Available via license: CC BY 4.0
Content may be subject to copyright.
1
=Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number
An AI-Assisted Framework for Lifecycle
Management of Beyond 5G Services
Alexandros-Ioannis Manolopoulos1, Viktoria-Maria Alevizaki1, Markos Anastasopoulos1,
Anna Tzanakaki1
1Department of Physics, National and Kapodistrian University of Athens, Athens, GR
Corresponding author: Alexandros-Ioannis Manolopoulos (e-mail: almanolop@phys.uoa.gr).
“This work has been financially supported by the EU projects 5G-TACTIC under grant agreement 101127973, 6G-SENSES under grant agreement 101139282
and ECO-eNET under grant agreement 101139133”
ABSTRACT Future mobile communication networks aim to offer services and applications in the most
flexible, adaptable and cost-effective manner. B5G networks aim at a fully softwarized network architecture,
where hardware and software programming is used for the design, implementation, deployment,
management, monitoring and maintenance of network equipment/components/services. Artificial
Intelligence (AI) and Machine Learning (ML) techniques are steadily being integrated into 5G systems,
offering intelligent automation, proactive network management, and resource allocation optimization. In this
environment, the role of Management and Orchestration (MANO) is vital to ensure efficient infrastructure
utilization and fulfillment of heterogeneous service requirements. Despite the development of various tools
and platforms to facilitate MANO in 5G systems, in most cases there is still the need of human intervention
and manual input for configuring the 5G elements according to service requirements. In this paper, a MANO
framework has been developed, that specifically targets the orchestration operations of 5G networks. The
proposed framework focuses on the lifecycle management of the 5G components, in order to achieve an
operational environment with minimal human intervention or manual configuration (Zero Touch Networks -
ZTN). Within this ecosystem, an Analytics & AI/ML Platform has comprehensive monitoring capabilities
and influences decisions across various layers or aspects of the infrastructure. This includes optimizing the
allocation and orchestration of both networking and edge/cloud computing virtual resources within the
infrastructure.
INDEX TERMS 5G, B5G,6G, MANO, slicing, NFV, LSTM, LCM, ZTN
I. INTRODUCTION
The introduction of 5G technology that promises faster data
rates, ultra-low latency, massive machine-type
communications and increased network reliability has
transformed mobile networks over the past few years. A key
enabler behind these advancements is the exploitation of cloud
computing to support the operation of 5G networks,
introducing the notion of 5G Cloud networks [1]. 5G Cloud
networks are able to offer innovative services and applications
with improved performance. This is achieved by taking
advantage of advanced networking technologies with
increased scalability and flexibility features that cloud
infrastructures inherently offer [2]. In addition, 5G
technologies introduce the feature of network slicing that
allows partitioning of the physical network infrastructure into
multiple virtual/logical network slices [3]. Each slice can
operate independently and can be configured to meet specific
service requirements. These requirements can be mapped to
various Quality of Service (QoS) classes corresponding to
different levels of bandwidth, latency, traffic priority, security,
reliability etc.
To achieve this, 5G networks adopt novel architectural
concepts such as microservices, network function
decomposition, as well as Control and User Plane Separation
(CUPS). The adoption of these concepts and approaches
enables the 5G infrastructure to become flexible and adaptable
increasing the efficiency with which resources are being
utilized. In the Radio Access Network (RAN) domain,
network functions are separated based on their roles (control
or user plane) and can be placed at different locations
according to their resource requirements and delay constraints.
The Core Network (CN) also adopts CUPS, leveraging
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
2
virtualization and softwarization. The overall CN architecture
follows the paradigm of the Service-Based Architecture
(SBA) involving a set of key Virtual Network Functions
(VNFs). [4]
While 5G delivers impressive advancements in data rates,
latency and connectivity, it still faces the challenge of
explosive growth in data-intensive and latency-sensitive
applications. To address this, Beyond 5G (B5G) is emerging,
that brings advancements beyond the current 5G standards,
setting the path towards 6G [5]. B5G is build on 5G
foundations and introduces key improvements, particularly in
integrating Artificial Intelligence (AI) and Machine Learning
(ML) for intelligent network management and predictive
maintenance, aiming to support a set of diverse services.
These services are related to various use cases, sectors and
vertical industries [6] spanning from Unmanned autonomous
Vehicles and automated production lines to entertainment,
Extended Reality (XR) and the Internet of Things (IoT).
Therefore, these networks are expected to affect and in some
cases even reshape various aspects of every day’s life as well
as the means of interaction between humans and technology
[7].
The heterogeneity and dynamicity of these complex
environments poses new challenges in terms of management
and performance optimization for these advanced systems. In
this context, AI and ML can play a key role [8]. More
specifically AI/ML tools can enable intelligent automation,
proactive network management and optimization in resource
allocation. By analyzing the massive volumes of data
generated by 5G networks, AI/ML algorithms can identify
patterns, predict network behavior, detect anomalies and
security threats as well as make informed real-time decisions,
leading to further optimization of network performance, QoS
enhancement, delivery of undisrupted services etc.
Furthermore, network operators can exploit AI/ML tools to
efficiently facilitate network planning and management of 5G
infrastructures in a cost-effective and efficient manner[9].
It is therefore clear that B5G networks are transforming into
open, flexible and efficiently shared infrastructures. In this
environment, traditional Network Functions (NFs),
softwarized according to the Network Function Virtualization
(NFV) [10] paradigm, are managed through a centralized
Management and Orchestration (MANO) Platform [11].
MANO introduces robust and centralized management of
service creation, offering network-wide service design,
configuration, deployment and monitoring. Additionally, it
enables automation in the arrangement and coordination of
network elements as well as scaling of resources and services.
This network-wide orchestration introduces the benefit of a
single integration point and a centralized representation of
distributed network resources, regardless of the volume of
resources involved or the location of these resources. By
automating, through the orchestrator, the infrastructure
configuration and monitoring processes, it is possible to
reduce the inherent complexity of delivering and
administering sophisticated and multi-featured services.
Solutions such as Open Source MANO (OSM)[12] and Open
Network Automation Platform (ONAP) [13] have emerged to
handle the lifecycle of the NFs, according to the standards set
by the European Telecommunications Standards Institute
(ETSI) and the Open Network Foundation (ONF) for NFV.
Both platforms aim to provide a versatile approach enabling
the onboarding of any application. In their typical design
approach, each vendor's application is paired with its own
Operator- a software component that encloses the application
along with comprehensive instructions for tasks such as
deployment, configuration, scaling, and integration on the
cloud. However, this often leads to discrepancies and
inaccuracies in the terminology and taxonomy associated with
cloud-native principles and MANO. Furthermore, the designs
and implementations of 5G elements from different vendors
often exhibit static behavior, even for basic tasks such as IP
address assignment and resolution. This reliance on static
setups necessitates human intervention that introduces
scalability challenges. Such malpractices are often justified as
minor engineering issues, but they are in fact a reflection of
carrying over a common practice from either legacy Physical
NFs (PNFs) or outdated design architectures that do not take
into consideration the nature of cloud and cloud native
deployments. More specifically, in case every vendor creates
a unique Operator, that is specific to the vendor applications,
inconsistencies may arise, introducing increased complexity.
To address this challenge, Operators are assigned to higher-
level, standardized network components—such as network
terminals, functions, or slices. This enables Operators to
manage logical network elements in a standardized way, rather
than being tied to specific applications from different vendors.
This approach allows to mix and match NFs, addressing
previous challenges arising from individual vendors being
locked with their own MANO and Operation Administration
and Maintenance (OAM) solution.
Going one step further, the vision for the B5G era is to further
extend the levels of automation and minimize human
intervention in network and service management. Zero Touch
Network (ZTN) and Service Management (ZSM) is
introduced with the aim to enable a network that is self-
configured, self-monitored, self-healing and self-optimized
[14]. To achieve this, ZSM strongly relies on tools and
attributes supported by AI/ML schemes and MANO.
Contributions: Aligned with the B5G vision for zero touch
system operation, this work leverages some of the key
technologies and concepts of 5G networks to develop a
framework that automates provisioning, deployment,
management and orchestration of network slices and services.
This framework can be used to easily deploy, manage, modify,
and delete 5G services and functions, while also autonomously
performing re-configuration actions, without human
intervention. The paper also presents the complete Lifecyle
Management (LCM) of CN components that allows 5G
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
3
systems to operate in an intelligent, adaptable, and flexible
multi-slice environment.
To achieve this functionality, a set of building blocks have
been implemented and are able to:
- host and instantiate various 5G deployment options
involving multiple network slices and sophisticated
topologies.
- handle LCM operations and manage network slices,
leveraging domain NFV Orchestrators (NFVOs) and
controllers.
- monitor the entire 5G system collecting a variety of
performance measurements for the virtualized network
functions and the underlying physical entities.
- take intelligent and fully optimized decisions leveraging
AI/ML algorithms. These algorithms are trained based on
data collected from the monitoring system recommending
optimal LCM actions.
To test the efficiency of the deployed system, a two-stage
evaluation process has been adopted. The first stage tests the
ability of the platform to perform a set of orchestrator-based
multi-slice network deployments. These tests evaluate the
efficiency of the platform to instantiate appropriate 5G
topologies (deployment options) across a diverse set of
operational scenarios. Once the 5G network has been
deployed, the second stage tests are used to evaluate the ability
of the system to perform optimal reconfigurations with zero
human interventions optimizing User Plane (UP) traffic
forwarding policies.
The rest of the paper is structured as follows: Section II
provides a brief overview and a literature review of the basic
components in 5G. This includes 5G architectural aspects,
mechanisms for network management with emphasis on
slicing and, finally, AI-based tools and algorithms for optimal
decision making. The progress and main innovations of the
proposed work compared to the state of the art is also
highlighted. Details of the proposed Overall System Design is
presented in Section III, while demonstration and
experimentation results are provided in Section IV. Finally,
Section V summarizes the conclusions of the paper and
proposes directions for future work.
II. BACKGROUND AND RELATED WORK
A. 5G System Architecture
The 5G ecosystem brings together a set of heterogeneous
applications and use cases such as enhanced Broadband
(eMBB), massive machine type Communications (MMTC)
and ultra-Reliable Low Latency Communications (URLLC)
[15][16], [17]. These use cases and the related applications
have extremely diverse and stringent requirements in terms of
network and compute resources, QoS aspects etc.
Consequently, 5G Systems introduce several architectural
advancements compared to previous generations of mobile
communication systems (3G/4G) and involve the adoption of
new concepts and technologies in order to meet the
requirements of the applications they are expected to support.
The overall 5G network architecture is shown in Figure 1. 5G
networks adopt the microservices architecture [18] and
network function decomposition [19], to provide a network
that allows for each service to scale or update without
disrupting other services in the network. These concepts have
been applied in the design of RAN and CN segments.
In the RAN, the concepts of vertical and horizontal functional
splits [20] is adopted in the design of both control and data
planes. The CUPS in 5G is known as “vertical split” while
“horizontal split” refers to the decomposition of the baseband
function stack into a set of individual independent functions
that can be allocated to different computational resources. The
most flexible RAN solution supports splitting of the baseband
function stack into a set of functions that can be independently
allocated to the Remote Unit (RU), the Distributed Unit (DU)
and the Central Unit (CU) according to the service
requirements, aiming to maximize resource and energy
efficiency as well as service performance. This architecture
enables a more flexible mapping of NFs to physical network
entities, depending on the use case and deployment constraints
[21].
In the 5G core segment, the concept of CUPS is adopted by
separating the control and user plane NFs [4]. Furthermore, it
relies on virtualization and softwarization which decouples the
various control functionalities from the underlying
hardware/infrastructure. This way, the 5G Core can benefit
from the advantages of Cloudified and cloud-native
deployments. The Control Plane (CP) of core network
comprises multiple VNFs that interact through service-based
interfaces (SBI) [22] and is accountable for decision-making
and network management. These VNFs include the Access
and Mobility Management Function (AMF) to facilitate user
registration and the Session Management Function (SMF) that
handles user connections. The UP of the CN includes the User
Plane Function (UPF), which manages the data path and traffic
policies. More specifically the UPF performs packet
inspection and routing, as well as UP QoS handling. The
communication of the UP elements with the CP elements is
achieved through point-to–point interfaces [23] where each
gNB
5G-RAN
CP
UP
CP
UP
CU DUs
CUPS Split
(vertical split)
CU/DU Split
(horizontal split)
Nnssf
Nnrf
Nausf
Namf
Nnwdaf Nudm
Nnef
Naf
NudrNsmf
SMFAMF UDRNEFNRF
AUSFNSSF UDMNWDAF AF
UE
Service Based
Architecture
N1
N2
N3
N4
5G-core Control Plane
5G-core User Plane
UPF DN
N6
N9
Figure 1: 5G Network Architecture
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
4
interface serves specific purposes, such as carrying user plane
traffic (N3, N9, N6), managing mobility and session
establishment (N1, N2), or managing the UPF nodes (N4)
[24].
B. 5G MANO AND NETWORK SLICING
A core prerequisite for 5G systems is the support of flexible
and configurable network architectures, so that they can adapt
to any use case and service requirements. To achieve this, 5G
systems embrace technologies such as NFV (already briefly
discussed) and Software Defined Networking (SDN) to enable
dynamic deployment of network functionalities, replacing the
need for manual, node-by-node configuration. This
centralized approach is related to the migration of individual
device configuration in favor of a more robust management
mechanism that can offer network-wide service design,
configuration, deployment, and monitoring. Such processes
require implicit autonomic control over all systems, resources,
and services as well as inherent intelligence.
In this context, NFV enables deployment of network functions
on virtual machines (VMs) or containers hosted on general-
purpose servers, rather than relying on vendor-specific
hardware. This approach allows the system to adapt
dynamically to varying network requirements and optimize
resource allocation based on actual end-users current and
future demands. A high-level view of the NFV architecture is
shown in Figure 2 comprising several key elements [25][21],
[26], [27] including: a) VNFs, which are software-based
instances that perform specific network functions; b) the NFV
Infrastructure (NFVI), that provides the necessary virtualized
resources, such as computing, storage, and network; and c) the
NFV MANO framework, that oversees the lifecycle
management of VNFs and coordinates the efficient use of
resources.
The MANO framework integrates various managers, such as
the Virtualized Infrastructure Manager (VIM), the VNF
Manager, and the NFV Orchestrator, to ensure automated
deployment and orchestration of network services. MANO
has been adopted by several research studies and projects to
automate network configurations in 5G networks. For
example, in [28] the authors adopt MANO to trigger allocation
of a set of resources and service scaling policies to meet
specific Service Level Agreement (SLA) requirements. Three
different scaling types are addressed, namely, application-,
resource- and scaling level. Their solution was implemented
in a proof-of-concept virtualized platform using in the wireless
access part an IEEE802.11p network verifying the ability of
MANO to automate network deployment and update service
instances. Unlike [28] that relies on a monolithic IEEE802.11p
RAN solution, in our work MANO is used to automate
deployments in a highly complex disaggregated 3GPP Rel.15
compliant 5G network for the RAN and core segments.
Similar studies have been also performed by the authors in
[29], [30] where MANO is used for the management of 4G
and WiFi network testbeds. The MANO concept has been also
adopted by the authors in [31] to automate the network service
lifecycle management with the use of knowledge management
and decision support techniques Their studies have been
carried out over a simulated (ns-3 based) 5G network showing
the ability of MANO to provide flexibility in service
deployment and decommissioning.
Network slicing is also a crucial concept in 5G networks,
enabling the partitioning of a single physical network into
multiple, isolated virtual networks or "slices"[32]. Each slice
is customized to meet specific service requirements, such as
enhanced connectivity, reliability, performance, and
scalability that allow the support of differentiated services
over the same infrastructure. Network slicing supports
multitenancy, where each tenant can operate independently
over customly configured virtualized network instances that
are either fully isolated ("hard") or share certain resources
("soft"). The specific characteristics of a network slice are
defined by attributes such as the Service Slice Type (SST) and
the Service Descriptor (SD). The SST is a high-level identifier
that represents a specific type of service that the network slice
is intended to provide (eMBB, URLLC or mMTC). The SD,
on the other hand, is a detailed specification that defines the
functional and non-functional requirements of the service,
including network functions, network topology, performance
metrics, and security policies.
To implement network slicing effectively, the Next
Generation Mobile Network (NGNM) structures the 5G slice
architecture into three distinct layers [32]. The 5G Service
Instance Layer (SIL) provides the service instances that need
support, while the 5G Network Slice Instance Layer (NSI)
defines the network requirements and configurations for these
services, potentially incorporating shared or dedicated
network slice sub-instances (NSSIs). Finally, the 5G Resource
Layer (RL) manages the allocation of both physical and
logical resources to each slice.
The 3rd Generation Partnership Project (3GPP) has
established a comprehensive framework for managing
Virtual Network Functions (VNFs)
NFV Infrastructure (NFVI)
Management and
Orchestration (MANO)
Virtualiz ation Lay er
Physical
Memory
Physical
Compute
Physical
Network
Virtual
Memory
Virtual
Compute
Virtual
Network
VNF VNF VNF VNF
Virtual I nfrastructure
Manager (VIM)
VNF Manag er
NFV Orchest rator
Figure 2: NFV Architectural Framework.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
5
network slices, encompassing a three-step lifecycle:
instantiation, configuration, and activation; run-time
operations; and decommissioning. During the first phase, all
necessary resources for the network slice are provisioned and
configured to become operational. In the run-time phase, the
slice is actively managed, monitored, and adjusted as needed
to maintain performance. Finally, in the decommissioning
phase, the slice is deactivated, and its resources are released,
ensuring efficient resource utilization across the network [33].
MANO and Network Slicing integration has been considered
by several studies to optimize performance and service
delivery. The authors in [34] present a highly automated
management framework for E2E network slices, designed for
multi-tenant 5G networks. The proposed framework is
designed to define slices from business models for network
slice providers. Their solution is prototyped and
experimentally validated in a large-scale 5G Non-Stand-Alone
(NSA) infrastructure. In [35] the design of a 5G network with
configured slices that offer low-latency services is being
presented. The implementation includes a cloud computing
platform, MANO and a 5G NSA platform, based on open
source tools. The implementation includes a cloud computing
platform, MANO and a 5G NSA platform, based on open
source tools. Our study differs from [34] and [35] in that we
use MANO for LCM of a Stand-Alone (SA) 5G Core
implementation combined with predictive analytics. Finally,
in [36] MANO is used to assist network slicing operations in
the RAN domain without involving network reconfiguration
and automation aspects.
C. ARTIFICIAL INTELLIGENCE/MACHINE
LEARNING
Over the past few years AI/ML techniques are increasingly
being adopted by the telecommunications industry, in an effort
to advance network optimization, security aspects, QoS etc.
Specifically, in the context of B5G/6G networks that are
expected to generate and handle enormous amounts of data,
the adoption of AI/ML techniques aims to transform this data
into knowledge. This knowledge can prove to be extremely
beneficial in automation as well as service lifecycle
management and reshape the relevant business models and
opportunities. For example, time series predictive models such
as network traffic forecasting is a field where the adoption of
AI/ML techniques can provide significant improvements.
Neural networks (NNs) have become a powerful tool in time
series data prediction, due to their ability to model complex,
non-linear relationships within data. By capturing temporal
dependencies and patterns, neural networks, such as Recurrent
Neural Networks (RNNs) and their variants like Long Short-
Term Memory (LSTM) networks [37], succeed in accurately
predicting future values based on past observations. Several
types of NNs and algorithms are well-suited for time series
data forecasting apart from RNNs, such as Convolutional
Neural Networks (CNNs) [38], attention mechanisms [39] and
Hybrid models [40]. These networks can learn intricate
sequences and trends over time, that makes them highly
effective for forecasting applications. By continuously
learning from new data, neural networks increase forecasting
accuracy and provide robust and adaptive models that
outperform traditional statistical methods in many scenarios.
AI/ML functionalities have also been introduced in different
domains of 5G Systems with their relevant interfaces and
functions [41]:
In MANO, AI/ML can enable optimization of network
resource allocation, network performance, efficient
analysis of failures and design of e2e network slices.
Additionally, it can support root cause analysis and alarm
correlation. An orchestration framework for the lifecycle
management and orchestration based on closed loop
optimization is presented in [42]. Specifically, apart from
the Day 0 operations that include VNF onboarding, the
authors introduce a zero-touch slice deployment with
intelligent decisions on optimal placement for Day 1
operations. Day 2 operations include analytics
functionalities and an optimization engine. A wide set of
use cases such as optimal resource allocation, dynamic
VNF placement and performance optimization are
supported by the proposed framework. Similarly, in [43]
a framework for the management of networks with
massive network slices is proposed. This MANO
framework achieves automation through the use of
multiple, distributed, AI-driven control loops that can
work at different levels (node, slice, interslice and
orchestration domain level). Finally, a MANO
framework for B5G vehicular edge service which is based
on closed-loop orchestration, is presented in [44].
The CN focuses on AI/ML services in the control plane,
targeting specific sessions, flows, or User Equipment
(UE). These services aim to analyze or predict users'
communication behavior, assess security risks, and
ensure desired network performance. For example, the
authors in [45] use analytics and ML techniques for three
different use cases. Firstly, an Extreme Gradiant Boosting
(XGBoost) model is implemented to obtain anomalous
behaviour in UPF nodes. Secondly, they consider RNN,
LSTM and Linear Regression (LR) models to predict load
traffic in 5G cells and thirdly, a closed-loop automation
model is implemented to predict SMF resource usage and
automatically instantiate SMF instances. The authors in
[46] propose the adoption of AI techniques in order to
optimize placement and scaling aspects in the 5G CN.
They explore how AI-based scaling algorithms combined
with functionality-aware placement can enable the design
of network slices. A mobility prediction ML-assisted
scheme which reduces the signaling-induced overhead, is
proposed in [47].
Within RAN, AI/ML utilizes real-time or near-real-time
data to predict and analyze user access and dynamic radio
conditions. The aim is to optimize tasks such as
scheduling, interference control, and radio resource
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
6
allocation. To this end, the authors in [47] propose an
intelligent Radio Resource Management (RRM) scheme
that aims to handle traffic congestion. The viability of
such solution is tested on a real-world dataset. Finally, a
network intelligence orchestration framework is
presented in [48], which is designed within the concepts
of Open-Radio Access Network (O-RAN) and
automatically computes the optimal set of data-driven
algorithms and their execution location.
Summarizing the above, although it is widely accepted that
MANO and AI can provide significant benefits in automating
and optimizing B5G systems operations, the majority of the
existing studies treat these concepts in a segmented way. To
the best of our knowledge this is the first study that addresses
jointly the following topics:
- Experimentally validates MANO over an operational 5G
testbed considering 3GPP Rel 15 5G components.
- Uses MANO to dynamically instantiate a wide range of
5G network deployment options, responding to a variety
of operational scenarios, considering the concepts of
functional split both at the RAN and at the Core segments.
- Integrates MANO with AI algorithms to perform real
time decision making, taking actions to rescale, modify or
even delete individual 5G system building blocks.
- Perform extensive experimentation quantifying resource
requirements and (re)-configuration times for the
deployed platform.
III. PROPOSED PLATFORM DESIGN
The main goal of this work is to design a framework that can
assist network operators to easily deploy, manage, modify,
reconfigure and delete 5G services. In order to do this, a
platform/environment that takes advantage of all technologies
and concepts entailed in the 5G ecosystem has been designed.
This section describes this environment along with the tools
and concepts that have been used to support its functionalities.
The detailed structure of the proposed 5G management
platform is illustrated in Figure 3 comprising the following
building blocks: (1) the 5G platform used to host the main 5G
network functions, (2) the Data Collection and Monitoring
block collecting statistics for the physical and virtual elements
of the system, (3) the Predictive analytics (AI/ML) block used
to assist decision making and (4) the Automated-LCM block.
These building blocks are integrated to serve the purposes of
5G MANO
5G Slice LCM
Predictive Analytics
2
4
Routing/Scaling/
Decision Making
3
5G NF Descriptors
5G CP 5G UP
SSH
Network Slice Configuration
NST 1 ... NST n
Instantiation/
Decomissioning
1
Runtime
Operations
VM/container
UPF
PCF
NRF
AUSF
UDR
SMF
AMF
UDM
NSSF
NEF
UPF
PCF
NRF
AUSF
UDR
SMF
AMF
UDM
NSSF
NEF
Control Plane User Plane
i-UPF
PCF
NRF
AUSF
UDR
SMF
AMF
UDM
NSSF
NEF
Control Plane User Plane
VM/container
VM/container
b-UPF PSA-UPF
Flexible 5G 5G CUPS Monolithic 5G
5G Cloud Platform
Figure 3: Proposed Framework with 4 components: 1) 5G Cloud Platform supporting flexible deployment options, 2) Monitoring Platform,
3) Predictive Analytics/Decision Making 4) Automated LCM.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
7
this work as stated above. In the following subsections each
block is presented in detail.
A. 5G CLOUD PLATFORM
5G Systems need to operate in highly heterogeneous
environments with stringent, varying and sometimes
conflicting service requirements. For this reason, traditional
all-in-one deployments adopted from the previous generations
of mobile networks have been replaced with more flexible and
adjustable deployment options. Following this approach, the
present framework supports a variety of 5G network
deployment options that can be flexibly adapted to meet the
requirements of the offered services. A graphical illustration
of the main deployment options that are supported by the
platform is shown in Figure 3. These include:
- The Monolithic deployment where all the functions are
collocated in the same site. This option provides a fast,
easy-to-deploy solution suitable for private 5G networks
since there are no shared elements and all NFs are fully
isolated in one environment, providing enhanced security.
- The CUPS deployment where the Control Plane Functions
are separated from the User Plane. Through this
deployment option the User Plane Function can be hosted
in physical/virtual machines that are located close to the
end-users, providing better support for delay-sensitive use
cases. In this option, the CP and UP elements of the 5G
platform are hosted in different machines1.
- The Flexible deployment option, where the concept of the
CP/UP split is expanded to the CP elements, as well as the
addition of extra UP nodes, Local Breakouts etc. Local
Breakouts [49] refer to the capability of routing traffic
flows directly from the edge of the network to its
destination without passing through the core network,
reducing latency and improving performance. This
deployment option can support implementation of
isolated network slices, applications with different QoS
requirements, node up/down-scaling for environments
supporting high mobility scenarios etc. This flexible
approach supports diverse applications by isolating these
through network slicing and enables services to adapt to
dynamic traffic requirements, including high-mobility
scenarios, through efficient scaling mechanisms.
B. DATA COLLECTION AND MONITORING
In order to monitor the complex technologies and
infrastructures deployed in the present study, the proposed
framework includes a set of Service Assurance (SA) functions
that comply with the 3GPP TS 23.501 standard [24]. By
integrating monitoring tools in the framework, we can extract
valuable data that map resource consumption to certain
network functionalities. This enables MANO to make more
informed network decisions, leading to improved
1 The term "machine" refers to either VMs or containers, depending on
the underlying virtualization layer
performance, increased security, and reliability. For example,
using data collected from network monitoring, malfunctions
can be identified that combined with alerting mechanisms can
trigger reconfiguration actions when needed.
In the current implementation, the monitoring platform
consists of network agents that are placed in every compute
and network component. These agents collect performance
metric values from the hosts and expose these to certain
interfaces. These values are then retrieved from the monitoring
server and are stored in a database for further analysis,
visualization, SLA and trending reporting. [50]. The
monitoring platform is based on open-source tools, such as
Prometheus and Grafana. Additionally, the devices are
connected to energy metering sensors that collect energy-
related data and store these to the Prometheus database. This
is a valuable component, as for all system operations, such as
service provisioning, network deployment and reconfiguration
etc., resource consumption can be measured in compute,
network and energy levels.
C. PREDICTIVE ANALYTICS
The predictive analytics block of Figure 3, interoperates with
the Data Collection and Monitoring component of the
platform to gather, process and analyze the retrieved data and
make predictions regarding future data traffic load in the
system. Data traffic forecasting is crucial for 5G networks, as
it can lead to resource allocation efficiency, optimized
network design and increased QoS management. By
accurately predicting data traffic patterns, we can dynamically
allocate resources depending on the demands and avoid
under/over-provisioning of resources in support of the
respective functionalities.
In this direction, we have developed a forecasting model that
is based on LSTM neural networks to realize this building
block. The model was developed in Python using the Keras
library. The model is implemented as follows:
a) Data Retrieval: The data are retrieved from the
Database through a simple Query
b) Data Preprocessing: The data are first normalized
and then split into a training set (67%) and a testing
set (33%)
c) Model Training: The model is trained through the
Keras library
d) Prediction: After completion of the model training
the model is ready to make predictions
D. AUTOMATED LIFECYCLE MANAGEMENT
Τhis block provides tools with which network operators can
easily deploy, manage, reconfigure and rearrange the network.
In this context, an LCM framework was developed that is able
to dynamically manage network slices, leveraging domain
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
8
NFVOs and controllers. The development of the framework
was performed in three stages:
5G Deployment Option Selection: The 5G platform is able to
instantiate a variety of deployment options spanning from the
monolithic all-in-one to the fully disaggregated approach
where CP and UP elements are separated. To achieve this, the
automated LCM framework is able to apply 5G network
configuration policies that can be used to instantiate:
- CP and UP entities that are fully separated,
- multiple UPF elements with different roles. The LCM
framework is able define the number and type of UPF
elements in the User Plane path. Therefore, it is able to
instantiate:
o a Packet Data Unit (PDU)-Session-Anchor (PSA)
UPF that acts as a single termination point for the
PDU Session.
o an Intermediate (I) UPF: This UPF is located within
the path between the RAN and the PSA-UPF and is
responsible for forwarding data between the RAN and
the PSA-UPF .
o a Branching (B) UPF: The B-UPF redirects uplink
traffic to the appropriate UPF that ends the PDU
Session and merges the downlink traffic from
different PSA-UPFs to the UE.
- Network Slices with different QoS characteristics
- Application Servers (AS) processing users’ data. AS can
be hosted at various locations i.e., on traditional Central
Clouds (CCs), or on computing resources that are closer
to the network edge (Mobile Edge Computing -MEC),
enabling low-latency, high-bandwidth applications in 5G
networks.
5G NF Automation: The next step maps the above elements to
the necessary descriptors that feed the NFVOs. In the current
implementation, OSM has been used as the main NFVO, since
its Information Model (IM) is ETSI NFV-aligned and is
agnostic of the underlying infrastructure, so that its models can
be used across various VIM types and transport technologies.
OSM uses configuration templates, called descriptors, to
describe the key characteristics of managed objects (e.g VNFs
or Network Services (NSs)) in a network. For each
component, descriptors specify how it will be deployed and
used, as well as how it will interact with other components.
Descriptors are written in YAML, a markup language
designed for data that is easy to read and understand.
Therefore, to automate the deployment process in 5G
networks the required VNFs for the main 5G elements, i.e.,
VNFs for 5G CP and UP have been implemented. The
specifications of the VMs that comprise VNFs along with their
connections are exposed to OSM through the VNF Descriptor
(VNFD). Thus, two VNFDs, one for the 5G-CP and one for
the UPF were created.
In order to be able to flexibly mix and match the two 5G planes
creating multiple 5G topologies, we considered each 5G plane
as a distinct Network Service in the context of OSM and two
Network Service Descriptors (NSDs) are created, one for each
plane. An NSD is a higher-level abstraction that defines the
structure and behavior of a network service composed of
multiple VNFs. NSDs reference one or more VNFDs to
VNF: 5G CP
VDU:
Name: cp
Resources:
- 2 vcpu
- 4GB RAM
- 40GB disk
- PARAVIRT interfaces
ens3
ens8
ens9
N6
N2
N4
int-cpd
ext-cpd
int-cpd
int-cpd
ext-cpd
ext-cpd
VNF: 5G UPF
VDU:
Name: upf
Resources:
- 1 vcpu
- 2GB RAM
- 20GB disk
- PARAVIRT interfaces
ens3
ens8
ens9
ens10
N6
N3
N4
N9
int-cpd
ext-cpd
int-cpd
int-cpd
int-cpd
ext-cpd
ext-cpd
ext-cpd
VNF: 5G CP
N2
ext-cpd
N6
ext-cpd
N4
ext-cpd
N2
N4
N6
NS: 5G CP
sapd
sapd
sapd
VLD
VLD
VLD
VNF: 5G UPF
N3
ext-cpd
N6
ext-cpd
N4
ext-cpd
N9
ext-cpd
N3
N4
N9
N6
NS: 5G UPF
sapd
sapd
sapd
sapd
VLD
VLD
VLD
VLD
Figure 4: Graphical illustration of the VNFDs and NSDs for 5G CP and UP respectively. In the VNFD we create the appropriate network
interfaces in the VMs that host the 5G NFs, and in the NSDs we connect these interfaces to the appropriate networks inside the private cloud.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
9
specify the VNFs that compose the network service, their
arrangement and the connections between them. In this work,
5G NSDs describe the network connections of the 5G VNFs
internal to the private cloud. Figure 4 shows a graphical
illustration of the 5G VNFDs and NSDs that were developed.
The created descriptors represent a generic implementation of
the two 5G planes, but they can be parametrized at the
instantiation of the slice, in order to be tailored to support its
specific requirements.
Slice Deployment: In the next step, the requirements of the
service slices are given as input. These are then automatically
converted into configuration parameters for the generic 5G
descriptors. Subsequently, the framework prepares a network
slice template (NST), that describes the characteristics,
requirements, and behavior of the network slice. The NST
comprises two main blocks, the NSSIs and the slice virtual
links (VLDs). The NSSIs include the 5G CP and UP NSs and
can be shared among multiple slices. The VLDs include the
network connections that are required for the slice. The NST
is provided to the NFVO that is responsible for the
instantiation of the slice.
The LCM of the slice can be performed either by the NFVO,
or directly by the LCM framework. In the first scenario, we
use the VNF Configuration and Abstraction (VCA) layer of
OSM that uses juju [51] software to achieve the 3-step MANO
lifecycle of 5G slices. Juju uses a set of generic scripts and
metadata, called charms, that encapsulate DevOps expertise
and can be adapted for various software deployments. Charms
are given to OSM along with the VNFDs and configure VNFs
by executing automated scripts via Secure Shell (SSH). They
are written in any language executable from the command line
and consist of YAML configuration files and "hooks" which
manage software installation, service control, charm
configuration, and interactions between charms.
Charms perform actions that are classified into:
actions that are automatically performed at the
instantiation of the slices (Day 1 actions)), and
actions that can be dynamically performed during the
deployment of the slice (Day 2 actions).
The charms that were developed for the appropriate operation
of the 5G control and user planes, perform the following
actions:
- Day 1:
1. Configure ssh access and IPs of the VMs
2. Manipulation of the configuration files for the proper
operation of 5G core and UPF
3. Role of UPF (i-upf or psa with N3/N9 interface)
4. Load required modules (e.g. General Packet Radio
Service (GPRS) Tunnelling Protocol User-part
(GTPU) tunnel for UPF)
5. Start UPFs
- Day 2:
Start/Stop 5G CP
Start/Stop UPFs
Charms can be allocated within the VMs of the VNFs (native)
or, most commonly, are hosted in LXD containers within the
OSM machine (proxy). When deploying a proxy charm,
several time-consuming steps occur by default. The LXD
container hosting the charm must be launched and configured
and the charm must be installed. For time sensitive scenarios,
in order to avoid delays, the developed LCM framework can
bypass the need of juju charms, by directly performing the
necessary actions for the slices appropriate functionality
through SSH connections. The NFVO in this scenario is
responsible for the allocation of the appropriate resources to
the slice, and the rest of the configuration and run-time
operations are performed through the LCM framework.
IV. TESTBED IMPLEMENTATION AND EXPERIMENTAL
RESULTS
The framework along with all the relevant building blocks that
was presented in Section III, aims to provide a tool that Mobile
Network Operators (MNOs) can utilize to support a wide
variety of use cases, services and applications. In this Section,
two comprehensive implementations are presented, both
carried out by the proposed framework. The first
implementation involves a multi-slice network deployment,
where each new slice can be instantiated on-the-fly, without
any disruption to other parts of the network. The second
implementation includes a real-time network configuration in
terms of UPF nodes, which is based on predicted compute
resource requirements provided by the analytics block of the
framework. The details for the two implementations are
provided in Subections IV.B and IV.C.
A. EXPERIMENTAL TESTBED
The lab testbed along with the related open source components
used is depicted in Figure 5. The physical infrastructure
comprises a set of servers, optoelectronic switches, routers and
physical links. All the physical resources are clustered into a
(openstack) private Cloud platform that is used to deploy and
host all 5G RAN & Core functions. The 5G CN is deployed
×
Physical Infrastructure
Energy
Monitoring
(Smart PDUs)
CPU, Network
Resource
Monitoring
Visualization
Monitoring System
Figure 5: Lab testbed overview.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
10
through the free5gc open-source platform while for the RAN
side, UERANSIM is used. The monitoring and data collection
platform stores, visualizes and monitors the system resources.
Metrics from all compute and network resources are collected
and exposed from agents to specific ports and are then
retrieved and stored in the Prometheus database. Additionally,
energy consumption metrics are retrieved from energy
metering devices. All the metrics are visualized through
Grafana. Finally, OSM is used as a MANO platform for the
automated deployment of Control Plane and User Plane Core
functions.
B. AUTOMATED LCM OF NETWORK SLICES
The first part of the implementation includes an automated
deployment of a functional 5G network with two configured
slices as shown in Figure 6. The environment is hosted in a
total of eight virtual instances. The first slice consists of an
SMF node, two UPF nodes and one Local Breakout loop
targeting delay sensitive applications that are served through a
local MEC node. The second slice is deployed with one SMF
and one UPF node. Following the SBA paradigm, dedicated
subnets are created for the external interfaces (i.e. N2, N3, N4,
N6, N9) and each virtual instance uses a separate virtual
network interface to enable isolation and ease the monitoring
procedure of each protocol. For the internal interfaces/NFs a
loopback network is used.
The 5G core is automatically instantiated through the LCM
platform. The LCM framework maps the CP and UP NFs from
free5gc to two different VNFs. Their combination to create a
specific slice is described in the NST. Figure 7 shows the two
NSTs that were developed for the two slices. The two slices
share the CP elements but have distinct UPF nodes.
The deployment is shown in Figure 8. First, the main 5G
network elements (CP, I-UPF, and PSA-UPF) for the
“Internet” slice are created, and the slice is instantiated (Figure
8a-b). A new slice, 'IMS,' is then requested. This requires the
addition of a new UPF node. The orchestrator is aware of the
already instantiated shared CP elements and only instantiates
the new UPF node for this slice. Once the 'IMS' slice is
instantiated (Figure 8c-d), the platform supports both the
'Internet' and 'IMS' slices.
In this procedure, initially OSM takes the responsibility to
instantiate the slice, by allocating the necessary resources.
Then the LCM framework performs a set of configuration
actions to enable the required functionality of the two 5G
planes. These include:
- Manipulation of the configuration files according to the
input parameters for appropriate operation of the 5G
core and UPF
- Configuration of the UPF role for the 5G UP (I-UPF or
PSA with N3/N9 interface)
- Load of the GTPU tunnel module to the VM that hosts
the 5G UP
MEC
DN2
gNB
I-UPF PSA1-
UPF
SMF1
PSA2-
UPF
SMF2
UE1
UE2 DN1
VM1
N3 N9
N4
N4 Slice 1
Slice 2
SBI
VM3
VM4
VM6
VM8
VM7
VM5
N3 N6
N6
Shared CP
elements
VM2
Figure 6: 5G topology considered for automated LCM. The topology consists of two slices that have distinct user plane paths, and SMFs,
but share the other CP 5G NFs.
VLD: N6
VLD: N2
VLD: N3
VLD: N4
VLD: N9
NSSI: CP
(NSD: 5G CP)
sapd
sapd
NSSI: UPF
(NSD: 5G UPF)
sapd
NSSI: iUPF
(NSD: 5G UPF)
sapd
NSSI: PSA
(NSD: 5G UPF)
N6 N6 N6 N6
N3 N3 N3
N4 N4N2
N9 N4 N9 N4 N9
NSI: IMS NSI: INTERNET
SHARED NS!
Figure 7: Graphical illustration of two NSTs for two 5G slices.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
11
Finally, the LCM starts the CP and UP NFs, and the slice
becomes operational. It is important to note that LCM can
perform this type of actions (start/stop NFs) dynamically
during the deployment of the slice.
C. ZSM MECHANISM FOR LOAD DATA TRAFFIC
MANAGEMENT
The second part of the implementation presented in this work
aims to employ a mechanism that follows the ZSM paradigm.
The mechanism relies on all four building blocks that were
presented in Section III; hence, the analytics block is added.
The idea is to develop a framework that optimizes resource
utilization without human intervention, in a dynamic
environment where data traffic load can vary significantly in
time. Specifically, the framework includes:
a) A simple 5G CN topology based on the CP/UP split
and one gNB.
b) City-wide mobile network traffic statistics: The
second set of measurements that are used for the
development of the AI/ML models include large
scale mobile network traffic statistics that are
available online [52]. The traffic statistics were
captured from a Base Station with a varying number
of connected users over a two-week period.
c) 5G related measurements collected from our lab
testbed. The network traffic statistics were mapped
to 5G network and compute resources through iperf
connections. For each user, an iperf connection was
made with a static allocated bandwidth. The
instances that host the 5G environment are
continuously monitored and the resulting data are
collected and stored in the Prometheus database.
d) A two-stage NN model implemented to proactively
react to traffic fluctuations. The first stage includes
c)
a)
d)
b)
NSSI
CP
NSSI
iUPF
NSSI
PSA
NSI
Internet
Creation of a new
service Slice
NSSI
UPF
NSSI
CP
NSSI
iUPF
NSSI
PSA
NSI
Internet
NSI
IMS
Figure 8: MANO-triggered Slice instantiation: a) NS Instances: The main 5G networks elements (CP, iUPF and PSA) implementing the
“Internet” Slice shown in Figure 7 are created. b) The “Internet” slice is instantiated. c) Creation of a new slice (IMS) is requested. This
requires a new UPF node to be added. d) The IMS slice is instantiated. The platform now supports the “Internet” and the “IMS” slices.
Time
CPU (%)
Figure 9: Recorded compute resource utilization measurements
for virtualized UPF. (blue line: raw data set, orange line: training
dataset, green line: test dataset)
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
12
an LSTM-based forecasting algorithm that collects
and processes the monitored data to predict future
load data traffic. The second stage includes a
Decision algorithm that evaluates the predictions and
triggers the instantiation of new UPF nodes to split
the data traffic.
A numerical example showing the prediction of the UPF CPU
load, based on the available history CPU data that are
extracted from our lab’s monitoring system is shown in Figure
9. The LSTM input vector corresponds to the total UPF load
at an arbitrary time step t, while the LSTM output vector
corresponds to the total load at time step t+1. To train the
LSTMs, the dataset containing history measurements of each
UPF is split into two parts, the training set and the test set. The
training set is used during the training of the LSTM network,
while the test set is used to validate the effectiveness of each
LSTM designed.
During the predictive stage, a critical parameter that should be
carefully considered in the decision-making process is the
prediction horizon. The prediction horizon should be at least
equal to the time needed by the system to calculate and then
apply the optimal reconfiguration and resource allocation
policies (i.e., modify/add/delete routing paths or compute
resources). To determine the prediction horizon, we measured
the time needed from OSM to instantiate a UPF instance in our
private cloud. The relevant results are shown in Figure 10.
The model at each time 𝑡 predicts the value of required
compute resources to accommodate the traffic at the next
timestep. If this value exceeds a predefined threshold, the
algorithm triggers instantiation of a new UPF node creating an
SSH connection to the orchestrator. The threshold depends on
the capacity of the UPF node (250 Mb/s in our case). The
prediction horizon of the algorithm is set to 4min so that the
MANO can apply appropriate configurations in time, as
discussed earlier.
Once the prediction exceeds the threshold, the new UPF node
is instantiated. The resulting topology has now two available
UPF nodes to handle the incoming traffic which is distributed
in both available instances. Figure 11 provides system
performance snapshots from Grafana. Figure 11a) shows the
distribution of traffic without the implementation of the
algorithm (single-VM) and in Figure 11b), the traffic is
distributed among two nodes. The proposed framework
optimizes the utilization of resources by allocating network
and compute resources when they are needed based on the
traffic predictions and releasing them when load traffic
decreases. The addition of the extra node leads to extended
network capacity for the system and avoids stretching the
available resources of a single node above its optimal levels,
leading to a smooth system behavior.
To validate the efficiency of the developed framework, we
consider a scenario with multiple gNBs each served by a
single UPF. Similar network patterns are assumed for all the
nodes in the system. Based on the data traffic statistics
extracted from the monitoring platform, a decision must be
made whether the deployment of additional VM(s) is needed
or not. For this decision we define three strategies (flat, time-
margined, predictive) and compare their performance in terms
of packet delivery. Specifically, the flat strategy generates the
deployment of a VM when network traffic reaches the
threshold (that was defined before), the time-margined
strategy generates the VM when the traffic levels reach the
threshold minus a time margin and the predictive strategy
initiates the VM deployment based on the predictions of the
(a) (b)
Figure 11: System Performance Evaluation a) Single-upf traffic distribution b) Two-upf traffic distribution
Figure 10: UPF instantiation time
0
10
20
30
40
50
60
70
80
0 2 4 6 8 10 12
Time (m)
Number of VMs
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
13
algorithm. The graph in Figure 12 illustrates the results of each
strategy for varying number of VMs in terms of packet loss,
where the flat strategy presents the highest levels of packet
loss, followed by the time-margined approach. The predictive
strategy achieves the best performance since packet loss is
only related to the prediction error. Similarly, Figure 13
presents the Packet Loss Ratio (PLR) for each approach. It is
worth mentioning that the time-margined strategy performs
very well, especially for a small number of VMs but at the
expense of risking an erroneous instantiation, i.e. the load
traffic reaches the time-margined threshold and then
decreases. In our dataset, twelve cases of erroneous
instantiation were observed.
V. CONCLUSIONS
In this paper, a MANO framework based on OSM has been
proposed and developed, that specifically targets the
orchestration operations of B5G networks. We have created
network descriptors for the core and the user plane network
elements. Combining those descriptors, we can successfully
deploy dynamic network slices. In order to test the validity and
performance of the proposed framework we demonstrated two
use cases. The first focuses on the dynamic deployment of
network slices on top of a softwarized multi-operator 5G
platform hosted in our private lab testbed. The second part of
the implementation concentrated on the demonstration of a
proactive UPF provisioning mechanism, ensuring that the
system can detect on time the compute and network demands
of a slice, that may change dynamically, and adapt to these
demands accordingly. Both cases highlight the ZSM approach
of our network, and its capability to directly and dynamically
manage 5G elements and optimize resource utilization.
ACKNOWLEDGMENT
This work has been financially supported by the EU projects
5G-TACTIC under grant agreement 101127973, 6G-SENSES
under grant agreement 101139282 and ECO-eNET under
grant agreement 101139133.
REFERENCES.
[1] Y. Mao, C. You, J. Zhang, K. Huang and K. B. Letaief, “A Survey
on Mobile Edge Computing: The Communication Perspective,” in
IEEE Communications Surveys & Tutorials, vol. 19, no. 4, pp. 2322-
2358, Fourthquarter 2017. Available: https://doi.org/10.1109.
[2] "Cloud RAN and MEC: A Perfect Pairing", Etsi.org, 2019. [Online].
Available:
https://www.etsi.org/images/files/ETSIWhitePapers/etsi_wp23_ME
C_and_CRAN_ed1_FINAL.pdf.
[3] Barakabitze, Alcardo & Ahmad, Arslan & Hines, Andrew &
Mijumbi, Rashid, “5G Network Slicing using SDN and NFV: A
Survey of Taxonomy, Architectures and Future Challenges.”, in
Computer Networks, Vol. 167, 106984, 2020. Available:
https://doi.org/10.101.
[4] Rommer, S., Hedman, P., Olsson, M., Frid, L., Sultana, S., Mulligan,
C., 5G Core Networks: Powering Digitalization, ISBN:
9780081030097, Elsevier Science, 2019.
[5] Imoize, Agbotiname & Adedeji, Oluwadara & Tandiya, Nistha &
Shetty, Sachin. (2021). 6G Enabled Smart Infrastructure for
Sustainable Society: Opportunities, Challenges, and Research
Roadmap. Sensors. 21. 1-57. 10.3390/s21051709.
[6] https://5g-ppp.eu/verticals/.
[7] “3GPP, TR 22.864, "Feasibility study on new services and markets
technology enablers for network operation; Stage 1",” [Online].
[8] M. E. Morocho-Cayamcela, H. Lee and W. Lim, "Machine Learning
for 5G/B5G Mobile and Wireless Communications: Potential,
Limitations, and Future Directions," in IEEE Access, vol. 7, pp.
137184-137206, 2019, doi: 10.1109/ACCESS.2019.2942390.
[9] J. Kaur, M. A. Khan, M. Iftikhar, M. Imran and Q. Emad Ul Haq,
"Machine Learning Techniques for 5G and Beyond," in IEEE Access,
vol. 9, pp. 23472-23488, 2021, doi:
10.1109/ACCESS.2021.3051557.
[10] ETSI NFV, see http://www.etsi.org/technologies‐
clusters/technologies/nfv.
[11] 3GPP, “TS 28.531 Management and Orchestration; Provisioning,”
January 2020.
[12] ETSI. Open Source MANO. [Online] OSM (etsi.org)
[13] The Linux Foundation. ONAP. [Online] https://www.onap.org/.
[14] Liyanage, M., Pham, Q. V., Dev, K., Bhattacharya, S., Maddikunta,
P. K. R., Gadekallu, T. R., & Yenduri, G. (2022). A survey on Zero
touch network and Service Management (ZSM) for 5G and beyond
networks. Journal of Network and Computer Applications, 203,
103362.
[15] 3GPP, TR 22.863, "Feasibility study on new services and markets
technology enablers for enhanced mobile broad-band; Stage 1"
[Online].
Figure 12: System Packet Loss for varying number of VMs. Three
different strategies have been employed for the instantiation of the
VMs (flat, time-margined, predictive).
0
20
40
60
80
100
120
140
160
0 2 4 6 8 10 12
PACKET LOSS (PACKETSx10^6)
NUMBER OF VMS
Flat Time-Margined Predictive
Figure 13: Packet Loss Ratio for varying number of VMs for
the three different approaches (flat, time-margined, predictive).
0
0.05
0.1
0.15
0.2
1 2 4 8 1 0 1 2
PACKET LOSS RATIO
NUMBER OF VMS
Flat Time-Margined Predictive
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
14
[16] 3GPP, TR 22.861, "Feasibility Study on New Services and Markets
Technology Enablers for massive Internet of Things; Stage 1",
[Online].
[17] “3GPP, TR 22.862, "Feasibility study on new services and markets
technology enablers for critical communications; Stage 1", [Online].
[18] "Microservices.io (2020). What are microservices? [Online].
Available: https://microservices.io/".
[19] Ian F. Akyildiz, Shuai Nie, Shih-Chun Lin, Manoj Chandrasekaran,
“5G roadmap: 10 key enabling technologies”, Computer Networks,
Volume 106, 2016, Pages 17-48, ISSN 1389-1286. Available:
https://doi.org/10.1016/j.comnet.2016.06.010.
[20] 5G; NR; NR and NG-RAN Overall description; Stage-2, ETSI TS
138 300 V16.4.0 (2021c). [Online]
[21] Network Functions Virtualisation (NFV); Terminology for Main
Concepts in NFV, ETSI GS NFV 003 V1.2.1 (2014). [Online].
[22] F. T. Kuhn, F. Schnicke and P. O. Antonino, “Service-Based
Architectures in Production Systems: Challenges, Solutions &
Experiences,” in ITU Kaleidoscope: Industry-Driven Digital
Transformation (ITU K), Ha Noi, Vietnam, 2020.
[23] View on 5G Architecture, 5G PPP Architecture Working Group,
Version 3.0, 06-2019. [Online]. Available: https://5g-ppp.eu/wp-
content/uploads/2019/07/5G-PPP-5G-Architecture-White-
Paper_v3.0_PublicConsultation.pdf
[24] System architecture for the 5G System (5GS), 3GPP TS 23.501
version 16.6.0 Release 16. [Online].
[25] Akyildiz, Ian & Nie, Shuai & Lin, Shih-Chun & Chandrasekaran,
Manoj, “5G Roadmap: 10 Key Enabling Technologies,” in Computer
Networks, Vol 106, 2016. Available:
https://doi.org/10.1016/j.comnet.2016.06.010.
[26] Network Functions Virtualisation (NFV); Architectural Framework,
ETSI GS NFV 002 V1.2.1 (2014a). [Online].
[27] Network Functions Virtualisation (NFV); Management and
Orchestration, ETSI GS NFV-MAN 001 V1.1.1 (2014b). [Online].
[28] X. Li et al., "Automated Service Provisioning and Hierarchical SLA
Management in 5G Systems," in IEEE Transactions on Network and
Service Management, vol. 18, no. 4, pp. 4669-4684, Dec. 2021, doi:
10.1109/TNSM.2021.3102890.
[29] Dreibholz, T. (2020). Flexible 4G/5G Testbed Setup for Mobile Edge
Computing Using OpenAirInterface and Open Source MANO. In:
Barolli, L., Amato, F., Moscato, F., Enokido, T., Takizawa, M. (eds)
Web, Artificial Intelligence and Network Applications. WAINA
2020. Advances in Intelligent Systems and Computing, vol 1150.
Springer, Cham
[30] N. Makris, C. Zarafetas, A. Valantasis and T. Korakis, "Service
Orchestration Over Wireless Network Slices: Testbed Setup and
Integration," in IEEE Transactions on Network and Service
Management, vol. 18, no. 1, pp. 482-497, March 2021
[31] R. Inam, A. Karapantelakis, K. Vandikas, L. Mokrushin, A.
Vulgarakis Feljan and E. Fersman, "Towards automated service-
oriented lifecycle management for 5G networks," 2015 IEEE 20th
Conference on Emerging Technologies & Factory Automation
(ETFA), Luxembourg.
[32] NGMN 5G Initiative Team, “A Deliverable by the NGMN Alliance:
NGMN 5G White Paper”. [Online]. Available:
https://www.ngmn.org/. wp-
content/uploads/NGMN_5G_White_Paper_V1_0.pdf.".
[33] Study on management and orchestration of network slicing for next
generation network (Release 15), Technical Specification Group
Services and System Aspects, Telecommunication management,
3GPP TR 28.801 V15.1.0, January 2018.
[34] E. Chirivella-Perez, P. Salva-Garcia, I. Sanchez-Navarro, J. M.
Alcaraz-Calero and Q. Wang, "E2E network slice management
framework for 5G multi-tenant networks," in Journal of
Communications and Networks, vol. 25, no. 3, pp. 392-404, June
2023, doi: 10.2.
[35] P. Vanichchanunt, O. Ritruechai, N. Wuttiananchai, P. Thossaporn,
L. Wuttisittikulkij and S. Paripurana, "Implementation of 5G
Network Slicing Using Open Source Software," 2024 12th
International Electrical Engineering Congress (iEECON), Pattaya,
Thailand, 2024, pp. 1-6, doi:
10.1109/iEECON60677.2024.10537902.
[36] X. Li, R. Ni, J. Chen, Y. Lyu, Z. Rong and R. Du, "End-to-End
Network Slicing in Radio Access Network, Transport Network and
Core Network Domains," in IEEE Access, vol. 8, pp. 29525-29537,
2020, doi: 10.1109/ACCESS.2020.2972105.
[37] X. Shen et al., "AI-Assisted Network-Slicing Based Next-Generation
Wireless Networks," in IEEE Open Journal of Vehicular Technology,
vol. 1, pp. 45-66, 2020, doi: 10.1109/OJVT.2020.2965100.
[38] B. Zhao, H. Lu, S. Chen, J. Liu and D. Wu, "Convolutional neural
networks for time series classification," in Journal of Systems
Engineering and Electronics, vol. 28, no. 1, pp. 162-169, Feb. 2017,
doi: 10.21629/JSEE.2017.01.18.
[39] Wen, Q., Zhou, T., Zhang, C., Chen, W., Ma, Z., Yan, J., & Sun, L.
(2022). Transformers in time series: A survey. arXiv preprint
arXiv:2202.07125.
[40] S. Zhao, S. Lin and J. Xu, "Time Series Traffic Prediction via Hybrid
Neural Networks," 2019 IEEE Intelligent Transportation Systems
Conference (ITSC), Auckland, New Zealand, 2019, pp. 1671-1676,
doi: 10.1109/ITSC.2019.8917383.
[41] K. Samdanis, A. N. Abbou, J. Song and T. Taleb, "AI/ML Service
Enablers and Model Maintenance for Beyond 5G Networks," in IEEE
Network, vol. 37, no. 5, pp. 162-172, Sept. 2023, doi:
10.1109/MNET.129.2200417.
[42] K. Govindarajan et al., "Closed loop optimization of 5G network
slices," 2023 15th International Conference on COMmunication
Systems & NETworkS (COMSNETS), Bangalore, India, 2023, pp.
186-188, doi: 10.1109/COMSNETS56262.2023.10041293.
[43] N. Slamnik-Kriještorac et al., "AI-Empowered Management and
Orchestration of Vehicular Systems in the Beyond 5G Era," in IEEE
Network, vol. 37, no. 4, pp. 305-313, July/August 2023, doi:
10.1109/MNET.008.2300024.
[44] L. Blanco et al., "AI-Driven Framework for Scalable Management of
Network Slices," in IEEE Communications Magazine, vol. 61, no.
11, pp. 216-222, November 2023, doi:
10.1109/MCOM.005.2300147.
[45] M. Ramachandran, T. Archana, V. Deepika, A. A. Kumar and K. M.
Sivalingam, "5G Network Management System With Machine
Learning Based Analytics," in IEEE Access, vol. 10, pp. 73610-
73622, 2022, doi: 10.1109/ACCESS.2022.3190372.
[46] A. Sheoran, S. Fahmy, L. Cao and P. Sharma, "AI-Driven
Provisioning in the 5G Core," in IEEE Internet Computing, vol. 25,
no. 2, pp. 18-25, 1 March-April 2021, doi:
10.1109/MIC.2021.3056230.
[47] J. Jeong et al., "Mobility Prediction for 5G Core Networks," in IEEE
Communications Standards Magazine, vol. 5, no. 1, pp. 56-61, March
2021, doi: 10.1109/MCOMSTD.001.2000046.
[48] S. D’Oro, L. Bonati, M. Polese and T. Melodia, "OrchestRAN:
Orchestrating Network Intelligence in the Open RAN," in IEEE
Transactions on Mobile Computing, vol. 23, no. 7, pp. 7952-7968,
July 2024, doi: 10.1109/TMC.2023.3342711.
[49] Service requirements for the Evolved Packet System (EPS), 3GPP TS
22.278 version 13.4.0 Release 13. [Online]".
[50] D. Giannopoulos, P. Papaioannou, C. Tranoris and S. Denazis,
"Monitoring as a Service over a 5G Network Slice," 2021 Joint
European Conference on Networks and Communications & 6G
Summit (EuCNC/6G Summit), Porto, Portugal, 2021, pp. 329-334,
doi: 10.1109/E.
[51] Juju, [Online] https://juju.is
[52] T. Italia, ""Telecommunications - SMS, Call, Internet - MI", Harvard
Dataverse, V1," 2015. [Online]. Available:
https://doi.org/10.7910/DVN/EGZHFV.
Alexandros – Ioannis Manolopoulos is a PhD
candidate in the Physics Department at
National and Kapodistrian University of Athens
(NKUA) and a researcher at the Institute of
Accelerating Systems and Applications (IASA).
His main research interests focus on mobile
networks, and in particular on Virtualization
technologies in Cloud- Core & Radio Access
Networks. He graduated from the Department
of Informatics and Telecommunications of the University of
Peloponnese. He holds a master’s degree in Electronics and
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
15
Radioelectrology, from the Interfaculty Program of Postgraduate
Studies (IPPS) at the NKUA, Greece.
Viktoria-Maria Alevizaki holds a PhD form the
National and Kapodistrian University of Athens
(NKUA), Greece and is a post-doctoral
researcher at the Institute of Accelerating
Systems and Applications (IASA) and NKUA.
Her main research interests center around
mobile networks (5G/B5G), and in particular
network design and resource management.
She holds a bachelor’s degree from the
Department of Physics of NKUA and a master’s degree in Electronics
and Radioelectrology, from the same University. She is the recipient
of a doctoral scholarship co-financed by Greece and the European
Union (European Social Fund-ESF).
Dr. Markos Anastasopoulos Markos
Anastasopoulos is an Associate Professor at
the National and Kapodistrian University of
Athens (NKUA),, Greece. Previously he was a
senior researcher at the High-Performance
Networks (HPN) Group of the University of
Bristol, UK. He is an author/co-author of more
than 100 papers in peer reviewed international
journal and conference proceedings.
Dr. Anna Tzanakaki is an Associate Professor
at the National and Kapodistrian University of
Athens, Greece. Previously she was Associate
Professor at the Athens Information
Technology, Greece and an adjunct faculty
member of the Information Networking Institute
of Carnegie Mellon University, USA. She was a
Senior Research Officer and a visiting lecturer
at the University of Essex and a co-founder of
ilotron ltd, a high-technology spin-off from the University of Essex.
She is a co-author of over 200 publications in international journals
and conferences and a co-inventor of several granted and published
patents. She is a senior member of the IEEE. She has participated to
a number of European and national research projects and is the
Technical Coordinator of the 5G PPP Phase 2 and Phase 3 projects
5G-PICTURE and 5G-VICTORI respectively. Her research interests
include converged networks and compute infrastructures,
architectures, technologies and protocols.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3507359
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
