计算机科学
强化学习
云计算
延迟(音频)
分布式计算
虚拟化
骨干网
边缘设备
计算机网络
人工智能
操作系统
电信
作者
Jiayin Zhang,Huiqun Yu,Guisheng Fan,Zengpeng Li
标识
DOI:10.1016/j.comcom.2022.09.004
摘要
Virtualized Network Functions (VNF), Service Function Chains (SFC) and Network Functions Virtualization (NFV) architecture are promising basis of modern network infrastructures. How to make the best use of the limited resource at the edge yet achieve acceptable latency performance is one of the key challenges. Further, cloud-native network functions (CNF) enables a more flexible architecture with container-based virtualization, yet brings the problem of cold-start handling, since transferring and booting a container image can bring an indispensable latency. We formulate the cold-start aware cloud-native SFC caching problem as a mathematical optimization problem with a set of constraints based on the resource limitation and performance requirement. To efficiently handle this problem, which has been proved to be NP-Hard, we design a deep reinforcement learning (DRL) approach, along with two graph neural network-based embedding networks for the extraction of backbone network graph and caching request information, respectively. The resulting DRL agent is able to learn caching decisions, aiming at optimizing the processing latency, sub-frame processing latency, and launch latency performance while maintaining the request acceptance ratio. Extensive simulations conducted on multiple backbone network structures and various request load suggest that the proposed approach outperforms the state-of-the-art solutions in request acceptance ratio, latency performance under high loads, and cold-start handling with little extra execution time overhead.
科研通智能强力驱动
Strongly Powered by AbleSci AI