1. Introduction

The rapid development of Internet technologies in the past two decades has enabled a broad spectrum of network applications with highly diverse service requirements and a wide variety of technologies for building heterogeneous network infrastructures. However, the ossification of the IP-based network architecture limits the Internet’s flexibility and agility to face the challenges introduced by the diversity in both network applications and infrastructures ^[1].

Network virtualization has been proposed as a key attribute of future networks for overcoming the limitation of the IP-based network architecture. The key idea of network virtualization is to decouple the network functions for service-provisioning from the network/compute capabilities for data transportation, processing, and storage, which allows alternative network architectures and protocols to coexist upon shared underlying infrastructures ^[2]. The Network Function Virtualization (NFV) paradigm follows the network virtualization principle to realize network functions as software instances that can be deployed upon commodity servers and storage devices, thus enabling virtual networks (network slices) customized for multiple tenants ^[3].

A Software-Defined Network (SDN) is another innovative networking technology that may address the ossification of IP-based networks. The key idea of the SDN lies in decoupling between the data plane and control plane to enable a logically centralized network controller with a global view of the entire network domain ^[4]. The SDN introduces a network operating system that provides an abstraction of data plane functionalities and resources upon which control applications may program network operations.

NFV and the SDN have rapidly become active research areas in networking that attract extensive interest from both academia and industry. Significant progress has been made in technology developments for enabling these two new paradigms in various networking scenarios, including data center networks, wide-area networks, and wireless mobile networks ^[5][6]. NFV and the SDN together form the foundation of future networks including the 5G and beyond networks. On the other hand, the current NFV and SDN implementations inherit some monolithic design patterns from the traditional network architecture that may constraint network and service flexibility and agility ^[7].

As the next step in the evolution of networking technologies toward fully exploiting the advantages of network virtualization and softwarization, cloud-native network design applies the service-oriented architectural principle together with virtualization and softwarization in networking, which enables network systems to be realized based on cloud technologies and network services to be provisioned following the cloud service model ^[8]. The cloud-native design principle has been embraced by all major network Standards Development Organizations (SDOs) in their latest work, for example ETSI NFV Release-4 and the 3GPP 5G core network architecture. Therefore, cloud-native design is expected to be a key attribute of future networks ^[9].

Cloud-native network design facilitates a significant transformation of communication networks from an infrastructure for data transportation to a versatile service platform for various industry verticals and customer segments, which plays a crucial role in supporting both cloud computing and the emerging edge computing paradigm. The same set of key technologies—virtualization and service-oriented architecture—now are widely applied in both cloud/edge computing and networking, thus allowing these two fields to converge together^[10]. Therefore, network-cloud/edge convergence with a holistic vision of end-to-end service provisioning across the networking and computing domains is expected to be a key attribute of the future cloud-native networks.

Flexible and effective management is one of the most critical challenges to cloud-native future networks. Given the large scale, heterogeneity, and complexity of the converged communication-compute-storage systems in future networks, management solutions must be highly automated and intelligent. The integration of dynamic intelligent network telemetry and closed-loop control mechanisms leveraging machine learning (ML) and data analytics techniques offers promising new management technologies. A wide spectrum of technologies, often developed based on various ML methods, has been proposed to enable intelligent and autonomous management for various aspects of future networks ^[11]. However, the heterogeneous nature of ML techniques and the unique characteristics of future networking technologies impose a variety of requirements for integration between these areas. The current disparate management mechanisms for ML functionalities and network functions may degrade the effectiveness of intelligent network operations.

Therefore, it becomes vital that a wide range of intelligent management technologies is applied based on the foundation of a consistent architectural framework with a common holistic vision. This calls for standardization of new management architectures that support seamless integration of various ML/data analytics techniques for addressing the management challenges in a holistic and interoperable way. Currently, various relevant network SDOs, including 3GPP, ETSI, and ITU-T, are actively working on this important subject and making encouraging progress.

The goal of this paper is to provide a big picture that reflects the latest developments of architectural frameworks for intelligent and autonomous management in cloud-native future networks. The paper first gives an overview of the technical trend toward cloud-native network design and network-cloud/edge convergence in future networks. Then, the paper surveys the latest representative progress in the standardization of the management architecture for future networks, including works by 3GPP, ETSI, and ITU-T, and analyzes how cloud-native network design may facilitate the architecture development for addressing the management challenges in future networks. Open issues related to intelligent and autonomous management from the architectural perspective are also discussed in the paper to identify some directions for future research and development.

2. Cloud-Native Network Architecture in Future Networks

2.1. Network Function Virtualization and Software-Defined Networking

ETSI has been a leading force for realizing network virtualization since it defined the NFV architecture in 2012. The ETSI NFV Industry Specification Group (ISG) community has evolved through several phases from defining the initial architectural framework to developing detailed standard specifications. NFV enables network functions to be realized as software instances and hosted by virtual machines or containers running on commodity network/compute equipment such as standard servers and switches. The ETSI NFV architecture comprises a set of Virtual Network Functions (VNFs) deployed upon a virtualized infrastructure. The Management and Orchestration (MANO) system of the architecture contains a Virtual Infrastructure Manager (VIM) for unified management of network-compute resources in the virtualized infrastructure, a VNF Manager (VNFM) for managing VNFs, and an NFV Orchestrator (NFVO) for life-cycle management of network services ^[12].

Recent developments of the NFV specifications in Releases 3 and 4 have focused on enriching the NFV architecture for global deployment and operations. The set of new features in NFV mainly includes support for the latest network technologies such as edge computing and network slicing, new operational aspects such as multiple administrative domains and policy frameworks, advances in acceleration technologies and lightweight (e.g., container-based) virtualization platforms, and enhancement of NFV automation capabilities for improving life-cycle management and orchestration of VNFs and network services.

Multiple network SDOs, including ONF, ITU-T, and IETF, have developed their own versions of SDN specifications, but they all share the same basic architecture—an architecture that comprises three planes (the data plane, control plane, and application plane) and two standard interfaces (the southbound interface between the data and control planes and the northbound interface between the control and application planes). The key principles behind both SDN and NFV are decoupling with abstraction, but focus respectively on the layer dimension (for NFV) and plane dimension (for SDN) ^[13]. Therefore, NFV and the SDN are expected to be integrated in future networks in order to fully exploit the advantages of both paradigms. For example, multiple SDN controller instances may be realized upon a network hypervisor in order to control different virtual networks (network slices) implemented by NFV ^[14].

2.2. Cloud-Native Network Design

The current specifications for realizing network virtualization and softwarization, including the NFV and SDN architectures, mainly change how network functions are realized and deployed (as software instances hosted on VMs and/or containers), but do not change much how network functions are designed. The state-of-the-art NFV implementations often replace monolithic hardware-based network functions with their monolithic software VNF counterparts. Such a monolithic VNF design may introduce a large number of common functionalities repeated across different VNFs, which causes some negative consequences that lead to sub-optimal resource usage and hinder network agility. In addition, the NFV and SDN architectures both comprise a set of predefined function blocks that are interconnected via point-to-point interfaces (reference points), which require standardization of a new set of reference points whenever a new function block is added into the architecture. Such monolithic and tight-coupling architectural features still introduce ossification, which limits the flexibility and agility of networking systems for service provisioning.

A promising strategy for addressing the ossification issue associated with the current NFV and SDN architecture is to enable finer granularity for network functions in the architecture and a common interface for loose-coupling interaction among network functions. The Service-Oriented Architecture (SOA) with its latest development as the Micro-Service Architecture (MSA) offers an effective approach to achieving this objective.

In general, the SOA principle advocates decomposing a large system into a collection of smaller units called services, which essentially are self-contained and platform-independent system modules that can be described, published, accessed, composed, and programmed through a standard interface and messaging protocol ^[15]. The MSA can be viewed as the second iteration of the SOA that aims to strip away unnecessary levels of complexity in service design in order to focus on the programming of simple services that effectively implement a single functionality ^[16]. Service instances in the SOA typically run on VMs, and the lightweight container-based virtualization offers an efficient platform for hosting microservices. The SOA and virtualization have been two key pillars of the foundation for cloud technologies ^[17], and the MSA is an enabler of the latest developments in cloud computing, including the emerging edge computing paradigm.

The service-oriented architectural principle has also been adopted in network design, which enables a Network-as-a-Service (NaaS) paradigm that abstracts various network functions and resources as services ^[18]. It is worth noting that the term “service” has a different meaning in the networking field than the “service” concept in the SOA/MSA context. In the networking context, a service typically refers to the data transport capability offered to a customer by an ordered set of network functions; while the service concept in the SOA/MSA emphasizes encapsulation of a system module into a self-contained component. On the other hand, the evolution of network design has been influenced by the SOA principle in the past decade ^[19], and the SOA service concept has been gradually adopted in the recent development of the network architecture with the NaaS paradigm ^[20].

In the ETSI NFV specifications, while a network service refers to an ordered set of (virtual) network functions specified by a service description (VNF forwarding graph), the SOA principle has been embraced by the NFV architecture at multiple levels. For example, NFV supports NFV Infrastructure-as-a-Service (NFVIaaS), Virtual Network Function-as-a-Service (VNFaaS), and Network Slice-as-a-Service (NSaaS), which all adopt the SOA service concept. In the 3GPP specifications for the 5G service-based architecture, usage of the term “service” becomes more aligned with the SOA service concept than its conventional definition in the networking field. Currently, the SOA-based NaaS paradigm encapsulates monolithic network functions into individual service components with coarse granularity. Adoption of the emerging MSA in network design allows the decomposition of monolithic network functions to fine-grained modules that can be flexibly orchestrated for achieving different functionalities for service provisioning.

The application of the virtualization and service-oriented principles in network design enables network systems to be realized based on cloud technologies and network services to be provisioned following the cloud service model. This emerging trend in networking technologies’ development is often referred to as cloud-native network design, which is expected to be widely adopted in future networks including the design of 5G/6G networks. An important aspect of the recent development in the NFV specification is to support cloud-native network design including the adoption of container-based virtualization technologies for deploying VNF components as microservices ^[21].

2.3. Service-Based Architecture for the 5G Network

The Service-Based Architecture (SBA) developed by 3GPP for the 5G core network is a representative case of cloud-native design in network architecture. The 5G SBA is based on network virtualization and softwarization: Network Functions (NFs) on the data plane and control plane are separated, and the NFs defined in the architecture may be realized as software instances deployed on a virtual infrastructure. On top of virtualization and softwarization, the 5G SBA adopts the service-oriented principle through the notion of NF services exposed by control plane NFs and the Service-Based Interface (SBI) for interactions among NFS services.

As depicted in Figure 1, the 5G SBA decouples the data plane and the control plane. The data plane comprises User Equipment (UE), the Radio Access Network (RAN), the User Plane Function (UPF), and the Data Network (DN). The interfaces between the data plane functions currently are still defined as reference points (N1–N6 in Figure 1). The 5G control plane comprises a set of NFs including the Access and Mobility Management Function (AMF), the Authentication Server Function (AUSF), the Session Management Function (SMF), the Policy Control Function (PCF), the Network Slice Selection Function (NSSF), Unified Data Management (UDM), the Unified Data Repository (UDR), the Network Repository Function (NRF), the Network Exposure Function (NEF), and the Application Function (AF) ^[22].

Figure 1. The Service-Based Architecture (SBA) for the 5G network.

Each control plane NF may expose one or multiple NF services as a service provider and also access the services exposed by other NFs as a service consumer. All the NF services are accessed through the SBI. With the SBI, the 5G SBA shifts from a point-to-point reference point-based interface toward web-based communications among NF services. The SBI may be realized by means of REST API calls using HTTP/2 on top of TLS/TCP as the transport protocol. The emerging transport protocol Quick UDP Internet Connections (QUIC) is paving the way toward HTTP/3 for supporting SBI communications ^[23].

The 5G SBA allows each service consumer to discover a suitable producer of a service instance. Service discovery in the SBA is supported by the NRF, which keeps a repository of all available NF instances and the services they expose. Each NF instance is required to register its profile at the NRF, and the NF profile contains relevant data about the NF including the services it provides and the binding information for accessing the services. The NEF may expose NF services provided inside the network to authorized external consumers such as AFs. The 3GPP introduced the service framework support function, referred to as Service Communication Proxy (SCP) in Release 16, which aims to extract the common NF processes related to service registration, discovery, selection, and communication binding into a unified platform ^[24][25].

The 5G SBA supports two modes of interaction between the NF service producer and consumer. In the request-response model, an NF service consumer sends a request message to the target service provided by a producer and receives the corresponding response message back from the producer. In the subscribe-notify model, an NF service consumer subscribes to an NF service for a (set of) certain event(s) and receives a notification message from the service producer whenever a subscribed event occurs.

Although the current 3GPP specification limits the SBA to the control plane of the 5G core network, the cloud-native design principle is expected to be applied in a broader scope in the future 6G network. A natural evolution beyond the current 5G SBA would be an integration of the SBA framework for the core network into an end-to-end architecture that comprises multiple planes, including the data, control, and application planes as defined in the SDN architecture, and across multiple domains including the access, transport, and core networks.

2.4. Edge Computing

The emerging edge computing paradigm essentially deploys decentralized cloud computing capabilities in the network infrastructures typically at the network edge ^[26]. The Multi-access Edge Computing (MEC) architecture developed by ETSI is a representative architectural framework for edge computing. The MEC architecture comprises two levels: the lower level consists of a set of individual MEC hosts, and the higher level forms an MEC system comprising the MEC hosts and the network connections among the hosts. Each MEC host contains an MEC platform built on top of the virtualized network-compute infrastructure, and various MEC applications are deployed on the platform. The host-level management is responsible for managing the infrastructure, platform, and MEC applications at each host. The core function of system-level management is an MEC orchestrator that maintains a global view of the entire MEC system for coordinating host management for end-to-end service provisioning ^[27].

The true impact of the edge computing paradigm relies on its interaction with networking. The foundation of the MEC architecture is a virtualized network infrastructure upon which the MEC platform and applications are deployed. Most network operators want to consolidate VNFs and MEC applications on top of a shared infrastructure to the maximum possible degree for enhancing resource utilization and improving service performance. Toward this direction, ETSI MEC ISG has developed a reference architecture for deploying MEC in an NFV environment, which specifies how MEC entities can be integrated in the NFV architecture by fully leveraging NFV MANO capabilities ^[28]. In this MEC-in-NFV reference architecture, both the MEC platform and MEC applications can be treated by NFV MANO in the same way as VNFs being managed, which allows NFV MANO to be leveraged for unified management and orchestration of both VNFs and MEC applications.

2.5. Convergence of Networking and Cloud/Edge Computing

The cloud-native network design adopts the virtualization and service-oriented architectural principles in networking essentially in the same way as the principles being applied in cloud/edge computing. Therefore, cloud-native network design enables network systems to be realized based on cloud technologies and network services to be provisioning in the cloud service model, which facilitates convergence between networking and cloud/edge computing. Such convergence calls for a holistic vision across the networking and computing domains that supports integrated management of networking-computing functionalities and unified provisioning of network and cloud/edge services.

The NFV architecture provides a unified virtualization layer for network-compute infrastructures and common management of virtual network/compute functions, thus forming the basis for network-cloud/edge convergence as reflected in the MEC-in-NFV reference architecture. Through its embrace of cloud-native design in the SBA, the 5G network introduces the very same characteristics about network function interactions, system operations, and service provisioning shared with cloud and edge computing ^[29]. Therefore, network-cloud/edge convergence with integrated network-compute infrastructures for end-to-end provisioning of composite network-cloud/edge services is expected to be a key feature of future networks.

The holistic architecture design, resource management, system operations, and service provisioning enabled by network-cloud/edge convergence may significantly improve resource utilization and service performance, lower capital/operational costs, and introduce opportunities for technical and business innovations. On the other hand, all these benefits can be realized only if the converged network-cloud/edge systems in future networks are properly managed. Standardization of the management architecture for cloud-native future networks plays a significant role in order to fully leverage various cognition techniques such as AI/ML to manage heterogeneous systems and integrate diverse management domains for end-to-end service provisioning. The following sections of this paper will review the representative standardization work on management architecture that addresses the challenges of intelligent and autonomous management in future networks.