PaperlayerView exact revision on arXiv

Source-linked AI summary

Holistic Network Virtualization and Pervasive Network Intelligence for 6G

Xuemin, Shen, Jie Gao, Wen Wu, Mushu Li, Conghao Zhou, Weihua Zhuang

arXiv:2301.00519v1cs.NIcs.AI

TL;DR

6G needs virtualization that covers both service provision and service demand, alongside AI integrated across network functions, layers, segments, and applications. The paper develops a conceptual architecture combining holistic network virtualization and pervasive network intelligence, with potential to improve management granularity, adaptivity, and AI-service QoS support. It also identifies implementation challenges and open issues for the proposed architecture.

  • Problem

    Existing virtualization primarily addresses service provision, while 6G requires attention to end user service demand and AI integration across the network.

  • Method

    The paper surveys prior work and proposes a conceptual 6G architecture combining network slicing, digital twins, pervasive AI, connected AI, AI slices, and hybrid data-model driven methods.

  • Results

    The proposed architecture has potential to improve user-centric networking and management granularity, support adaptive hybrid management methods, and provide QoS guarantees for AI services.

  • Takeaways & Limitations

    Holistic virtualization and pervasive network intelligence provide a conceptual basis for more flexible, scalable, adaptive, and intelligent 6G networks.

Abstract

from arXiv · show

In this tutorial paper, we look into the evolution and prospect of network architecture and propose a novel conceptual architecture for the 6th generation (6G) networks. The proposed architecture has two key elements, i.e., holistic network virtualization and pervasive artificial intelligence (AI). The holistic network virtualization consists of network slicing and digital twin, from the aspects of service provision and service demand, respectively, to incorporate service-centric and user-centric networking. The pervasive network intelligence integrates AI into future networks from the perspectives of networking for AI and AI for networking, respectively. Building on holistic network virtualization and pervasive network intelligence, the proposed architecture can facilitate three types of interplay, i.e., the interplay between digital twin and network slicing paradigms, between model-driven and data-driven methods for network management, and between virtualization and AI, to maximize the flexibility, scalability, adaptivity, and intelligence for 6G networks. We also identify challenges and open issues related to the proposed architecture. By providing our vision, we aim to inspire further discussions and developments on the potential architecture of 6G.

I. INTRODUCTION

The paper argues that 6G requires a new architecture extending virtualization toward service demand and integrating AI throughout networking. It proposes holistic network virtualization and pervasive network intelligence as complementary foundations for flexible, scalable, adaptive, and intelligent networks.

  • 6G network architecture must support increasingly heterogeneous networks, diverse services, and stringent QoS/QoE requirements.
  • Holistic network virtualization combines network slicing for service provision with digital twins for service demand, linking service-centric and user-centric networking.
  • Pervasive network intelligence integrates AI across networks through both AI for networking and networking for AI.
  • The proposed architecture enables interplay between digital twins and network slicing, data-driven and model-driven methods, and virtualization and AI.
  • The tutorial surveys existing virtualization and AI research, proposes original 6G ideas, and identifies challenges and open issues for further study.

D. Structure of the Paper

The paper reviews network virtualization, its historical development, benefits, techniques, and remaining scope limitations before presenting the paper’s organization. It emphasizes virtualization’s role in programmable, flexible, scalable, and cost-effective network management.

  • The paper is organized around holistic network virtualization, pervasive network intelligence, their integrated 6G architecture, related challenges, and conclusions.
  • The paper reviews virtualization techniques and benefits before introducing holistic network virtualization.
  • Network virtualization evolved from early virtual LANs toward NFV and network slicing, which creates multiple end-to-end virtual networks over shared infrastructure.
  • Virtualization can abstract nodes, links, resources, and networks, enabling virtual entities and resource pools to share physical infrastructure.
  • Common virtualization characteristics include abstraction, co-existence, and isolation, supporting simpler management, resource utilization, reliability, security, scalability, and QoS.
  • Existing network virtualization improves programmability, flexibility, scalability, and cost effectiveness but focuses mainly on infrastructure and resources rather than end users.

B. End User Virtualization

End user virtualization extends network virtualization beyond infrastructure by representing users and other network entities with digital twins. The paper identifies their benefits, requirements, implementation questions, and integration with network slicing.

  • Digital twins offer a potential paradigm for end user virtualization, complementing earlier network-hosted avatars and virtual objects.
  • Existing network research mainly studies digital-twin applications, while digital twins for network architecture and network management remain limited.
  • End user digital twins can provide service-demand and user QoS/QoE data, complementing infrastructure virtualization’s characterization of network status and service provision.
  • Digital twins should flexibly represent heterogeneous devices and applications, remain compatible with network slicing, and use customizable attributes and data.
  • Digital-twin deployment must account for the network resources consumed by creating and maintaining the twins.
  • Implementation requires decisions about hosting location, slice affiliation, data attributes and history, synchronization frequency, and model control.
  • The proposed holistic virtualization architecture integrates digital twins and network slicing through six layers spanning data collection to digital-twin model control.

C. Holistic Network Virtualization

Holistic network virtualization integrates digital twins with network slicing through a six-layer architecture spanning data collection, abstraction, processing, control, and model management.

  • C. Holistic Network Virtualization: The six-layer architecture integrates digital twins into network virtualization to improve network management and service provision.VL 1 collects data, while VL 6 controls digital twin models.
  • C. Holistic Network Virtualization: VL 1 collects prescribed end-user data through access points under the control of local edge controllers.Collection settings include data precision, uploading method, and frequency.
  • C. Holistic Network Virtualization: VL 2 forms level-one digital twins of individual end users from collected data and hosts them near local controllers.The abstraction can aggregate sources, update historical data, and add new end users.
  • C. Holistic Network Virtualization: VL 3 processes level-one twins at the edge for user prediction, service decisions, and partial edge-network emulation.Examples include traffic and mobility prediction, computing offloading, content delivery, and link-layer adaptation.
  • C. Holistic Network Virtualization: VL 4 creates level-two digital twins for network slices from aggregated user data and slice-specific configuration, utilization, and SLA information.VL 5 then supports service-demand forecasting, slice decisions, emulation, admission, reservation, and coverage control.
  • C. Holistic Network Virtualization: VL 6 updates both digital twin models using resource availability, management and service-provision performance, and dynamic spatiotemporal service demands.The architecture interfaces with local controllers at VL 1–3, the centralized controller at VL 4–5, and both at VL 6.
  • C. Holistic Network Virtualization: Digital twins enhance network slicing with organized end-user data and slice abstraction, while two-level twins avoid duplicate user twins and synchronization burden.The architecture is presented as a configurable design that exploits data for management and network emulation.

D. Holistic Network Virtualization: A Summary

Holistic network virtualization extends 5G-era service-provision virtualization to include end-user service demand through configurable digital twins and network slicing. A six-layer architecture integrates these paradigms and addresses digital-twin hosting, data, and management.

  • D. Holistic Network Virtualization: A Summary: The section identifies current network virtualization as insufficient and develops holistic virtualization to combine network and end-user virtualization.This extends virtualization from service provision toward both service-centric and user-centric networking.
  • D. Holistic Network Virtualization: A Summary: Digital twins characterize individual end users’ status and service demand using configurable assemblies of historical, real-time, collected, and generated data.Their data collection and processing are also configurable.
  • D. Holistic Network Virtualization: A Summary: The proposed six-layer architecture systematically integrates digital twins and network slicing as a reference design for holistic network virtualization.It addresses where digital twins are hosted, what data they contain, and how they are managed.

III. PERVASIVE NETWORK INTELLIGENCE

Pervasive network intelligence places AI across end users, edge, and cloud, covering both AI for networking and networking for AI. The paper organizes this vision with a four-level architecture and reviews relevant learning approaches and applications.

  • A. AI Techniques: An Overview: The overview covers unsupervised, supervised, and reinforcement learning, along with centralized, decentralized, federated, and multi-agent approaches.Reinforcement learning iteratively optimizes policies through network-state sensing and feedback, while federated learning trains on distributed edge data.
  • III. PERVASIVE NETWORK INTELLIGENCE: Pervasive network intelligence describes AI penetrating end users, the network edge, and the cloud.The trend is associated with advances in machine learning, data collection, edge and cloud computing, and programmable network control.
  • III. PERVASIVE NETWORK INTELLIGENCE: Service-oriented AI provides end-user services, while management-oriented AI supports network tasks such as power allocation and slice resource reservation.Service-oriented applications require storage, computing, and communication resources for training, inference, data collection, and model uploading.
  • III. PERVASIVE NETWORK INTELLIGENCE: The paper distinguishes AI for networking from networking for AI, expanding 6G’s AI scope beyond 5G’s focus on AI in communications.AI for networking manages networks, while networking for AI designs networks that facilitate AI services.
  • B. Motivation and AI Architecture: The proposed four-level AI architecture assigns ALs 1 and 2 to service-oriented applications and ALs 3 and 4 to management-oriented applications.End-user-hosted and edge-hosted service-oriented applications occupy the lower levels.
  • B. Motivation and AI Architecture: End users can process AI tasks locally, with partial workloads offloaded to nearby edge servers when local resources are insufficient.Examples include next-word, traffic-demand, and vehicle-trajectory prediction.

C. AI for Networking

AI for networking applies AI techniques to network management, including network slicing, while connected AI coordinates models across control functions. The paper proposes intelligent modules that combine distributed neural sub-networks and conventional model-based techniques.

  • AI-Based Network Slicing: Network slicing separates long-timescale resource reservation from short-timescale resource scheduling for end users.Planning reserves resources for slices, while operation allocates reserved resources to users.
  • AI-Based Network Slicing: AI analyzes network data and dynamics to derive network slicing strategies for both planning and operation stages.
  • AI-Based Network Slicing: Local controllers schedule resources using centralized reservations and instantaneous user data from level-one digital twins.Relevant data include service type, user location, and user mobility.
  • AI-Based Network Slicing: Slice digital twins aggregate user-level data into service-level information that centralized controllers use for resource reservation.Proactive reservation must address time-varying traffic while avoiding over-provisioning or under-provisioning.
  • Connected AI Solution for Network Management: Connected AI links models across network control functions to capture their interplay and jointly make network control decisions.This balances training complexity and network performance, while addressing repeated processing of shared data.
  • Connected AI Solution for Network Management: The proposed intelligent modules may use a DNN alone or combine a DNN with conventional model-based techniques such as water-filling power allocation.DNN splitting and nested DNN techniques distribute connected sub-networks across network entities.
  • Connected AI Solution for Network Management: In mobile edge computing, SBS sub-networks make offloading decisions and connect to an MBS sub-network that produces service migration policies.The migration module receives SBS decisions and sub-network parameters, enabling cooperation across modules.

D. Networking for AI

Networking for AI designs and optimizes networks to support AI services throughout data collection, model training, and model inference. Its challenges include distributed data, constrained resources, and heterogeneous, dynamic network conditions.

  • Motivation: AI can operate as a network service, while networking for AI designs and optimizes networks to facilitate such services.AI services reside at levels 1 and 2 of the proposed AI architecture.
  • Motivation: Networking for AI requires interdisciplinary work to develop communication standards and technologies that support AI services at scale.The topic is receiving attention in both academia and industry.
  • Networking for AI Factors: Distributed data create spectrum scarcity and privacy concerns when traditional cloud-based collection and centralized training are used.Data are generated across end users and network-edge nodes.
  • Networking for AI Factors: Constrained network-node resources make deploying complex AI models costly because execution consumes computing resources and exchanging sub-model data consumes communication resources.Model partition can distribute complex models across multiple network nodes.
  • Networking for AI Factors: Network heterogeneity and dynamics complicate resource allocation for AI tasks across nodes with different communication, computing, and storage capabilities.Time-varying channels and spatiotemporal service demands further affect allocation.
  • AI-Service Lifecycle: Networking for AI spans data collection, model training, and model inference across the AI-service lifecycle.Communication links support data collection for training and subsequent AI-service operations.

2) State-of-the-Art Approaches:

State-of-the-art networking-for-AI research addresses data collection, distributed model training, and model inference. Representative approaches optimize communication, aggregation, resource allocation, data resolution, model choice, partitioning, and compression.

  • State-of-the-Art Approaches: Research on networking for AI remains at an early stage and is organized around data collection, model training, and model inference.
  • Data Collection: Data-collection methods increasingly account for model-training needs rather than maximizing only reliability or the amount of collected data.Data samples can have different importance levels for model training.
  • Data Collection: AI-centric data collection uses data-importance-aware resource allocation and protocols that incorporate sample importance into transmission decisions.Data uncertainty, including entropy, can represent data importance; retransmission protocols address poor channel conditions.
  • Model Training: Federated learning trains local models at end users and aggregates them at edge servers, supporting communication efficiency and user-data privacy.Growing model sizes make model uploading a strain on spectrum-constrained wireless networks.
  • Model Training: Federated-learning performance is optimized through framework design, spectrum-efficient aggregation, and resource management.Resource management includes end-user selection, local update counts, and model importance levels.
  • Model Inference: Model-inference approaches trade accuracy, latency, computation, energy, and communication through data-resolution, model-selection, model-partition, and model-compression decisions.Partitioned inference collaborates between end users and edge servers, while compressed models reduce local computation.

3) Research Challenges:

Networking for AI faces complex implementation choices and multi-dimensional QoS requirements. The paper proposes AI slices that separate training and inference into logically isolated subslices while coordinating them through shared resources.

  • Research Challenges: AI services can use different model structures, training procedures, and inference processes, each consuming different communication, computing, and storage resources.Implementation selection must suit service characteristics and network dynamics.
  • Research Challenges: AI-service QoS spans multiple dimensions, including model accuracy and service latency.Autonomous-driving examples require latency below 100 ms and ultra-high 3D-object-detection accuracy.
  • AI Slice: An AI slice contains a training subslice for model training and an inference subslice for model inference, with logically isolated resources.The two stages can have different goals, motivating their separation.
  • AI Slice: Training and inference subslices share a resource pool and coordinate to support the AI service while receiving separate QoS requirements.For object detection, example requirements are 99% detection accuracy and 100 ms service latency.
  • AI Slice: The training subslice configures data collection and model training and schedules resources to meet target accuracy, with retraining possible as data distributions change.
  • AI Slice: The inference subslice configures model inference and input compression according to service demand and deploys model variants based on base-station resources.Uncompressed models may serve resource-abundant BSs, while partitioned and pruned models may serve resource-limited BSs.
  • AI Slice: A vehicular video-analytics example uses federated learning in the training subslice and user-edge orchestration in the inference subslice.The example allocates camera computing and network spectrum for training and uses model partition or frame-rate reduction for inference.

E. Summary

The proposed 6G architecture combines holistic network virtualization with pervasive network intelligence. It integrates physical and cyber spaces, digital twins, network slices, connected AI, and AI slices to support flexible network management and AI services.

  • The architecture integrates digital twins and network slicing into holistic network virtualization, alongside connected AI and AI slices for pervasive network intelligence.
  • Its physical space contains end users and network infrastructure, while its cyber space contains network slices, digital twins, and connected AI.
  • Network slices support varied services, and digital twins represent end users and slices through data collected from the physical network.
  • AI manages network slices and digital twins through connected intelligent modules and also supports dedicated AI slices for AI services.
  • By combining virtualization and intelligence across networks and end users, the architecture aims to improve flexibility, scalability, adaptivity, and intelligence.
  • The architecture can apply to diverse physical networks, but implementations require customization for network-specific deployment, data flow, and digital-twin migration.

D. Implementation

A vehicular-network case study demonstrates implementation through network-slice establishment, digital-twin construction, and AI-module deployment for autonomous-driving services.

  • Multiple slices support autonomous-driving services with different QoS requirements, including conventional slices and AI slices with training and inference subslices.
  • Digital twins are constructed for vehicle users, roadside base stations, and network slices from extensive physical-entity data.
  • Vehicle-user and roadside-base-station twins reside at edge servers, while network-slice twins reside at a cloud server.
  • AI modules at centralized and local controllers provide intelligent network management, including planning and resource reservation decisions.

E. Interplay between Digital Twin Paradigm and Network Slicing

The architecture links digital twins with network slicing to combine user-centric and service-centric networking. It also enables hybrid data-model-driven methods that exploit the complementary strengths of both approaches.

  • Interplay between Digital Twin Paradigm and Network Slicing: Digital-twin virtualization focuses on end-user data management, whereas network slicing focuses on network management.
  • Interplay between Digital Twin Paradigm and Network Slicing: Digital twins enable user-centric networking, while isolated network slices enable service-centric networking through service-specific management.
  • Model-Driven and Data-Driven Methods: Model-driven methods are generally explainable and broadly applicable, but may become inaccurate or inapplicable in complex or rapidly changing networks.
  • Model-Driven and Data-Driven Methods: Data-driven methods can address problems too complicated for model-driven methods when data and stationarity are sufficient, but may perform questionably in non-stationary environments and generalize poorly.
  • Model-Driven and Data-Driven Methods: The proposed hybrid methods include backup or switching, task division, refinement, and mixing across network-management or AI-service functions.

G. Interplay between Virtualization and AI

The architecture couples virtualization and AI through data, control, and shared resources. It presents this coupling as part of a broader 6G architecture whose practical realization still has open issues.

  • Digital twins provide organized end-user and network-service data to intelligent modules and AI-slice training or inference subslices.
  • AI functions use digital-twin data to make network-management and service-provisioning decisions, which feed back to physical networks, slices, and digital-twin updates.
  • Virtualization and AI both require extensive computing and communication resources, creating resource-sharing considerations between their functions.
  • Digital-twin model control may reduce AI resource consumption by supplying only high-importance data, while creating a twin for every end user may be too resource-demanding.
  • Integrating digital twins with network slicing facilitates user-centric networking and improves the granularity of network management.
  • Hybrid data-model-driven methods are presented as potentially improving adaptivity and granularity, while AI-oriented network slicing can target QoS performance guarantees.
  • Practical implementation remains open around digital-twin performance characterization, model selection, migration, and data security.

B. Network Management Oriented Data Abstraction and Processing

The section identifies data abstraction, intelligent-module configuration, and training–inference coordination as central challenges for managing network data and AI services. It also highlights energy efficiency and unresolved questions in hybrid data-model driven methods.

  • Data abstraction: Different network management decisions require different data granularities: high granularity for operation strategies and low granularity for planning strategies.Selecting appropriate granularity and decision timescales remains an open issue.
  • Intelligent-module configuration: Connected AI modules must be configured across algorithms, input and output attributes, and inter-module connections while limiting communication and processing overhead.Finding settings that maximize data utilization remains challenging.
  • Training and inference coordination: AI-service performance depends on coordinating resource reservations between training and inference subslices sharing a virtualized resource pool.Overprovisioning training can increase accuracy but leave inference resources insufficient, whereas underprovisioning can reduce model and inference accuracy.
  • Energy efficiency: Large AI models consume substantial energy and can impose significant environmental and economic costs as accuracy improves.The paper notes that model compression and hybrid data-model driven methods may reduce data, computation, and energy requirements.
  • Hybrid data-model driven methods: Hybrid data-model driven methods remain exploratory, with open problems in option selection, switching, component interactions, additional designs, explainability, and robustness.The four proposed options are presented as initial ideas rather than a complete solution.
  • Conclusion: The proposed 6G architecture integrates holistic network virtualization and pervasive network intelligence through designs including digital-twin virtualization, connected AI, AI slices, and hybrid methods.The authors state that this architecture has potential for greater scalability and flexibility.
Loading 2301.00519v1…