Cost-Aware Collaborative Task Offloading in Computation Reuse-Enabled Industrial Edge Networks

Abstract

Collaborative edge computing (CEC) enables edge servers to cooperatively process computation-intensive tasks from industrial Internet of Things devices, while computation reuse (CR) leverages cached task results to reduce response delay and energy consumption for similar tasks. However, existing studies fail to adequately consider both the time validity of cached results and retrieval costs, leading to suboptimal system performance. This paper investigates the joint optimization of task offloading, computing resource allocation, result retrieval, and result caching (T3R) in CR-enabled CEC networks with stochastic task arrivals. The problem presents unique challenges due to the complex interplay between early-stage caching decisions and subsequent retrieval choices. We formulate the problem as a long-term cost minimization problem incorporating result validity periods and retrieval costs under computing and cache capacity constraints. Our solution first employs a Lyapunov-based method to transform the problem into a real-time stochastic optimization problem. To tackle the graph structure of the network and the challenge of a large action space, we employ a graph convolutional network (GCN)-enhanced deep reinforcement learning (DRL) algorithm. Moreover, we maintain task arrival statistics to enable the agent to capture the potential relationship between result caching and retrieval strategies. Extensive experiments demonstrate significant improvements in system cost reduction compared to baseline methods.