【摘要】 随着数字经济战略和数据要素改革的深化，计算需求持续爆发式增长，由此引发的能源消耗急剧增加，为“碳中和”“碳达峰”带来巨大挑战。“东数西算”通过对跨地理空间的计算网络资源进行协同调度，解决计算需求与资源供给不平衡问题，对网络带来了前所未有的挑战和演进动力。文章分析检索查询类、渲染交互类、深度学习类和区块共识类等四种算力需求场景，归纳总结算力网络场景特性及其路由资源需求模型，借鉴工业界和学术界在跨地理空间的计算资源协同调度方案，遵循以网络为中心且对现有网络协议改动最小的原则，探讨应用层算力任务预配分解，网络层路由优化和报文复制消除相结合的算网融合统一调度机制，并设计算力资源发布、算力任务预配和算网融合调度的协议流程，以平衡算力网络负载与算力服务质量。更多还原

【Abstract】 Computing network resources across geospatial, which has brought unprecedented challenges and evolution power to the network. This paper analyzes four kinds of computing power demand scenarios, such as retrieval and query, rendering interaction, deep learning, and block consensus, summarizes the characteristics of computing power network scenarios and their routing resource demand models, draws lessons from the collaborative adjustment scheme of computing resources across geospatial in industry and academia. The paper also discusses the pre-allocation and decomposition of computation tasks at the application layer based on taking the network as the center and minimizing the changes to the existing network protocols.The unified scheduling mechanism of computing network integration is designed based on network layer routing optimization and package replication and elimination. The protocol processes of computing resource release, task pre-allocation, and computing network integration scheduling are designed to balance the computing network load and service quality.更多还原