Optimal Latency and Energy-Aware Task Scheduling in In-Network Computing Paradigm: A Deep Reinforcement Learning Approach

To support the escalating traffic demands in the 6 G era, the novel computing paradigm of in-network computing (INC), where tasks can be processed on the forwarding path, is emerging with enhanced network performance and improved service quality. Considering the large network scale with high dynamics, effective task scheduling in INC paradigm becomes imperative but challenging. In this work, we investigate the task scheduling in INC paradigm to minimize both the task delay and network energy consumption, while considering constraints on task latency, traffic dynamics, and available communication and computing resources. We first construct a novel computing and communication model considering the traffic variation in network nodes on the transmission path. To solve the task scheduling problem, we propose an algorithm, named as NBFNDRL, which is a deep reinforcement learning (DRL) algorithm based on Neural Bellman Ford networks (NBFNet). NBFNet can learn high-dimensional correlated graph structural information, utilize message-passing mechanisms to represent changes in traffic between adjacent nodes, predict scheduling paths, and provide a basis for DRL decision-making. The DRL agent trains and updates NBFNet through interaction with the environment. Finally, we present simulation results to demonstrate the effectiveness of our proposed approach in comparison to benchmark algorithms and various computing paradigms.