CaRCS: Joint Optimization of Computing-aware Routing and Collaborative Scheduling in Computing Power Networks

With the rapid advancement of technology, intelligent applications increasingly require higher computing power, driving the evolution of the end-edge-cloud Collaborative Scheduling (CS) paradigm. However, this paradigm primarily focuses on vertical collaboration, neglecting horizontal coordination among ubiquitous heterogeneous computing resources, leading to imbalanced resource utilization. Fortunately, the Computing Power Network (CPN) has been introduced to facilitate ubiquitous resource coordination through pervasive networks. The core issue in CPN is the CS of network-wide computing tasks, yet most existing research has largely concentrated on offloading decisions while critical aspects of network routing remain underexplored. Therefore, this paper studies the optimization of joint routing and scheduling within large-scale CPNs. Initially, we develop a CS framework that accommodates the return of task computing results. We then incorporate computing information into the routing domain and introduce a CS mechanism based on Compute-aware Routing (CaR), which narrows the scope of routing exchanges and collaborative scheduling. Ultimately, we present a Deep Reinforcement Learning (DRL)-based algorithm to optimize the long-term CS problems based on balanced CaR, enhancing the success rate of computing tasks and ensuring balanced utilization of computing and network resources. The effectiveness of our proposed mechanism and algorithm is confirmed through simulation experiments.