您的浏览器禁用了JavaScript(一种计算机语言,用以实现您与网页的交互),请解除该禁用,或者联系我们。[新华三技术有限公司]:DDC 技术白皮书 - 发现报告

DDC 技术白皮书

AI智能总结
查看更多
DDC 技术白皮书

i目录1概述······························································································································· 11.1产生背景······················································································································11.1.1传统交换机的限制·································································································11.1.2 DDC技术············································································································11.2技术优势······················································································································22DDC技术实现·················································································································· 32.1 DDC物理架构··············································································································32.1.1 DDC物理设备概述································································································32.1.2 DDC的设备连接要求·····························································································42.2 DDC的数据转发机制·····································································································42.2.1概述···················································································································42.2.2技术优势·············································································································52.2.3数据转发表项的生成和同步·····················································································62.2.4基于信元(Cell)的转发·························································································82.2.5基于VOQ的拥塞控制····························································································92.2.6 Cell网络数据转发流程························································································· 103H3C实现的技术特色······································································································· 113.1去中心化···················································································································· 113.2开放性······················································································································ 124DDC在人工智能数据中心(AIDC)中的典型组网应用··························································· 13 1概述1.1产生背景1.1.1传统交换机的限制随着大数据、云计算和人工智能技术的快速发展,数据中心遭遇了流量激增的挑战。如何快速处理这些流量,给数据中心的核心交换机带来了巨大的压力。数据中心的核心交换机均为传统框式交换机。传统框式交换机通常是封闭式的、集中式的大型机箱(Chassis),就像一个“大铁柜”。其中所有的组件如主控引擎、交换板、接口板等都集中在一个柜子里。这种设计虽然集中且易于管理,但存在以下局限性:•扩展性困境传统机框式交换机端口密度受限于机框槽位数量,一旦槽位用尽就必须整机更换或购买新机。•能效挑战传统的大型框式交换机在交换芯片技术不断进步、交换容量不断增大(从100G迈向400G)的同时,也带来了功耗的显著提升。一个16槽位、全400G端口的框式交换机可能需要高达4到5万瓦的电力供应,这对许多老旧机房的设备升级构成了巨大挑战,尤其是当机柜的电力供应无法满足这一需求时。1.1.2 DDC技术DDC(Diversified Dynamic-Connectivity,多元动态联接)技术是一种创新的网络架构设计,它打破传统的集中式机框交换机设计,采用分布式解耦的方法来提高数据中心网络的灵活性和可扩展性。•物理层面:DDC将一台庞大的机框结构分解为多台盒式交换机。如图1-1所示,DDC技术通过将传统的大型网络交换机拆分为更小的、独立的模块化组件,即盒式交换机,从而相当于将“大铁柜”分解成“积木”,实现了网络功能的分散部署。这些盒式交换机可以担任转发接口板或交换板的角色,它们被分散安装在多个机柜中,从而提 1 2供了更好的散热管理、功耗控制,并且克服了设备升级和空间扩展的局限,大大增强了网络部署的灵活性和便捷性。图1-1传统交换机和DDC设备对比•数据转发层面:DDC将多台盒式交换机设备之间的物理连接整合成一个Cell(信元)转发网络,使得业务报文在Cell网络内的转发就像在框式设备内部转发一样快速、高效。1.2技术优势DDC具有以下显著优势:•灵活扩展如同积木可以自由组合,DDC允许在无需更换核心硬件的情况下,轻松添加新的盒式交换机,灵活支持网络的动态扩展,以适应不断增长的业务需求。•极速转发DDC基于VOQ(Virtual Output Queue,虚拟输出队列)和Cell交换等先进硬件技术。VOQ技术能够保证数据在DDC内部转发时不丢包,Cell转发技术能够保证数据流在DDC内部转发时负载更加均衡,提升了DDC内部盒式交换机之间链路的利用率和吞吐量,充分满足HPC(High Performance Computing,高性能计算)业务对传输网络提出的低转发时延、低丢包率的严格要求。•超强可靠DDC提供亚秒级故障恢复能力,可完美适配万卡AI集群使用场景。•绿色环保DDC架构可以提高能源效率,因为它允许精确的容量规划和动态电源管理,从而降低了电力消耗和冷却需求。综上所述,DDC可提供卓越的灵活性、扩展性和鲁棒性,是现代数据中心网络架构的优选方案。NCFNCPNCFNCPNCP 2DDC技术实现2.1 DDC物理架构2.1.1 DDC物理设备概述图2-1DDC物理架构如图2-1所示,H3C的DDC方案将传统的集中式机框分布式解耦成两种类型的物理设备:•NCF(Network Connectivity Fabric,网络交换单元)NCF类似于框式交换机的交换网板,用于透传报文。如果报文的入接口和出接口分布在不同的NCP,则需要通过NCF将入口NCP收到的报文转发给出口NCP。NCF上无业务接口,使用专属物理接口SFI(SerDes Framer Interface)接口来传输NCP之间的数据报文。•NCP(Network Connectivity Processor,网络业务单元)NCP类似于框式交换机的主控板加业务板,负责处理协议报文并转发业务报文。NCP上的SFI接口用来连接NCF。NCP上的业务口用来连接业务网,是整个DDC对外的输入、输出接口。S12500AI系列设备的端口情况如表2-1所示。表2-1S12500AI系列设备的端口情况产品系列机箱类型H3C S12500AI系列NCFNCPIngress入报文接口板主控板 3产品型号接口情况S12500AI-128EP-NCFN•128个OSFP800 SFI口S12500AI-36DH20EP-NCPN•20个OSFP800 SFI口•36个QSFP 112业务口 DDC组网中,建议每种类型的物理设备至少部署两台,从而提高网络的可靠性。2.1.2 DDC的设备连接要求在DDC中,NCP之间的报文通过NCF透传,控制报文和数据报文共用传输通道,连接要求如下:•每个NCP均通过SFI接口直接连接至NCF,无需在NCP之间或NCF之间增加额外连接。•一台NCF上有128个SFI口,每台NCP上有20个SFI口,一台NCF满配可实现6台NCP的互联互通。•DDC组网中,建议每种类型的物理设备至少部署两台,每个NCP至少和两台NCF相连,从而提高网络的可靠性。图2-2DDC连接拓扑图2.2 DDC的数据转发机制当前DDC仅支持三层转发,本文仅描述DDC的转发原理。2.2.1概述DDC通过将框式交换机解耦为多台盒式交换机,继承了盒式设备灵活扩容、易维护升级的优势,同时分散了机房承重、供电和散热压力。但其核心技术挑战在于:如何确保业务报文在分布式盒式交换机间的转发效率达到传统框式交换机的水平?这一问题的解决依赖于DDC的创新数据转发机制。NCP 1 4NCF 1NCF 2NCP 4NCP 2NCP 3 5如图2-3所示,传统Spine-Leaf架构通过业务接口构建了一张普通以太网转发网络,DDC基于SFI接口以及内部协议的交互,相连的NCP和NCF会自动组建Cell转发网络。在这个网络中,所有的NCP和NCF设备在转发平面对外相当于一台设备,网络内部的设备间则采用Cell转发方式进行数据转发。•传统Spine-Leaf架构的局限性依赖五元组(5-tuple)的逐流ECMP负载均衡,容易因哈希冲突导致流量分配不均,尤其是在大流量(如大象流)场景下,可能造成部分链路拥塞而其他链路闲置。为避免拥塞,通常需要预留额外带宽并调整收敛比,导致资源利用率降低。•DDC的创新解决方案Cell交换技术:将报文切分为固定大小的Cell单元,并在多路径上并行转发,实现跨节点(NCP)的