通用在网计算系统架构及协议设计

摘要：生成式人工智能大模型训练以及海量的大数据实时处理对任务处理性能提出了更高要求。在网计算在数据完成尽力而为转发的基础上，进一步将计算相关的操作卸载至转发节点，可以有效提升系统计算效率。针对当前在网计算系统设计碎片化问题开展研究，提出了满足多种在网计算场景的通用系统架构，并进行相关协议设计。通用在网计算架构兼顾了系统实现的灵活性以及应用开发友好性，为进一步推进在网计算的规模化应用提供了新思路。

关键词：在网计算；算力网络；网络架构；网络协议

Abstract: The large model training of generative artificial intelligence and the real-time processing of big data need higher requirements for task processing performance. On the basis of completing the best-effort forwarding of data, in-network computing (INC) further offloads computing-related operations to network forwarding nodes, which can effectively improve the system computing efficiency. Aiming at solving the problem of design fragmentation of INC systems, a generic system architecture is proposed to meet different requirements of various INC scenarios, and related protocols are designed. The generic INC architecture takes into account the flexibility of system implementation and the friendliness of application development, and puts forward a new idea for further improving the scalability of INC.

Keywords: in-network computing; computing power network; network architecture; network protocol