The rapid proliferation of end devices and their hosted applications has significantly increased the demand for data processing. However, lightweight edge endpoints often suffer from limited computational capabilities due to power and battery constraints. Consequently, computation-intensive tasks are typically offloaded to cloud servers located in data centers for processing. While existing studies often abstract the edge–cloud connection as a direct link, the actual backbone network features complex topologies and constrained bandwidth, which intensify as workload scales up. To address this challenge, this paper investigates routing optimization within the backbone network under task offloading scenarios, and proposes a workload-aware deep reinforcement learning (DRL)-based routing algorithm for joint minimization of transmission and processing delays. Specifically, we first develop a detailed system model using discrete event simulation. Then, we design an integrated DRL-based decision-making scheme that simultaneously determines optimal routing paths and target data centers. Extensive experiments on diverse real-world network topologies and heterogeneous task workloads demonstrate that the proposed algorithm significantly outperforms conventional baselines in reducing both transmission and processing delays.

The Internet today hosts a multitude of communication sessions from diverse vertical industries, each with distinct and increasingly stringent quality-of-service (QoS) requirements across multiple performance metrics. However, QoS-aware routing remains a significant challenge in traffic engineering, as existing solutions struggle to adapt to dynamic network conditions and meet these rigorous QoS demands. To address this issue, this paper proposes a multi-agent graph reinforcement learning-based routing algorithm that provides differentiated treatment for multiple services. First, we explore both nodebased and link-based graph reinforcement learning paradigms for performance comparison. Second, two key mechanisms, i.e., a packet sacrifice mechanism and a Tchebycheff-based reward function, are designed to realize adaptive sacrifice behavior patterns, aiming to optimize the lower bound of service satisfaction rates and enhance fairness across services. Furthermore, to ensure practical applicability, we devise a distributed computing architecture featuring neighborhood-restricted data acquisition and asynchronous historical information retrieval. Extensive simulation results demonstrate that our proposed algorithms significantly outperform benchmark methods regarding the minimum service satisfaction rate, even under unseen networks. Besides, the distributed computing architecture is proven to incur no performance penalty, which can be generalized to other resource-constrained applications.