1、可预期数据中心网络周维，席永青阿里云资深技术专家，阿里云高级技术专家Predictable Datacenter NetworkWei Zhou,Yongqing XiContents目录01可预期高性能的数据中心02可预期网络质量Predictable high-performance data centerPredictable network quality可预期高性能的数据中心build Datacenter Network with Predictable High performance01数据中心网络data center network网络服务：高效稳定的链接Network S

2、ervice计算服务：低时延高带宽的IOComputing IO ServiceNetworkServicesComputing IO数据中心发展趋势AI/BD workload growthAI算力需求持续增长，依赖网络释放更大算力，解除通信规模和时延的束缚Heterogeneous&high performance storageARM+RISC-V+GPU，accelerator新型算力模型和存储介质open inter-conn to break the closed systemCXL等inbox开放互联衍生出新的开放式multi-host互联，异构+池化数据中心趋势端网协同DC a

3、s a Computer，Network as IO带宽持续升级、极限低时延技术inbox PCIe+outbox Eth-inbox、inrack、inbuilding分级、分池端网协同、软硬一体trend of data center network development数据中心发展趋势trend of data center developmentDC as a Computerscale upMoores Law增速放缓爆炸半径大规模分布式Best-effort IO互联分布式计算效率高带宽 vs 低时延数据传输的能耗有效算力能耗高性能的同时，要求可预期的结果scale outene

4、rgy数据中心高性能网络现状静态动态拥塞1us1us1us1us1usNusNus1us1us1us1ususmsNusNus0.5us/100m0.5us/100mqueuing&scheduling静态时延、吞吐的要点0 copy，DATA DMA协议/CC-硬化单跳低时延，减少跳数大带宽更短的路径距离23445动态时延、吞吐的要点网络语义和操作调度高效精确的流控算法更好更精细的拥塞反馈均衡的负载和优化的路径1345协议/CCAPPmem copy协议/CCAPPmem copy123454网卡硬件卸载拥塞控制算法物理网络架构端网协同互联协议典型距离150m，静态时延约20us拥塞调度St

5、atus of High-Performance Networks in Data Centers传统方式AAPPDDDPPAAProtoAPPProtoAPP网络端网协同AAPPDDDPPAAProtoAPPProtoAPP拥塞链路故障网络路径拥塞程度感知（道路拥堵）故障状态感知（道路维修）路径信息感知（导航选择）端侧网络栈对物理fabric状态不感知端端网端端网端+网+全视角无导航，盲开blind driving,No navigation导航反馈，城市大脑navigation system，City Brain阿里巴巴端网协同数据中心网络Alibaba data center netwo

6、rk with terminal networks integration阿里巴巴高性能可预期数据中心实践实时状态感知IO模型的拓扑抽象拥塞优化HPCCRDMAAlibabas high-performance predictable data center network practiceReal-time status awarenessHPCC to optimize congestion controlTopological abstraction for a specific IO patternDC as computer，Network as IO低时延全栈优化和卸载端网协同流控

7、充分满足 计算、存储、大数据、AI等，分布式大规模系统的集群IO互联性能SLA可用性SLA目标转变为性能SLA目标稳定IO网络事件的可视和感知，ms级异常切换基于端网协同的可预期高性能数据中心软硬一体全自研数据中心系统自研协议Solar-RDMA自研流控HPCCHigh Availability,Intelligence,and Low latencyHAIL数据中心架构全栈自研全自研的网卡、交换机、光互联自动化平台NET可视化监控北斗、telemetry阿里巴巴高性能可预期数据中心体系Alibabas high-performance predictable data center netw

8、ork systemTORCPU xPU CPU xPU CPU xPU CPU xPU CPU xPU CPU xPU CPU xPU CPU xPU CPU xPU CPU xPU CPU xPU CPU xPU CPU xPU CPU xPU CPU xPU CPU xPU TOR1122m距离，1us100m距离，5n*10us2m距离，1usEthernetPCIe/CXLPCIe/CXLEthernet path 层端网协同低时延、可预期网络服务可预期高性能数据中心的未来Expect the future of Predictable&high-performance data c

9、enters Networks全栈协同，软硬一体可预期的性能和服务化Full stack collaboration and integration of software and hardwarePredictable performance and servicing可预期网络质量Predictable Network Quality02曾经的网络监控Network Monitoring in Old Days故障视角Focus on Network Outage事后响应Passive response缺乏网络质量指标Lack of Meaningful Network Quality M

10、etricsNetworkQualityMetricsSYSLOGSNMPPINGGRPCCLI可预期网络质量Predictable Network Quality网络质量路径丢包时延PathLatencyPacket LossNetwork Quality流级别精确路径Flow level network path纳秒级端到端，网络节点间时延Nanosecond level end-to-end,node-to-node latency measurement丢包关键信息及原因Packet loss information and root cause关键技术Key Technologies

11、Precision Time Protocol精确路径时延丢包In-band Network TelemetryIn-band Network TelemetryMirror on RetranSmartflowAggregate DropPathLatencyPacket LossHashlib关键技术-INTKey Technologies流级精确路径Flow level network path网络节点间时延Node to node latency 队列级信息Queue infoASWPSWDSWPSWASWCollectorINT probeINT metadata原始报文路径INT报

12、文路径DSCP标识原始报文INTINT反弹报文路径INTGRES-NCD-NC关键技术-PTPKey Technologies纳秒级端到端延时Nanosecond level end to end latency网卡硬件时间戳，去除协议栈抖动Use Hardware timestamp to improve accuracy 关键技术-MORKey Technologies镜像重传数据包Mirror on Retransmission PacketS-NCASWPSWDSWPSWASWCollectorRetrans PktGRED-NCRetrans PktGRE精细网络质量SLA虚拟网络全覆盖物理网络全覆盖未来规划Future PlanFull Coverage of Physical NetworkFull Coverage of Virtual NetworkMore Granular Network SLAFrom Best effort to Deterministic QoS THANKS