Simulation And Performance Analysis Of Hybrid Memory System For In-memory Database

With the rapid development of the information age,computing power and storage requirements have grown rapidly,presenting greater challenges to database systems in different application scenarios and areas.In order to meet the growing demand,researchers have explored the development of new types of in-memory database systems to adapt to current needs and adopted new storage devices to meet performance optimization goals.In the process of exploration,the development and gradual maturity of new non-volatile memory(NVM)can theoretically solve many problems in the current database architecture.However,NVM devices have not yet been mass-produced on a large scale.In order to analyze the impact of NVM characteristics on database performance,it is necessary to simulate NVM latency and design a corresponding hybrid memory system,and analyse bottleneck at the same time.The basic features of non-volatile memory and comparison of different hybrid memory architectures are thoroughly studied,and this paper analyses the application characteristics of database systems called H-Store and N-Store.A series of tools have been developed for system simulation and performance analysis.Design a tool to track the dynamic tracking of memory information by Intel Pin and code annotations.The function of NVM latency simulation is designed by using Linux system functions,and its accuracy was evaluated.Design a tool with the function of extracting and analyzing the accessing information,and the characteristics of the database are analyzed from the perspective of accessing.To apply NVM reasonably in the database and combine database features,a hybrid memory system based on Anti-Caching and NVM as external memory was proposed to solve the problems of small memory capacity and power consumption.A hybrid memory strategy of NVM and DRAM at the same level was analyzed and designed,and the simulated hybrid memory system experiment and performance analysis were performed for different access granularity.Through the above experiments and tools for performance analysis,it is concluded that in-memory database,the execution cost of the storage engine is reduced in the entire system and is no longer a bottleneck in the system.At the same time,distributed control and extensibility design become the bottleneck of the current system.