Performance Analysis and Evaluation ofInfiniBand FDR and 40GigE RoCE on HPCand Cloud Computing SystemsIntroductionNETWORK-BASECOMPUTLABOROutlineIntroductionCloud Computing applicationsPerformance Analysis and Evaluation(4 ConclusionIEEE Hot Interconnects August 23, 20123/28IntroductionNETWORK-BASECOMPUTLABORIntroductionCommodity clusters continue to be very popular in HPC andcloudsHPC applications and cloud computing middleware(e.gHadoop) have varying communication and computationcharacteristicsNew Pcle gen 3 interface can now deliver speeds up to 128psHigh performance interconnects are capable of deliveringspeeds up to 54 GbpsNew Mellanox's ConnectX-3 Fdr (54 Gbps /roce 40 gigeIEEE Hot Interconnects August 23, 20124/28IntroductionOverview of Network Protocol StacKs K aHigh Performance Computing /Cloud Computing ApplicationsSocketsVerbsernel spaceRDMARDMEthernet DriverUser SpacUser spa10/40 GigEQDR/FDR10/40GgEQDR/FDRInfiniBandInfiniBandAdapterAdapterAdapte10/40 GigEQDR/FDR10/40GgEEthernetInfiniBandPolBROCERoCE: allows the rdma of infiniBand to run over EthernetConnectX-2: 10 GigE in Roce mode or QDR (32 Gbps) in IB modeConnectX-3: 40 GigE in RocE mode or FDr (54 Gbps )in IB modeIEEE Hot Interconnects August 23, 20125/28IntroductionNETWORK-BASECOMPUTLABORProblem statementHow much benefit can the user of a hPc/ cloud installationhope to see by utilizing IB FDR/ RocE 40 GigE over IBQDR and RocE 10 GigE interconnects, respectively?How does Infini band compare with roce in terms ofperformance?IEEE Hot Interconnects August 23, 2012Cloud Computing ApplicationsNETWORK-BASECOMPUTLABOROutlineIntroductionCloud Computing applicationsPerformance Analysis and Evaluation(4 ConclusionIEEE Hot Interconnects August 23, 20127/28Cloud Computing ApplicationsNETWORK-BASECOMPUTLABORCloud Computing MiddlewareCloud computing economies have gained significantmomentum and popularityRequired the highest performance and reliability availableApache Hadoop is the most popular framework for runningapplications on large cluster built of commodity hardwareMajor components of Hadoop usedHDFSHBaseIEEE Hot Interconnects August 23, 2012Cloud Computing ApplicationsNETWORK-BASECOMPUTLABORHDFSHadoop Distributed File System (HDFS)is the underlyingfile system for Hadoop frameworkHDFS is designed for storing very large files on clusters ofcommodity hardwareTwo main types of nodesNameNode: responsible for storing and managing themetadataDataNode: act as storage for hdfs filesFiles are usually divided into fixed-sized (64 MB)blocks andstored as independent unitsEach block is also replicated to multiple(typically threeDataNodes in order to provide fault tolerance and availabilityIEEE Hot Interconnects August 23, 2012Cloud Computing ApplicationsNETWORK-BASECOMPUTLABORHBase and YCsBHBasedeveloped as part of the apache Hadoop project●Java- based databaseo Runs on top of the Hadoop frameworke Used to host very large tables with many billions of entriese Provides capabilities similar to Google's Big TableYahoo! Cloud Serving Benchmark● Used as our work|oadFacilitates performance comparisons of different key /value-pair and clouddata serving systemsDefines a core set of benchmarks for four widely used systems: HBaseCassandra, PNUTS and a simple shared MySQL implementationIEEE Hot Interconnects August 23, 201210/28