Joint optimization of task scheduling and computing resource allocation for VR video services in 5G‐advanced networks

Abstract

AbstractWith the development of 5G networks and related hardware equipment, the demand for various VR video services including panoramic video live broadcast and immersive game is increasing day by day. Due to the explosive growth of data volume and the trend toward lightweight VR equipment, traditional local computing architecture faces problems of insufficient computing power and high cost. As an emerging computing paradigm, the remote computing architecture of cloud‐edge collaboration frees users from cumbersome local restrictions. However, edge computing resources are often limited, and the remote cloud has inconsistent network latency. Therefore, how to adaptively schedule and allocate computing resources is the key to determining VR service quality. Based on an in‐depth analysis of the characteristics of VR video services and network environments, this article first builds a cloud‐edge‐network coorperation computing and transmission architecture for VR video services. This architecture can adaptively offload concurrent computing tasks and allocate computing power to them in VR scenes based on network conditions provided by the 5G‐Advanced networks via network information exposure interface. This paper proposes a general optimization problem to maximize user experience. And a single‐slot delay optimization model was constructed. Combining convex optimization theory and deep learning algorithms, this paper designs a task scheduling and computing power allocation strategy based on deep reinforcement learning. The policy gradually learns optimal task scheduling and resource allocation decisions to maximize delay‐related rewards in the long run. This article further introduces the graph neural network under the attention mechanism to solve the problem of changes in the number of concurrent tasks. Finally, through Cloudsim simulation, this article verifies the superiority of task scheduling and computing power allocation strategies based on deep reinforcement learning compared to other common strategies in terms of resource utilization, client latency, scalability and decision time.