration of the requested network slice at mechanisms (compare differentiated services) and IP
support systems (OSS/BSS). Depending on the node B central unit user plane (gNB-CU-UP) – instantiation time. Further, the enablers make it ranges as the entry point. This involves configuring
scenario, different combinations of enablers will be hosting PDCP (Packet Data Convergence Protocol) possible to dynamically decide to use dedicated (or end points in the RAN and core to map the traffic into
required to engineer the appropriate network slice(s). – may be selected for a given S-NSSAI to fulfill delay shared) user-plane and control-plane NFs based on preexisting transport capabilities. Additional marking,
and security requirements. policy at attach or session requests. such as IPv6 flow labels, can be used to enable the
Cloud infrastructure Specific layer 1/layer 2 configurations can be Decisions at design and instantiation times set the observability and mapping of individual slices. The
Cloud infrastructure provides great versatility with tailored for slice policies. A framework for Radio context in which the main parameter S-NSSAI transport network can be configured further to map
multiple enablers. The physical infrastructure can be Resource Management (RRM) policies is used to (network slice selection assistance information) and the slices to transport VPNs with traffic-engineered
managed, allowing servers, for example, to be allocate resources and assign QoS levels per slice. secondary parameter Data Network Name (DNN) characteristics. Dedicated transport VPNs can be
allocated to a network slice. Furthermore, the For example, the RRM function may use the Radio together with user identities will steer which used to satisfy slices with highly specific requirements.
virtualized infrastructure manager enables a set of Resource Partitioning (RRP) capability to allocate a dedicated or shared NFs that will be used at attach Coordinated management is essential between the
infrastructure resources to be shared by making use specific partition to a DRB associated to a slice, and session requests. RAN/core and the transport domain to ensure the
of traditional virtualization. Container-as-a-service according to its requirements. Such partitioning may The user-plane function (UPF) is the most valuable end-to-end (E2E) SLAs, which may include cross-
provides a more granular and dynamic approach to vary depending on slice requirements. Hard NF to dedicate, not only because of the general domain orchestration. The transport VPNs can be
sharing resources (by using Kubernetes, for partitioning restricts resource usage to a specific independency values (optimal redundancy level and tuned and weighed by the assurance capabilities of the
example). slice; soft partitioning allows resources to be used by no risk of interruption from other services), but also OSS. Transport and mobile network capabilities
The cloud infrastructure is distributed from any slice when they are not utilized by the slice that is because it ensures low latency through distributed should be harmonized to ensure that mobile network
central data centers (DCs) to the customer on nominally assigned to them; shared resources can deployment, which makes it possible for the user-data capabilities are not compromised by limitations in the
premises and does not have direct awareness of also be defined for resources accessible by all slices, traffic to stay close to customer premises. A dedicated transport network.
network slices. However, processes can ensure on demand. UPF also ensures that established sessions can
that a cloud platform can fulfill the requirements Further, prioritization between DRB traffic can be survive for a period of time, even when the connection Operations support systems/business support
associated to slices using identifiers of resources. achieved by means of QoS policies. Such policies allow to control-plane NFs is lost. systems
Network slices will have different needs in terms of for the differentiation of DRB traffic within a slice or The control-plane NFs are the second-most The enablers within the OSS/BSS area relate to the
isolation, distribution and resource guarantees. The between slices when shared resources are used. valuable NFs to dedicate, starting with the session management of service characteristics specified by the
cloud infrastructure can provide highly dedicated While it may seem logical to define RRP for each management function (SMF) and followed by the SLAs included in contracts. This results in a
resources to slices that need it, while other slices slice supported at the RAN, this is, in practice, access and mobility management function (AMF). requirement for assurance and analytics capabilities
share resources. The resources can be optimally suboptimal. There is a tradeoff in performance With dedicated distributed SMF and AMF, it is based on business policies, SLA fulfillment and
used without unnecessary tradeoffs, controlled between the gain of dedicating resources to specific possible to make changes to established sessions and operations. As both the specific KPIs and the approach
through orchestration and policies to satisfy the slice services and the overhead in maintaining establish new sessions for a period of time, even if to monitoring them will differ between slices, there will
demands of network slices. numerous resource partitions. The balance is to connection to data-plane NFs is lost. The final step is be a growing need for customization.

4
66 E R I C S S O NE RTIECCSHSN
OONLTOEGCYH R
N EOVLIOE G
WY✱R EFVEIBERWU A✱R Y#101 1, 22002211 F#E0B1 R2U0A
2R1 Y✱1 1 ,E2R0I 2C1S S✱O E
NRTI C
ECSH
SONN
O LTOE G
CYH NRO
EVLO
I EGWY R E V I E W 5
67
✱ APPLIED NETWORK SLICING APPLIED NETWORK SLICING ✱

Orchestration enables the automation of manual characteristics, particularly in terms of service

tasks. For example, if an enterprise requires services in coverage and enterprise integration depth. Network Campus Radio-access Campus On-premises DC WAN transport Central DC
a city location, orchestration automatically configures slices can be loosely grouped into three slicing site network
Hybrid private 5G networks for enterprises
all the cell sites that provide coverage in that location, categories: campus-based, wide-area and
generates the configurations required to meet limited-area. RAN
capabilities/
A. Network slice for
critical services
RAN
NF 5G Mgm
enterprise service SLAs and automatically provisions In campus-based scenarios, services are mainly NFs NFs Appl. Appl.

the respective cells by applying the configurations. consumed locally. In the case of a slice for smart
This process can adapt policies and NF selection manufacturing plant services, for example, only the RAN
capabilities/
B. Network slice for
latency-sensitive
RAN
NF 5G 5G
based on slice load and the number of slices supported base stations covering the plant need to support the NFs services NFs Appl. NFs

to deliver a solution that can react to operational plant slices. Core NFs can be deployed in edge DCs
conditions while fulfilling SLAs. Orchestration within or tightly integrated with the plant, typically RAN
capabilities/
C. Network slice for
non-critical services 5GC/EPC
provides faster service provisioning across all the together with RAN functions. Depending on the NFs NFs

nodes through the automation of every step, including complexity of the particular scenario, either a
instantiation and provisioning of all necessary standalone or hybrid private 5G network is likely to
RAN D/E. Network slices for other
network functions. be the best option in these cases. Though not capabilities/ (wide-area/limited-area) services
Transport
partition 5GC/EPC
As the use cases, deployment models and optimal, a virtual private 5G network would also be a NFs NFs

business models are diversified, it must be possible possibility.

to customize and repeat the actions of the OSS/BSS In wide-area scenarios, services are consumed in
layer, which drives the adoption of a model- and a large part of the network. A typical example of this Wide-area RAN
intent-driven approach, where templates and scenario would be services in the transportation/
policies dictate the actions. These templates reflect automobile sector that require wide-area coverage.
the SLAs and are used to orchestrate the From a radio perspective, this requires RAN slicing Figure 2 Example of network slicing in a campus deployment (hybrid private 5G network)
deployment of NFs, the system’s capabilities, to be set up across the entire network. From a core
configuration and policies. This is essential for speed network perspective, it may require core NFs to be Figure 2 illustrates a campus deployment with a protecting resource utilization for critical services.
and cost-efficiency. deployed in strategically placed edge DCs, hybrid private 5G network that addresses several To ensure survivability, the critical service slice relies
Monetization and the timely delivery of services potentially together with RAN functions, use cases. A critical slice (A) supports robotics with on multiple instantiations of RAN and CN functions
are vital. These, together with a need for cost- throughout the network to guarantee the strong requirements on both latency and to increase redundancy. The network slice for critical
efficiency, drive demand for automation and characteristics and performance of critical functions. survivability. A latency-sensitive slice (B) supports services will have a complete on-premises core
flexibility across the OSS/BSS layer. Exposure In these scenarios, a virtual private 5G network is augmented reality/virtual reality application, installation allowing services to survive.
enablers are required to allow customers to influence the ideal choice. possibly with lower requirement on survivability. A The latency-sensitive slice (B) uses QoS profiles,
and monitor the service. In limited-area scenarios, services are consumed non-critical slice (C) supports people working at the enabling each service to be served even during high
within a geographically limited area, such as a sports site (using handheld devices, for example), including load. This will avoid long latencies and service
Network slicing categories arena or massive event. All base stations within this mobility across the enterprise sites. interruptions, possibly at the expense of throughput
The main differences between network slices have to area need to be configured to support the slice, and The different types of use cases are supported by per service. Most of the functions handling RAN UP
do with the geographical area within which their there needs to be a DC close by for critical functions one slice each, and each slice is assigned a unique traffic are collapsed into one node, while functions
services can be reached and their specific service for the core network and potentially for the RAN. A S-NSSAI and DNN. These slices can be assigned governing control plane traffic are more centralized.
virtual private 5G network is a good fit in these dedicated RAN capabilities/NFs to satisfy their The core UPF will be deployed locally to ensure that
scenarios. unique requirements. For example, services within any delays that occur are minimal.
NETWORK SLICES CAN each slice could use different QoS profiles. The non-critical slice (C) can have the complete
Campus-based scenarios
BE LOOSELY GROUPED INTO Campus-based services require tightly integrated
Within the critical service slice (A), hard priority is
given (through QoS assignment) to services essential
core installed in the central DC. The RRP covers
several different services, whose allocation of
THREE SLICING CATEGORIES: solutions such as indoor coverage and local edge to the smart plant operation. The RAN will then resources will be regulated by their QoS profile.
CAMPUS-BASED, WIDE-AREA DCs to support low latency, data isolation and
service assurance during normal operation and fault
prioritize such essential services in a resource
shortage situation. Similarly, RRP in the critical
Such partitions are provided with “soft” borders,
meaning that if resources are not used by the slice
AND LIMITED-AREA situations. service slice case outlines strict policies for services, they are reused by other services.

68
6 E RTI ECCSH
ERICSSON SO N LTOEG
NO CYHN
ROE VL IOEG
WY ✱R EFVEI BE RWU A
✱R Y# 1011, 2 0 2 1 # 0 2A1R Y✱1 1E, 2R0I C
F E0B1 R2 U 2 1S S✱
ONE RTIECCSHS N
OON LTOEG
CYH N
ROE VL IOEG
WY R E V I E W 69
7
✱ APPLIED NETWORK SLICING APPLIED NETWORK SLICING ✱

networks. These offerings can be realized as one or NETWORK SLICING

Aggregated
several network slices targeting different use cases.
SCENARIOS VARY
Radio-access sites Edge DC WAN transport Central DC
transport
The four use cases shown in Figure 3 – stadium,
RAN
capabilities/
NFs
A. Non-critical slice for MBB Transport
partition 5GC/EPC
NFs
smart city, logistics company and consumer MBB – CONSIDERABLY DUE TO
each have different requirements in terms of
Public service (consumer MBB)
geographical reach and mobility. Network slicing THE DIVERSITY OF USE
RAN B. Mobility-enabled Transport RAN Transport
enablers provide each use case with its required CASES AND CUSTOMER
capabilities/ low-latency slice NF 5GC
functionality and characteristics supporting
REQUIREMENTS
partition partition 5GC/EPC
NFs for logistics NFs Appl.
NFs
different levels of SLAs.
Virtual private 5G network for a logistics company A non-critical slice (A) is used for traditional MBB
services, providing wide-area service with strong
C. High-availability slice
RAN
capabilities/ for stationary devices
RAN
NF 5GC
Transport
partition 5GC/EPC
requirements on mobility, interworking with earlier start with a few slices with wide-area coverage and a
NFs for a smart city NFs NFs
standards, roaming and so on. A mobility-enabled high degree of sharing, and then evolve to more
Virtual private 5G network for a smart city (limited area)
low-latency slice (B) serves the wide-area use case dedicated, isolated and limited-area ones. It is
represented by the logistics company in the form of a advisable to start with non-critical use cases, and
RAN D. Throughput-optimized RAN Transport virtual private 5G network. A high-availability slice then move on to latency-sensitive ones later.
capabilities/ slice for a stadium NF 5GC 5G
NFs NFs
partition
(C) is used for stationary devices, serving use cases in
a limited area for the smart city use case, as well as in Conclusion
Virtual private 5G network for a stadium (limited area)
a virtual private 5G network. Finally, a throughput- Network slicing scenarios vary considerably due to
optimized slice (D) serves use cases in a hotspot for the diversity of use cases and customer
virtual private 5G networks for the sports stadium. requirements. Engineering appropriate slices for
Figure 3 Slicing in wide-area/limited-area deployment (supporting virtual private 5G networks)
Different use cases and customers may need to be each case requires a solid understanding of an
assigned different network slices across the network. evolving set of enablers in cloud infrastructure, the
As there can be at least as many slices as there are enterprise’s core is installed, either centralized or The non-critical slice for MBB (A) uses RRP to RAN, the core and transport networks and in OSS/
customers, the OSS quickly need to scale to manage on-premises. These slices are configured to allow guarantee bandwidth. Slice-specific QoS regulates BSS. In light of this, we at Ericsson believe that the
a large number of instances. Critical and latency- mobility in and out of the campus. To enable this resource access among different services. The core starting point for pursuing network slicing should be
sensitive traffic would typically stay within the configuration, slice-based mobility features need to network needs to serve full mobility across the the business needs and use cases rather than the
campus, which implies little need to depend on be activated, where appropriate target cell selection coverage but does not require additional technology behind them. To fully harness the power
transport slicing for these use cases. ensures slice service continuity. The transport distribution. of network slicing, operators will need to embrace a
Figure 2 also shows other customers being network is critical to ensure the right E2E The mobility-enabled low-latency slice (B) uses new and transformational approach to building and
supported at the site. A slice supporting regular characteristics for these services. RRP available throughout the wide area. It has a operating networks.
mobile broadband (MBB) service reachable from It is preferable to introduce the various enablers in dedicated core network with distributed UP and
within the campus (D) allows workers as well as phases, in response to business needs. One high-availability configuration. Mobility is enabled
visitors to access the service with full mobility in/out approach could be to start with few campuses and with robust policies. The high-availability slice for
of the campus. Another slice (E) supports virtual then scale up gradually. It is prudent to start with just stationary devices (C) uses RRP in a limited area
private 5G network services for the same enterprise a few, static slices and then increase the number of with policies for fairness. The core network is
or a different one (such as a logistics company), slices and the level of “dynamicity” in terms of deployed locally without mobility considerations.
accessible from within the campus and with mobility orchestration and adaptability later on. It is usually The throughput-optimized slice (D) also uses RRP in
in/out of campus. wise to focus on less critical or advanced services dedicated sites. The core network is optimized for
A specific set of S-NSSAIs and DNNs are used for first, before moving to more critical ones. high throughput.
the MBB slice (D). The core and potentially some The main challenge for OSS relates to topology.
RAN functions are deployed in a central DC to Wide-area and limited-area scenarios This is due to the fact that even if there are fewer
provide ubiquitous access. For the wide-area virtual The example in Figure 3 shows a combination of network slices in comparison to the campus case,
private network (E), a specific set of S-NSSAI and wide area network and limited-area use cases. In this there may be a large number of sites and areas.
unique DNNs are used to direct to where the context, network slicing enables virtual private 5G From a phasing perspective, it may make sense to

70
8 E RTI ECCSH
ERICSSON SO N LTOEG
NO CYHN
ROE VL IOEG
WY ✱R EFVEI BE RWU A
✱R Y# 1011, 2 0 2 1 # 0 2A1R Y✱1 1E, 2R0I C
F E0B1 R2 U 2 1S S✱
ONE RTIECCSHS N
OON LTOEG
CYH N
ROE VL IOEG
WY R E V I E W 71
9
✱ APPLIED NETWORK SLICING APPLIED NETWORK SLICING ✱

area of packet core

the authors

architecture and technology,

with a focus on automated
orchestration of 5G
networks and how to use the
possibilities of network
slicing. Lemark holds an
M.Sc. in electrical
engineering from Chalmers
Angelo Centonza network slicing can act as a University of Technology in
◆ joined Ericsson in 2011 key enabler. He has more Gothenburg, Sweden.
after working for several tier- than 25 years of experience
1 telecom vendors in the in the telecom industry
areas of IEEE/3GPP across a wide range of
standardization, technology areas and
telecommunication and positions, including packet
defense systems. He is a core networks, cloud
principal researcher, focused technologies and OSS.
on the areas of RAN Basilier holds a M.Sc. in
automation, network slicing, computer science and
network architecture and technology from Linköping Thomas Åsberg
interface design. He also University, Sweden. ◆ joined Ericsson in 1987
serves as a 3GPP and currently serves as an
standardization delegate. expert implementation
Centonza holds an M.Sc. in architect, operation and
electrical engineering from manitenance (O&M), within Further reading
the Politecnico di Bari, Italy, the Technology Management
❭ Ericsson, 5G for business, available at: https://www.ericsson.com/5g/business
and a Ph.D. in hybrid department at the Ericsson
broadcast/ CTO office. He has held a ❭ Ericsson, Network slicing, available at: https://www.ericsson.com/network-slicing
telecommunication variety of roles – developer, ❭ Ericsson, 5G RAN slicing, available at: www.ericsson.com/ran-slicing
networks from Brunel lead architect, team leader, ❭ Ericsson, What is 5G?, available at: https://www.ericsson.com/en/5g/what-is-5g
University London in the UK. Jan Lemark project manager and line
❭ Ericsson Technology Review, Critical IoT connectivity: Ideal for time-critical communications, June 2, 2020,
◆ joined Ericsson in 1994 manager – in the areas of
Fredrik Alriksson, Lisa Boström, Joachim Sachs, Y.-P. Eric Wang, Ali Zaidi, available at:
Henrik Basilier and has worked in hardware and software R&D
https://www.ericsson.com/en/reports-and-papers/ericsson-technology-review/articles/critical-iot-connectivity
◆ joined Ericsson in 1991. He technology areas including with systems and architecure
is an expert in network packet core, user data evolution, with a focus on the ❭ Ericsson white paper, Critical capabilities for private 5G networks, available at:
https://www.ericsson.com/en/reports-and-papers/white-papers/private-5g-networks
architecture evolution, management, IMS and O&M area and value for the
focusing on 5G networks platforms. He currently user. ❭ GSMA, An Introduction to Network Slicing, available at: https://www.gsma.com/futurenetworks/wp-content/
and applications and how serves as a developer in the uploads/2017/11/GSMA-An-Introduction-to-Network-Slicing.pdf

10
72 ERICSSOE
NRTI E
CCSH
SONO
N LTOE G
CYH NRO
EV I EGWY ✱R EFVEI BE R
LO WU ✱
A R Y# 0
1 11, 2 0 2 1 0 2A1R Y✱1 1E, R
F E0B1 R2 U
# 2 1S S✱
2 0I C ONE RTIECCSHSN
OONLTOEGCYH N
R EOVLIOE G
WY R E V I E W 11
73
ISSN 0014-0171
284 23-3353 | Uen

© Ericsson AB 2021
Ericsson
SE-164 83 Stockholm, Sweden
Phone: +46 10 719 0000