摘要
随着生成式人工智能(artificial intelligence generated content,AIGC)大模型和智能计算应用的高速发展,面向广域远程直接存储器访问(remote direct memory access,RDMA)的确定性网络技术成为近期智算中心互联的研究热点。首先分析了广域RDMA技术发展路线、实现方案和技术挑战,然后研究广域RDMA部署时面临的各类网络问题,通过搭建实验室和外场测试环境开展了一系列测试验证,分析提出了广域RDMA的确定性需求特性和边界性能指标,最后对面向广域RDMA的确定性网络技术进行了总结和展望。
With the rapid development of AIGC large models and intelligent computing applications,deterministic networking technology for wide-area RDMA has become a recent research hotspot.The development path,implementation solutions and technical challenges of RDMA-WAN were analyzed,by setting up a laboratory and an outfield testing environment,with a series of testing and verifications.The deterministic demand characteristics and boundary performance indicators of RDMA-WAN were proposed.Finally,the deterministic networking technologies for RDMA-WAN were summarized and prospected.
作者
赵俊峰
李芳
叶晓峰
江凇
ZHAO Junfeng;LI Fang;YE Xiaofeng;JIANG Song(China Academy of Information and Communication Technology,Beijing 100191,China;Jiangsu Viscore Technologies Co.,Ltd.,Suzhou 215000,China;State Grid Jiangsu Electric Power Co.,Ltd.,Information and Telecommunication Branch,Nanjing 210024,China)
出处
《电信科学》
2023年第11期39-51,共13页
Telecommunications Science
基金
国家电网公司总部科技项目(No.5700-202318269A-1-1-ZN)。
关键词
广域RDMA
确定性网络
边界性能指标
RDMA-WAN
deterministic networking
boundary performance indicator
