Dependency-aware computation-offloading and cooperative service-caching are pivotal technologies in mobile edge computing (MEC) systems. However, existing works often overlook the interdependencies among tasks/subtasks and the limited caching capacity of MEC servers, leading to suboptimal energy-time cost (ETC). In this paper, we investigate the joint optimization of computation-offloading, service-caching, and resource-allocation in an environment with multiple MEC servers, where multiple Internet of things devices (IoTDs) generate tasks composed of interdependent subtasks which are modeled as directed acyclic graphs (DAGs). Considering the constraints of limited caching capacity and computation resources at MEC servers, we formulate a mixed-integer nonlinear programming (MINLP) problem aimed at minimizing the weighted sum of task completion delay and energy consumption across all IoTDs, i.e., ETC. To solve this challenging problem, we design a scheme by integrating graph convolutional networks (GCN) with deep reinforcement learning (DRL), in which GCN extracts subtask features and dependencies among subtasks, and combines them with the features of the system environment as the state of our DRL method. We then employ an advantage actor-critic (A2C) framework with hybrid action space handling capability to perform Markov decisions and obtain optimal caching, offloading, and resource-allocation actions. Third, we conduct extensive experiments to evaluate the effectiveness of our proposed scheme. The simulations demonstrate that, compared with state-of-the-art baselines, our scheme achieves the lowest ETC in most cases, converges faster, and maintains robustness under diverse experimental conditions, including varying transmission rates, numbers of IoTDs, subtask counts, and caching capacity. Moreover, by conducting comparative analyses with relevant literature, we further confirm that our joint optimization framework has significant advantages in balancing energy efficiency and latency in the MEC environment.
Wang et al. (Tue,) studied this question.