将“工作”放在网络中：网络内流数据分析.pdf

1、OCP Regional Summit|April 19,2023|Prague,CZPutting the work in network:in-network stream data analyticsMonica ChiosaSystems Group,Dept.of Computer Science ETH ZurichIn collaboration with Thomas Preusser,Michaela Blott and Gustavo AlonsoMotivationCLOUDCOMPUTINGStorageServersMobileApplicationsDatabase

2、sPrivateHow can we process the data in the cloud?FPGAASICHW Specialization and Power EfficiencyGeneralityFPGAmemorybusPCIeFPGACo-processor KernelsDMA EngineFPGAmemorybusPCIeFPGADDR ControllerNetworkStackVideo CtrlSound CtrlCommunication CtrlPass-through KernelsEthernetHBM ControllerStorage CtrlDMA E

3、ngineSetupPass-through KernelsApproximation computeExact computedata analyticsKernel setupNetwork setupNetworkStackTCP/IPRDMANetwork Setup:RDMAAPPLICATIONSIB Transport ProtocolIB Network LayerInfiniBandLink LayerIBInfiniBandManagementIB Transport ProtocolIB Network LayerEthernetLink LayerRoCEv1Ether

4、net/IPManagementRDMA Software StackIB Transport ProtocolUDPEthernetLink LayerRoCEv2Ethernet/IPManagementIPIB Transport ProtocolEthernetLink LayeriWARPEthernet/IPManagementTCPIPKernel Setup:Exact compute-Correlation coefficient computationCharacterize the linear relationship between two attributes(st

5、reams);It is quantizable,results belong to-1,1.Pearson Correlation Coefficient uses numerical values;Kernel Setup:Exact compute-Pearson Correlation coefficientWhere can it be used?MLDBPrivacyMedicinePhysicsEconomyKernel Setup:Exact compute Pearson Correlation coefficient sum of squares:sum of produc

6、ts:sum of elements:Kernel Setup:Exact compute Block DiagramData Streams Correlation CoefficientsKernel Setup:Exact compute ArchitectureCompute kernel can process imput data at every clock cycle(II=1).CPU baseline16-core AMD EPYC 7302PMulti-threaded baseline using Eigen Library32-