摘要
AI大模型训练、高性能存储等业务应用场景提出了海量规模的计算需求,与传统数据中心业务相比,在流量模型和网络需求方面有着显著区别,驱使传统的数据中心网络向智算中心和无损网络转型。从智算中心和无损网络的发展背景入手,分析了当前智算中心网络存在的问题,探索了智算中心网络流量调度的关键技术,并进行了流量调度平台的研发实践,为智算中心网络发展和应用提供思路。
Scenarios such as AI large-scale model training and high-performance storage have proposed massive computing demands.Compared with traditional data center services,there are significant differences in traffic models and network requirements,driving traditional data center networks to transform towards intelligent computing centers and lossless network.From the development of intelligent computing centers and lossless network,it analyzes the problems existing in the current intelligent computing center,explores the key technologies of network traffic scheduling in the intelligent computing centers,and conducts research and practice on the traffic scheduling platform,which provides ideas for the development and application of intelligent computing centers.
作者
韩博文
徐博华
曹畅
刘千仞
Han Bowen;Xu Bohua;Cao Chang;Liu Qianren(China Unicom Research Institute,Beijing 100048,China;China United Net-work Communications Group Co.,Ltd.,Beijing 100033,China)
出处
《邮电设计技术》
2024年第4期12-19,共8页
Designing Techniques of Posts and Telecommunications
关键词
智算中心
无损网络
大模型训练
Intelligent computing centers
Lossless network
AI large-scale model training