Overselling-Aware Task Scheduling in MEC With Deep Reinforcement Learning

Abstract

Multi-access edge computing (MEC) enables low-latency computing services, yet third-party resource providers integrated into MEC systems may oversell their computing capacity, severely degrading task scheduling performance. To address this, we propose an overselling-aware task scheduling scheme with deep reinforcement learning (DRL) for delay-sensitive tasks in overselling environments, aiming to maximize system profit and improve on-time task completion. The proposed scheme incorporates an Overselling Awareness capability that enables the scheduler to be aware of servers’ overselling behaviors and estimate their actual computing power, thereby mitigating overselling effects on scheduling decisions. Furthermore, to handle dynamic variations in task and server quantities, we design a novel policy network architecture, Chimera, for the DRL agent in place of conventional policy networks. Chimera integrates weight sharing, pooling, and attention mechanisms to produce scheduling decisions under variable input lengths without altering its parameter size or network structure. Experimental results demonstrate that the proposed scheme achieves approximately 58.14-60.72% higher system profits and 74.30-84.70% better on-time completion rates compared to baseline methods in overselling environments.