Efficient task migration and resource allocation in cloud–edge collaboration: A DRL approach with learnable masking

Abstract

The paper addresses the challenges of task migration and resource allocation in heterogeneous cloud–edge environments, where dynamic and stochastic conditions complicate efficient scheduling. To tackle this, the authors propose a novel scheduling algorithm combining soft actor–critic (SAC) agent with masked layer and graph convolutional network (GCN), namely MGSAC algorithm. MGSAC utilizes GCN to extract hidden structural features from the environment, enabling better adaptation to dynamic changes. Additionally, a learnable mask layer filters out ineffective actions, refining the selection of scheduling strategies and improving overall performance. By evaluating MGSAC on the real-world Bit-Brain dataset and simulating it using Cloud-Sim, experimental results demonstrate its superiority over existing algorithms in energy consumption, task response time, task migration time, and task Service-Level-Agreement violations rate, showcasing its effectiveness in real-world scenarios.

Keywords

• Cloud–edge computing
• Deep reinforcement learning
• Resource allocation
• Task migration