Source-linked AI summary
AI-Native Network Slicing for 6G Networks
Wen Wu, Conghao Zhou, Mushu Li, Huaqing Wu, Haibo Zhou, Ning Zhang, Xuemin, Shen, Weihua Zhuang
TL;DR
6G requires slicing that can handle integrated networks, diverse stringent-QoS services, and emerging AI workloads. This paper proposes an AI-native architecture combining AI for slice management with slicing for AI services, and its case study reports about 15% lower one-day cumulative system cost than a myopic solution. It also identifies prediction errors as an open research issue for prediction-enabled slicing.
Problem
6G's heterogeneous infrastructure, stringent service requirements, and emerging AI services create challenging slice-management and QoS-support requirements.
Method
The paper proposes an AI-native architecture using AI for slicing and customized slices, AI instances, and resource management for AI services.
Results
Around 15% lower cumulative overall system cost within one day is achieved by the DDPG-based solution versus the myopic solution.
Takeaways & Limitations
AI for slicing can reduce network-management complexity and adapt to dynamic environments, while slicing for AI accommodates emerging AI services.
Abstract
from arXiv · showhide
With the global roll-out of the fifth generation (5G) networks, it is necessary to look beyond 5G and envision the 6G networks. The 6G networks are expected to have space-air-ground integrated networks, advanced network virtualization, and ubiquitous intelligence. This article presents an artificial intelligence (AI)-native network slicing architecture for 6G networks to enable the synergy of AI and network slicing, thereby facilitating intelligent network management and supporting emerging AI services. AI-based solutions are first discussed across network slicing lifecycle to intelligently manage network slices, i.e., AI for slicing. Then, network slicing solutions are studied to support emerging AI services by constructing AI instances and performing efficient resource management, i.e., slicing for AI. Finally, a case study is presented, followed by a discussion of open research issues that are essential for AI-native network slicing in 6G networks.
I. INTRODUCTION
6G networks introduce integrated space-air-ground infrastructure, diverse stringent-QoS services, and ubiquitous intelligence, making intelligent slice management necessary. The article proposes AI-native slicing to support both network management and emerging AI services.
- 6G combines space, air, and ground networks to provide global coverage and on-demand services.
- Network slicing creates logically isolated virtual networks for diverse services and manages them across preparation, planning, and operation.
- SAGIN coordination, stringent QoS requirements, and emerging services complicate slice management in 6G networks.
- The proposed architecture applies AI to manage slices and uses slicing to support customized AI services.
II. AI-NATIVE NETWORK SLICING FOR 6G NETWORKS
6G network slicing extends 5G virtualization and SDN to support larger performance demands, integrated coverage, diversified services, and intelligence distributed toward the network edge and users.
- Network Slicing: Network slicing creates multiple logically isolated virtual networks over shared infrastructure for flexible and adaptive management.
- 6G Features: 6G is expected to require 1 Tbps peak data rate, 0.1 ms end-to-end latency, 10 million devices/km2, and near 100% coverage.
- 6G Features: SAGIN integrates space, air, and ground segments to provide global coverage and support on-demand, high-rate, low-delay services.
- 6G Features: 6G services impose different stringent QoS requirements, including high data rates for mobile VR and ultra-high reliability for autonomous driving.
- 6G Features: Ubiquitous intelligence pushes AI, caching, and computing capabilities toward the network edge and end users.
C. AI-Native Network Slicing Architecture
The AI-native architecture integrates SAGIN and ubiquitous intelligence while using AI to manage slices and network slicing to support emerging AI services.
- C. AI-Native Network Slicing Architecture: The architecture addresses SAGIN scale, dynamic resources, diversified services, and stringent QoS requirements through AI-native slicing.
- C. AI-Native Network Slicing Architecture: AI for slicing integrates AI into SDN controllers to manage stringent-QoS slices efficiently and cost-effectively.
- C. AI-Native Network Slicing Architecture: Slicing for AI constructs customized slices for emerging AI services over common physical infrastructure.
- C. AI-Native Network Slicing Architecture: The architecture deploys centralized cloud SDN control for slice management and local access-point control for end-user resource scheduling.
- C. AI-Native Network Slicing Architecture: The article examines AI solutions across preparation, planning, and operation, including information exchange among network entities.
A. Network Slicing Lifecycle
The network slicing lifecycle has preparation, planning, and operation phases, with controllers constructing slices from service, traffic, user, and virtual-resource information.
- A. Network Slicing Lifecycle: The lifecycle consists of preparation, planning, and operation, with centralized control handling preparation and planning and local control coordinating operation.
- Preparation: Preparation constructs and configures slices using service requirements, traffic, user information, and virtual network resource availability.
- Preparation: Service requirement extraction classifies services using QoS attributes such as delay, priority, throughput, and reliability.
- Preparation: Virtualization pools communication, computing, and caching resources and separates network functions into virtualized network functions.
- Preparation: After these tasks, the SDN controller constructs network slices for each admitted slice request.
2) Planning Phase:
The planning phase reserves virtualized network resources for slices over longer planning windows, using service and network information plus feedback to adjust reservations.
- Planning Phase: Planning reserves network resources for slices during windows lasting several minutes to several hours.Window duration depends on service demand and network dynamics.
- Lifecycle Context: The network slicing lifecycle separates centralized preparation and planning from locally coordinated operation.
- Planning Phase: The SDN controller collects service demands, channel conditions, and user mobility patterns for resource-reservation decisions.
- Planning Phase: Reserved virtualized resources are mapped to the physical network and adjusted according to monitored slice performance.
B. Roles of AI in Network Slicing
AI-based network slicing addresses the management cost of numerous slices by applying distinct AI tasks across preparation, planning, and operation phases.
- Overview: AI-based network slicing is presented as a potential solution to the significant management cost of operating many slices in 6G networks.
- AI for Preparation: During preparation, AI predicts service demand and supports slice admission for network-resource utilization.Predictions based on historical data can inform planning decisions.
- AI for Planning: During planning, AI supports dynamic VNF placement and resource reservation to address time-varying service demands.Deep learning methods are proposed to enhance resource utilization in dynamic network environments.
- AI for Operation: During operation, reinforcement learning can support dynamic resource orchestration, while AI selects radio access technologies for users.
C. Procedure of Information Exchange
AI for slicing exchanges information among end users, access points, and the SDN controller, while slicing for AI supports AI services through implementation selection and resource management.
- Information Exchange: Access points collect user-level demand, mobility, and channel information, then translate it into service-level information for the SDN controller.AI can support data abstraction, fusion, and analysis during this translation.
- Information Exchange: The SDN controller uses service-level information for AI-based planning, sends decisions to access points, and receives their enforcement.
- Information Exchange: End users report real-time service information, enabling access points to run AI-based operation algorithms for resource allocation.
- Slicing for AI: Slicing for AI supports AI services while satisfying QoS requirements through AI-instance construction, selection, and resource management.
- AI Instance: An AI service can have multiple implementation options, making selection of an appropriate option a primary support issue.Options may differ in algorithm, training manner, and network-resource allocation.
- AI Instance: AI-instance management constructs candidate implementations from available resources and service requirements before selecting one for the service.
B. Resource Management in AI Service Lifecycle
AI service performance depends on data collection, model training, and model inference, so network resources should be allocated jointly across the lifecycle; a case study targets long-term system cost.
- AI Service Lifecycle: AI services proceed through data collection, model training, and model inference stages.Training may be centralized or distributed, while inference may use device-edge collaboration.
- AI Service Lifecycle: AI service performance depends on all three lifecycle stages, with inference accuracy affected by data quality, training iterations, and inference approach.
- Resource Management: The three stages consume multidimensional network resources, motivating joint allocation to optimize AI-service performance.
- Case Study: The case study examines AI-assisted resource reservation with the aim of reducing long-term overall system cost.
A. Considered Scenario
The scenario models resource reservation for autonomous-driving services in an air-ground integrated network, minimizing a weighted system cost over planning windows with DDPG.
- A. Considered Scenario: The air-ground scenario supports autonomous driving on a 2 km highway with two BSs and one UAV.The BSs are separated by 1 km, while the UAV hovers centrally at 100 meters.
- A. Considered Scenario: Resource reservation decisions minimize overall system cost while accounting for vehicle traffic dynamics.
- A. Considered Scenario: The overall system cost is a weighted sum of resource reservation, slice reconfiguration, and delay violation costs across planning windows.Resource reservation cost covers reserved spectrum and computing resources; reconfiguration cost reflects consecutive-decision differences, while delay violation incurs a penalty.
- A. Considered Scenario: The model allocates spectrum in 5 MHz subcarrier units and computing resources in VM instances processing 10 × 10^9 cycles/sec.
- A. Considered Scenario: The cost weights are ωr = 1, ωs = 20, and ωd = 200, with one-hour planning windows.
- A. Considered Scenario: A DDPG-based solution minimizes overall system cost using fully connected actor and critic networks, and is compared with myopic resource reservation.Both networks have four layers, hidden-layer sizes of 128 and 64, and learning rates of 2 × 10^-4 and 2 × 10^-3.
B. Simulation Results
Using real-world highway traffic traces, the DDPG resource reservation solution converges after 4,000 training episodes and lowers one-day cumulative system cost by around 15% versus the myopic solution.
- B. Simulation Results: The DDPG-based resource reservation solution converges after 4,000 training episodes.Raw simulation points are processed with a five-point moving average to highlight the convergence trend.
- B. Simulation Results: The DDPG-based solution achieves low system cost in the evaluated resource reservation scenario.
B. Data Management Framework
The paper identifies distributed data collection and analysis as necessary for AI-native slicing, while emphasizing that imperfect traffic predictions can degrade slicing performance.
- B. Data Management Framework: AI-native network slicing depends on abundant distributed data for training AI models.
- B. Data Management Framework: A data management framework is needed to collect and analyze distributed data despite non-negligible communication costs.Historical user behavior data can be analyzed to predict spatio-temporal service demand distributions.
- C. Prediction-Empowered Network Slicing: Imperfect traffic prediction can degrade network slicing performance, motivating evaluation of prediction errors and corresponding mitigation solutions.
- C. Prediction-Empowered Network Slicing: The proposed architecture combines AI for slicing to reduce management complexity with slicing for AI to accommodate emerging AI services.