Abstract: Computing power network(CPN) is a new network system that solves the contradiction between computing power supply and demand, computing power network transmission problems, and computing power resource universalization problems. According to the supply capacity of computing power resource providers and the dynamic resource requirements of application demanders, the computing, storage, network and other multi-dimensional resources of the underlying computing power infrastructure in the region are integrated to provide users with personalized computing power resource services and realize efficient management and on-demand allocation of computing power resources. In order to improve the utilization and reliability of computing power network resource matching and scheduling, this paper proposes a highly reliable matching method (resource measurement and rescheduling matching method, RMRMM) based on multi-dimensional resource measurement and rescheduling. In order to achieve high-utilization resource scheduling, RMRMM designs a resource measurement matching scheme based on entropy weight TOPSIS and deep reinforcement learning, comprehensively measures the structural eigenvalues, computing power, storage capacity, and network communication capacity of nodes, and narrows the resource matching range to improve the matching accuracy and resource utilization; RMRMM considers the failure of nodes due to attacks, and designs a rescheduling module based on large neighborhood adaptive search. When the matching result fails, nodes and tasks are rescheduled to improve the task reception rate and enhance the overall reliability. In order to evaluate the proposed scheme, this paper conducted simulation experiments on the OMNet++ platform. The experimental results show that the resource utilization and request reception rate of the RMRMM scheme are better than other matching strategies, proving that the RMRMM scheme is more efficient and reliable.

Key words: Computing power network, resource metrics, rescheduling, high reliability, deep reinforcement learning

摘要：  算力网络(CPN)是一种解决算力供需矛盾、网络传输问题以及算力资源普惠问题的新型网络体系。根据算力资源提供方的供给能力和应用需求方的动态资源需求，对区域内算力基础设施底层的计算、存储、网络等多维资源进行整合，为用户提供个性化的算力资源服务，实现算力资源的高效管理和按需分配。为了提高资源匹配调度的利用率和可靠性，本文提出一种基于多维资源度量和重调度的高可靠匹配方法(RMRMM)。为了实现高利用率的资源调度，RMRMM设计了基于熵权TOPSIS和深度强化学习(DRL)的资源度量匹配方案，对节点的结构特征值(SFV)、计算能力、存储能力、网络通信能力进行综合度量，缩小资源匹配范围以提高匹配的精准性和资源的利用率；同时RMRMM考虑节点遭受攻击失效的情况，设计基于自适应大邻域搜索(ALNS)算法的重调度模块，在匹配结果失效时进行节点与任务的重新调度，提高任务的接收率以增强整体的可靠性。为了评估所提出的方案，本文在OMNet++平台上进行了仿真实验，实验结果表明，RMRMM方案的资源利用率和请求接收率均优于其他匹配策略，证明RMRMM方案更加高效可靠。

关键词: 算力网络, 资源度量, 重调度, 高可靠, 深度强化学习