A Computing and Network Resource Scheduling Approach in Compute First Networking Based on Deep Reinforcement Learning

According to the supply capacity of computing resource providers and the dynamic resource requirements of application demanders, matching and scheduling the multidimensional resources such as computing, storage, and network at the bottom of the computing infrastructure are key problems in compute-first networking applications. In this paper, we propose a two-level model for computing-networking fusion scheduling with a centralized scheme to achieve computational offloading. We utilize a Gate Recurrent Unit (GRU) network to capture the predicted network link state information for the purpose of network link pre-screening. We present four scheduling strategies by considering energy consumption, device redundancy, price, and service quality in the optimization goal. The computing power scheduling is formulated as a Markov Decision Process problem, and a computational task offloading algorithm is proposed based on the Deep Deterministic Policy Gradient (DDPG). The results suggest that the proposed algorithm has advantages in terms of time required for computational task offloading, resource utilization, and reward value compared to several baseline algorithms, demonstrating the effectiveness of the algorithm.