基于虚拟网络嵌入的异构算力网络资源管理

Heterogeneous computing network resource management based on virtual network embedding

随着人工智能大模型引领的新应用和新需求的蓬勃兴起,算力规模和计算技术正经历着前所未有的快速演进与多元化创新。然而,在算力网络呈现集群化和异构化趋势的同时,算力需求迅猛增长与资源利用低效性之间的矛盾日益凸显。如何实现对异构算力的统一高效管控以提升资源利用率,已成为当前研究的重要课题。基于网络虚拟化(network virtualization,NV)技术,提出了一种基于虚拟网络嵌入(virtual network embedding,VNE)异构跨域算力资源分配方法。具体而言,构建了一个基于深度强化学习(deep reinforcement learning,DRL)模型的策略网络,负责精准计算候选算网节点与链路,以实现资源优化分配。通过一系列仿真实验,验证了该方法的有效性,为解决异构算力管理问题提供了新的思路和方法。

With the booming development of new applications and demands led by large AI models,computing scale and technology are experiencing unprecedented rapid evolution and diversified innovation.However,as computing networks show a trend of clustering and heterogeneity,the contradiction between the rapid growth of computing demand and the inefficiency of resource utilization has become increasingly prominent.How to achieve unified and efficient management of heterogeneous computing to improve resource utilization has become an important research topic.Based on network virtualization(NV)technology,a heterogeneous cross-domain computing resource allocation method based on virtual network embedding(VNE)was proposed.Specifically,a policy network based on a deep reinforcement learning(DRL)model was constructed to accurately calculate candidate computing nodes and links for optimal resource allocation.Through a series of simulation experiments,it verifies the effectiveness of this method and provids new ideas for solving the problem of heterogeneous computing management.