黄光平 史伟强 谭斌

摘要：提出一种以IP网络为中心的算力网络架构，即在网络域创建云池算力资源和服务的状态，从而实现网络层的算力编排和调度。算网一体编排和路由，是该算力网络架构的核心特征。针对算力网络中的服务多实例应用场景，所提架构方案对SRv6或基于SRv6的业务功能链（SFC）做功能增强和扩展，以满足单服务对应动态多实例的算力路由需求。控制面架构方案采取一种分级分层状态表的维护机制，将不同颗粒度的算力资源和服务状态在不同的网络域做同步通告，并创建对应的分级路由表，从而压缩节点的状态表和边界网关协议（BGP）的通告频率。转发面则执行算力服务标识语义封装，承载网骨干节点仍然保持无状态转发。

关键词：算力网络；SRv6；算力状态；分级路由

Abstract： An IP network-based architecture of computing power network is proposed， which creates the state of cloud pool computing power resources and services in the network domain to realize the computing power arrangement and scheduling of the network layer. Integrated computing network arrangement and routing are the core features of the computing power network architecture. For the service multi-instance application scenario in the computing power network， the proposed architecture scheme enhances and extends SRv6 or SRv6-based service function chaining （SFC） to support the single service routing requirements for dynamic multi-instances. The control surface architecture scheme adopts a maintenance mechanism of hierarchical state tables， which synchronously notifies the computing power resources and service states of different granularity in different network domains， and creates the corresponding hierarchical routing table， to compress the state table of the node and the notification frequency of the border gateway protocol （BGP）. Accordingly， a dual-semantic encapsulation with IP topology and computing service identification in the forwarding plane would also be proposed， while the backbone network nodes would remain unaware of computing power metrics.

Keywords： computing power network； SRv6； computing status； classified routing

在互联网协议（IP）承载网络域，通过精细化动态感知，网络控制器或网络节点可以创建基于多云池内算力资源及服务状态的算力路由表，并据此进行算力资源和服务的编排调度。这是以网络为基础平台的算力网络架构的核心要素。也就是说，在IP拓扑路由的基础上，新增算力资源和服务路由，使路由策略约束机制由当前的IP拓扑单约束演变为IP拓扑和算力双约束。这给网元控制面、转发面和管理面均带来新的挑战，也是算力网络为IP网络引入的全新议题。

当前主流的云侧应用级跨云池计算资源调度系统，如内容分发网络（CDN）、AWS（亚马逊公司的云计算服务）等，均与特定应用或应用集群硬绑定。……