《智能網絡加速 5G 云服務.pdf》由會員分享,可在線閱讀,更多相關《智能網絡加速 5G 云服務.pdf(35頁珍藏版)》請在三個皮匠報告上搜索。
1、NVIDIAACCELERATE TELCO CLOUDWITH SMART NETWORK#page#DIGITAL TRANSFORMATION JOURNEYS FOR TELCOThe Years of 5GRise of Artificial Intelligence歡Race to the Edge of the NetworkExponential Growth ofThe Cloud Native TelcoIoT-Driven Devices#page#ACCELERATING TELCO WORKLOADSModern Telecom NetworkAcCe55R5G UP
2、F acceleration withTime-Triggered TransportEdgeTechnology for Telco (5T)for 5GGTP Offload(RAN0CoreCloud-Native deliver high-口Provision Bare-Metal Like aperformance Kubernetes forCloud Giant with DPUTelco EdgeUncompromised SDN Accelerationwith ASAP2 technology#page#SDN ANDCLOUDNATIVEACCELERATION#page
3、#SDN網絡的性能挑戰Software-defined networking (SDN) has major performance burdensLimited ThroughputHigher LatencyHigh CPU Utilizatiion#page#嵌入式交換功能是智能網卡卸載的核心(ESWITCH)Hardware offload for a virtual switch logice-switchAction BClassificationClassificationClassificationPackets InProcessed Packets OutBANAction
4、 AAction N#page#兩種OVS加速方式:SR-IOVVS.VIRTIOHOStHostVMVMVMVMVMVMVirtOirtloHypervisorNetwork adapterNetwork adapterVIRTIO ACCELERATIONSR-10V ACCELERATIONThis hardware acceleration provides increased networkThis hardware acceleration provides SR-loy networkperformancewhilemaintainingthesoftwareVirtioperf
5、ormance and enable SDN and Overlay support for theimplementationVNICs to theVM#page#CONNECTX-6DX支持更高性能的SDN加速Accelerate any cloud infrastructureSRIOV support with up to 90M PPSAccelerate Virtualized and CloudAccelerate Virtio by 3x with Zero CPU %Native EnvironmentsSupport legacy guestOswith seamless
6、 integrationConnection Tracking Windows Validation客Orchestration and Commercial LinuxopenstackkubernetesInbox in RHELand Red Hat OSPK8sintegration(Roadmap)Virtio acceleration25120%Connectx100%20OvSNeutron80%Plusin15Open vSwitc60%1040%20%DRASAP2VirtioVirtioCompute Node#page#KUBERNETES的SDN加速Accelerati
7、ng Bare-Metal Kubernetes Networking,The Right WayAccelerating all pod-to-podHost(x86)K8s Networkcommunication,ClusteripContainerservice,etc.PluainethoOVSFlexible solution forrepaccelerating the primarynetwork(Kubernetes API)orsecondary network withmeta CNI (Multus)ConnectXLeveraging advancedvSwitch
8、Offload RuleseSwitchoffloadfunctionalitiesincluding overlay networkencap/decap,connectiontracking and NATUpliink#page#共同推進電信云的SDN加速生態Accelerating EverythingSDNOrchestration來nuagenetworksANTREAkubernetesopenstack.OVN開QUOvSSDNtungstenfabricHereu! suloeid NOS injamod ainos-uado Jo asuel e saelajae ppoi
9、yO SAO JINHJBUSbare-metal,virtualizedand containerizedenvironmentsanVIDIA#page#BAREMETAL CLOUD#page#裸金屬云的主要業務需求Ability to provide bare-metal like performanceOfload control& data planes from the host=lsolate & secure compute resourceAbility to virtualized storage & networkresourcesEnable flexibility
10、and scalability=In server provisioning= In server configuration (network and storage)Private/Public cloud service in which the customer rents the entire serverresources from a cloud providernVIDL#page#SMARTNIC支持BARE-METAL場景Full-featured SDN capabilitiesASoftware-definedFull orchestration through ups
11、tream OpenStack Neutron APIsNetworkingNo installation of network driver in bare-metal hostSame flexibility as virtualized storage,same performance as local storageStorage-DefinedOperating System is agnostic; requires only a NVMe driverStorageDynamically allocates cloud storage and back-ups in the st
12、orage cloudComplete isolation of security control enforcement from the tenants workloadBare-Metal CloudEnabling diverse cyber security solutions, enhancing data-center securitySecurityNo need to install agents on servers or any impact on server performance#page#什么是DPU(DATA PROCESSING UNIT)Software D
13、efined Data Center Infrastructure-on-a-Chip換山理NVIDIA NICNVIDIADPU withArm Cores&AcceleratorsSofwedefinedStorasTo Software Defined InfrastructureTo Software Defined InfrastructureFrom Hardwareon CPUApplianceson DPU#page#BLUEFIELD-2DATA PROCESSING UNITData Center Infrastructure-on-a-ChipACCELERATONENG
14、INEXXNN200Gb/s Ethernet8 InfiniBandNEWORKNTERFACEPROGRAMMABLARM CORESPowered by ConnectX-6 Dx8ARM A72 CPUs subsystemSecurity Networking, Storage accelerationsACCELERATIONPCIEGEN40ENGINESWITCHOpen-source reference codeDORAMEMORYNNTERFACE#page#采用DPU實現基礎平臺能力Software Defined NetworkingSoftware Defined S
15、torageSoftware Defined Security(181))SavRoutervSwitchNAT/LoadVMs8Telco/NFVNVMe-oFDataDistributed IDS/IPSRoot ofOJWWDDOSElasticCompression DeDupStorageDirect EncryptionStorageSegmentatfonPreventionBalancer ContainersNG FirewallTrustDPUDPUDPUDPUDPUDPU#page#NVIDIADOCA組件DPU SDK- Developer Community- Acc
16、elerated TTMAPPLICATIONSMultiple OSSoftware DefinedSoftware DefinedSoftware DefinedInfrastructure時學StorageSecurityManagementNetworkingAllArchitecturesDOCA SDKUser ModeManagementStorageSecurityNetworkingTelemetrySPDKDPDKDPDK/P4armALL Programmable ResourcesProgrammableRDMAResource VirtualizationFul Ap
17、plication on DPUNVMeRoTASAP2RDMACryptoKernelSNAP3Open Source, Openn APIs#page#DPU實現裸金屬SDNBare-metal HostBare-metal HostOperating-SystemOperating-SystemTenants DomainAirtioProviders DomainH復品eld SmartNICielan中TOR SwitchTOR SwitchTOR Switch NetworkingSDN IntegrationFull-featured SDN hardware-accelerat
18、ed capabilitiesxLimited to no SDN capabilitiesxOrchestration through proprietary TOR switchFull orchestration through upstream OpenStackvendor pluginsNo installation of network driver in bare-metal hostrinstallationirbare-metnVID#page#Compute PlatformsDPU加速存儲業務Disrupting Enterprise Cloud EconomicsHO
19、ST OsNVMeVirtio-blkEmulate remote storage to appears as local to the host OSBlueField-2 SNAP FrameworkDynamically assigned storage,not bound by physical capacityVirtualized or Bare Metal CloudOver-provisioning, scaled to rack/clusterInbox standard driversOS agnostic supports legacy OSsRemote Storage
20、#page#DPU將安全帶入主機側Cloud ServerCloud Server記IDS店WorkloadWorkloadWorkloadWorkloadDevOpsDevOpsL4-L7 InspectionNGFWWorkloadWorkloadFirewall / Micro-segmentationWITHOUT DPUWITH DPUEncryption(Software)Anti-MalwareWorkloadWorkloadSoftwareSoftware DefineddefinedNetworking (SDN)Storage (SDS)IsolationOptionalL
21、4-L7 IspectionITOpsIT OpsU-Segmentation/NGFW/CnyptoSDN&SDS品等會意Core Data Center Perimmeter#page#DPU提升裸金屬上的資源利用和性能DPU free up server resource for its primary application tasksVirtualized&Software DefinedBare MetalSoftware DefinedHardware AcceleratedCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreC
22、oreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreCoreSmart NICCoreMoney Maker ApplicationCorMoney Maker ApplicationCoreSof
23、tware-Defined NetworkingDone by HardwareCoreCoreDeep Packet InspectionEncryption23nvio#page#ROCE AND LOSSLESS#page#RDMA(ROCE)與無損網絡技術RDMA vS.TCP/IFRDMARoCETCP/IPServer-InitiatorServer-TargetServer-InitiatorServer-TargetApplicationApplicationApplicationApplication000SocketsSocketsNIC DriverNIC DriverR
24、oCERNICBufeRNICBufeNICNICDuHBar#page#RDMA加速存儲與AICaffe2fhadoopcephROCEicrosoftSAPHANAORACLEAzUrGTensorFlowsBenwwaredataonExceleroHewtPackardpriseRDMAIBMn NetAppNUTANXS0verEthernetWesternOPURESTORAGEWDDigitalNVIDIA#page#ROCE加速分布式存儲WindowsServer2016S2D-WriteloPsSATSSDs(100s)WindowsServer2016S2D-ReadIop
25、司Storage is Getting a Lot Fasterl#page#RDMA加速AIRDMA Supercharges Leading AI FrameworksMicrosoftSG Caffe22IBMPaddlePaddleTTensorFlowUp toUp to60%2.5X95%50%Higher ROIBetterScalingSavings on CapitalPerformanceEfficiency8 Operation CostTencent騰訊lnl ByteDanceBaid百度DRACLE中國D賽新SONY2RaH歡NVIDIAnVID#page#無損網絡
26、(ROCE)對交換機的要求Simple Configuration1-Command CLI Config=1-ctick GUI ConfigHigh Performance=High PPS f Low LatencyROCEFair 8 Predictable PerformanceAdvanced Congestion ControlEarly Detectiona PreventionRoCE over VXLANExtensive VisibilitySingle Pane-of-Glass=Realtime RoCE Telemetry#page#RoCE一鍵式配置NVIDIA
27、Mellanox “Do RoCE”ROCE Configuration otherwise公樂區TE西術大switch(config)#roce24 Linesl#page#如何保證RoCE實現優化的性能Fairness 8 QoSAvoidable Packet LossCongestion ManagementSpectrumSpectrumMicroburstAbsorption131Competition0.3新蘇12864B1.5BPacketSizeCompetitionSpectrumCompetitionLowest LatencyTEST50%REPORTTm/tallly
28、#page#RoCE支持基于VXLAN及云化網絡場景RoCE Performance with VXLAN FabricROCE + VXLANVXLAN FabricQuality of Service within VXLAN tunnelsPreserve ECN even with VXLAN電用:Unique Mellanox differentiatorOther vendors dont copy QOS between outer 8Inner headersUse Cases: VXLAN +GPU PODSMachine LearningNVMe over Fabricsm
29、Storage Spaces DirectmSparkROCE#page#無損網絡應具備錯誤診斷能力Show RoCE Interface:Show RoCE Global:= RoCE Traffic= RoCE Configurationm Packet 8 Byte Count= RoCE Mode (Lossless/lossy)=Packet dropsmBuffer settingsCurrenta max buffer usagem Port congestion control settingsmPFC settings=Congestion Notification Pack
30、ets:=Priorities to Buffers MappingPacket a ByteCountPacket dropsTCmappingCurrenta max buffer usageWhich Queues are for ROCEmETS settingsECN:Buffer PoolsECN marked packetsBuffer memory allocationmMax buffer usage=Pause (PFC):Type:Lossy/losslessm Pause packet countPause duratiion#page#更多ROCE參考技術材料Nvidia Mellanox RoCE home pagehttps:/ guidance for RoCE configurationshttps:/ RoCE success cases and deployment guideshttps:/ vendor tested configurationsRoCE configuration for Arista switchesRoCE configuration for Cisco swittchees#page#NVIDIA