面向算力网络的端网协同RDMA拥塞控制OA
Congestion control for RDMA with end-network collaboration in computing power network
为解决远程直接内存访问(RDMA)技术跨域互联场景下的长控制回路及混合流量拥塞问题,提出了一种面向算力网络的拥塞控制方法WRCC.采用基于输入速率的公平速率计算策略,由交换机精确计算拥塞队列的端口公平速率.结合近源交换机双控制回路与带内网络遥测技术,实现端网协同的速率控制,快速响应拥塞.仿真实验表明,与现有商用方法相比,WRCC能将平均流完成时间降低8%～47%,还能将尾流完成时间降低10%～70%.原型系统测试表明,与英伟达CX7相比,WRCC将短距离场景下尾时延降低7%～49%.在640 km长距离场景下,WRCC将平均时延降低2%～7%,尾时延降低45%～49%,平均吞吐量提升26%～90%.
To address the long control loop and hybrid traffic congestion issues caused by cross-domain interconnection scenarios of remote direct memory access(RDMA)technology,a congestion control method for computing power net-works,named WRCC(WAN RDMA congestion control),was proposed.A fair rate computing strategy based on input rate was employed,enabling switches to accurately calculate the port fair rate of congested queues.Combined with dual control loops on the near-source switch and in-band network telemetry technology,it achieved end-network collaboration rate control and rapidly responded to congestion.Simulation experiments demonstrate that compared with existing com-mercial methods,WRCC reduces the average and tail flow completion time by 8%～47%and 10%～70%.Prototype sys-tem tests indicate that compared with NVIDIA CX7,WRCC reduces the tail latency by 7%～49%in short-distance sce-narios.In long-distance scenarios of 640 kilometers,WRCC reduces the average and tail latency by 2%～7%and 45%～49%,while achieving the average throughput improvement of 26%～90%.
刘亚萍;严定宇;方滨兴;许名广;张硕;杨智凯
广州大学网络空间安全学院,广东 广州 510006||鹏城实验室,广东 深圳 518108鹏城实验室,广东 深圳 518108||北京邮电大学可信分布式计算与服务教育部重点实验室,北京 100876广州大学网络空间安全学院,广东 广州 510006鹏城实验室,广东 深圳 518108广州大学网络空间安全学院,广东 广州 510006||鹏城实验室,广东 深圳 518108广州大学网络空间安全学院,广东 广州 510006
信息技术与安全科学
拥塞控制远程直接内存访问算力网络端网协同
congestion controlremote direct memory accesscomputing power networkend-network collaboration
《通信学报》 2026 (2)
109-124,16
新一代人工智能国家科技重大专项基金资助项目(No.2025ZD0122203) The New Generation Artificial Intelligence-National Science and Technology Major Project(No.2025ZD0122203)
评论