Dynamic Edge Loading Balancing with Edge Node Activity Prediction and Accelerating the Model Convergence

Abstract

In mobile edge computing networks, achieving effective load balancing across edge server nodes is essential for minimizing task processing latency. However, the lack of a priori knowledge regarding the current load state of edge nodes for user devices presents a significant challenge in multi-user, multi-edge node scenarios. This challenge is exacerbated by the inherent dynamics and uncertainty of edge node load variations. To tackle these issues, we propose a deep reinforcement learning-based approach for task offloading and resource allocation, aiming to balance the load on edge nodes while reducing the long-term average cost. Specifically, we decompose the optimization problem into two subproblems, task offloading and resource allocation. The Karush–Kuhn–Tucker (KKT) conditions are employed to derive the optimal strategy for communication bandwidth and computational resource allocation for edge nodes. We utilize Long Short-Term Memory (LSTM) networks to forecast the real-time activity of edge nodes. Additionally, we integrate deep compression techniques to expedite model convergence, facilitating faster execution on user devices. Our simulation results demonstrate that our proposed scheme achieves a 47% reduction in terms of the task drop rate, a 14% decrease in the total system cost, and a 7.6% improvement in the runtime compared to the baseline schemes.

Keywords

• deep reinforcement learning
• load balancing
• mobile edge computing
• resource allocation
• task offloading