嵌入式網(wǎng)絡(luò)協(xié)議引擎ETA加速網(wǎng)絡(luò)應(yīng)用程序處理

嵌入式網(wǎng)絡(luò)協(xié)議引擎嵌入式網(wǎng)絡(luò)協(xié)議引擎(ETA)加速網(wǎng)絡(luò)應(yīng)用程序處理加速網(wǎng)絡(luò)應(yīng)用程序處理作者：林明峰日期：2003-12概要 今天的網(wǎng)絡(luò)環(huán)境和網(wǎng)絡(luò)操作系統(tǒng) Linux 的tcp/ip stack 的性能 嵌入內(nèi)核的tcp 加速模塊 ETA的現(xiàn)實(shí)應(yīng)用今天的網(wǎng)絡(luò)環(huán)境和操作系統(tǒng)10G的以太網(wǎng)絡(luò)帶來(lái)什么 通用處理器必須每67奈秒處理一個(gè)數(shù)據(jù)包 超過(guò)通用處理器處理極限 協(xié)議占用太多時(shí)間NP是唯一的解決方案嗎?Intel Ixp1200帶來(lái)好處 網(wǎng)絡(luò)處理器對(duì)軟件系統(tǒng)結(jié)果帶來(lái)的問(wèn)題 靈活性和高性能的矛盾系統(tǒng)中斷 內(nèi)存拷貝緩沖區(qū)管理系統(tǒng)呼叫Tcp/ip協(xié)議應(yīng)用程序只能占用3%誰(shuí)占用的cpu的時(shí)間.Linux 的tcp/ip協(xié)議內(nèi)核的resource manager (ckrm)控制 diskIO 的資源分配 控制網(wǎng)絡(luò)的IO分配 目前還不成熟 http:/ 區(qū)分出獨(dú)立的協(xié)議處理層 固定一個(gè)cpu來(lái)專(zhuān)門(mén)處理協(xié)議 制定一個(gè)異步的通訊接口 新的socket庫(kù)來(lái)利用ETAETA System ArchitectureLANStorageIPCETA Host InterfaceIP StorageDriverFileSystemKernelApplicationsUser Socket ApplicationsSocketProxyNetwork FabricPacket Processing EngineDirect Transport InterfaceETAPacket Processing EngineNICApplication(Kernel or User)Adaptation LayerDTIEventQueueDTIRxQueueDTITxQueueAnonymousBufferPoolDTIDoorbellShared Host MemoryAppBuffersNICDTI Operation Model DTI operations:Connection requests(Connect,Listen,Bind,Accept,Close,)Data transfer requests(Send,Receive)Misc.operations(Set/Get Options,)EVENT AEVENT BEVENT CEventQTxQOP AOP CRxQOP BOP DDTIDoorbellProcess OperationServiceDoorbellDe-Queue OperationDescriptorPostCompletionEventPostETA InterruptEvent(if waiting)Host ApplicationAdaptation layerETA PPESoftwareGigabit NICs(5)ETA Host InterfaceKernelTestProgramCPU 0Host2.4 GhzCPU 1PPE2.4 GhzOff-the-shelf Linux ServersHostMemoryClientsClientsClientsClientsTestClientsKernelAbstraction LayerETA Test Environment2P SMP ProfileApp3%Kernel97%Idleapp3%tcp/ip26%bufmgt14%driver20%syscall7%Interrupt10%copy12%other8%idleProcessing requirements in multiple componentsTCP/IP is the largest single component,but is small compared to totalThe copy overhead is required to support legacy(synchronous)socket semanticsInterrupts and system calls are required in order to time-share the CPU resourcesETA Profile(1 host CPU+1 PPE)host app7%host kernel3%PPE51%idle39%host app8%host kernel3%PPE-tcp/ip12%PPE-bufmgt7%PPE-driver11%PPE-Queuing12%PPE-Other8%Idle39%Processing times are compressedIdle time represents CPU resource that is usable for applicationsAsynchronous queuing interface avoids copy overheadInterrupts avoided by not time-sharing CPUSystem calls avoided by ETA queuing modelETA2P SMPProfile Comparisonshost app8%host kernel3%PPE-tcp/ip12%PPE-bufmgt7%PPE-driver11%PPE-Queuing12%PPE-Other8%Idle39%app3%tcp/ip26%bufmgt14%driver20%syscall7%Interrupt10%copy12%other8%idleNormalized CPU UsageETA的應(yīng)用 基于內(nèi)核級(jí)別的ids程序 獨(dú)立于操作系統(tǒng)的包獲取 接管網(wǎng)卡