Traffic-cognitive Slicing for Resource-efficient Offloading with Dual-distillation DRL in Multi-edge Systems

Abstract

In edge computing, emerging network slicing and computation offloading can support Edge Service Providers (ESPs) better handling diverse distributions of user requests, to improve Quality-of-Service (QoS) and resource efficiency. However, fluctuating traffic and heterogeneous resources seriously hinder their broader application in multi-edge systems. Existing solutions commonly rely on static configurations or prior knowledge, lacking adaptability to changeable multi-edge environments and thus causing unsatisfying QoS and improper resource provisioning. To address this important challenge, we propose SliceOff, a novel resource-efficient offloading framework with traffic-cognitive network slicing for dynamic multi-edge systems. First, we design a new traffic prediction model based on self-attention to capture traffic fluctuations among different edge regions. Next, an adaptive slicing strategy based on random rounding is devised to adjust the resource configuration according to the traffic and demands of edge regions. Finally, we develop an improved Deep Reinforcement Learning (DRL) method with a dual-distillation mechanism to address the complex offloading problem, where twin critics networks and dual policy distillation are integrated to improve the agents exploration and updating efficiency in huge decision spaces. Notably, we carry out rigorous theoretical analysis to prove the effectiveness of the proposed SliceOff. Using the real-world testbed and datasets of user traffic, extensive experiments are conducted to verify the superiority of the proposed SliceOff. The results show that the SliceOff improves resource inefficiency and ESP profits under dynamic multi-edge environments, which outperforms state-of-the-art methods on multiple metrics under various scenarios.