Incremental DRL-Based Resource Management for Dynamic Network Slicing in an Urban-Wide Testbed

TL;DR

This work addresses resource management for dynamic network slicing in multi-access edge computing by modeling the problem as an evolving MDP and solving it with an incremental cooperative MADDPG framework. Each network slice is assigned a DRL agent, sharing a common reward to capture inter-slice competition, while an incremental learning mechanism preserves learned policies when slice sets change, reducing re-training costs. The approach is validated on a city-scale OpenStack–Kubernetes testbed, demonstrating superior latency and energy performance over baselines and substantial training-energy reductions (up to ~50%) compared to retraining from scratch. The results indicate strong practical potential for real-time, energy-aware orchestration of edge resources under dynamic slicing conditions, with implications for scalable, SLA-compliant MEC deployments.

Abstract

Multi-access edge computing provides localized resources within mobile networks to address the requirements of emerging latency-sensitive and computing-intensive applications. At the edge, dynamic requests necessitate sophisticated resource management for adaptive network slicing. This involves optimizing resource allocations, scaling functions, and load balancing to utilize only essential resources under constrained network scenarios. However, existing solutions largely assume static slice counts, ignoring the re-optimization overhead associated with management algorithms when slices fluctuate. Moreover, many approaches rely on simplified energy models that overlook intertemporal resource scheduling and are predominantly evaluated through simulations, neglecting critical practical considerations. This paper presents an incremental cooperative Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm for resource management in dynamic edge slicing. The proposed approach optimizes long-term slicing benefits by reducing delay and energy consumption while minimizing retraining overhead in response to slice variations. Furthermore, we implement an urban-wide edge computing testbed based on OpenStack and Kubernetes to validate the algorithm's performance. Experimental results demonstrate that our incremental MADDPG method outperforms benchmark strategies in aggregated slicing utility and reduces training energy consumption by up to 50% compared to the re-optimization approach.

Figures (10)

• Figure 1: Network slicing in MEC networks with constrained computing and bandwidth resources.
• Figure 2: Dynamic power consumption of a MEC server caused by the resource allocation of previous and future requests. The top left corner showcases an example of energy consumption by a task, highlighting its co-dependence on past or future issued tasks besides its own resource allocation.
• Figure 3: The power consumption and temperature of a Dell PowerEdge R760 bare-metal MEC server vs. CPU load percentages.
• Figure 4:
• Figure 5: Architecture overview of the proposed MADDPG approach.