How to Architect a Cloud-Native Memory-Optimized Database Engine-马文龙.pdf

1、Security Level:Security Level:HUAWEI TECHNOLOGIES CO.,LTDHiEngineHiEngine:How to Architect a Cloud-Native How to Architect a Cloud-Native Memory-Optimized Database Engine Memory-Optimized Database Engine Wenlong MaCloud Database Innovation Lab of Cloud BU,Huawei Research CSecurity Level:Security Lev

2、el:HUAWEI TECHNOLOGIES CO.,LTD3ContentContent Background and Motivation HiEngine Architecture Storage and Index Transaction Management Experimental Evaluation Summary4Memory-centric Computing ArchitectureMemory-centric Computing ArchitectureNew persistent memory provides DRAM-level performance and F

3、lash-level capacityIn the SCM era,computing architectures tend to be memory-centricPooled memory appliance has gradually become the direction of explorationAcademia:HydraDB,RAM Cloud,NAM-DB,Hotpot,DDC,InfiniswapIndustry:FaRM,SAP HANA,SRSS SCM,WSCs,DAOSCustomers of in-memory databases migrate applica

4、tions to the cloudIt becomes necessary for vendors to offer cloud-native,memory-optimized OLTP solutions OLTP Database JungleOLTP Database JungleOn-premise Storage-centric Engines MySQL/PG/SQL ServerOn-premise Memory-centric Engines Hekaton/ERMIACloud-native Memory-centric Engines?Prototype:NAM-DB/F

5、aRM Production:HiEngineCloud-native Storage-centric Engines Aurora/PolarDB/GaussDB(for MySQL)Block-based IOPage-orientedBuffer Pool Most of in-memory databases are designed based on on-premise solutions Do not directly work in the cloud and they are untappedByte-addressable IOTuple-level No Buffer P

6、oolHuawei Cloud InfrastructureHuawei Cloud InfrastructureHardware Trends and ChallengesPersistent memory in the decoupled architectureOpportunity:equip persistent memory on compute nodesChallenge:goes against the stateless nature of compute nodesARM-based manycore processorsOpportunity:Better price