Research On Performance Optimization Of In-memory Database For Hybrid Memory

With the booming development of Internet technology in the era of big data,lots of emerging applications with high concurrency and low latency have emerged.They bring new challenges to the performance of traditional computer systems.At the same time,the amount of application data is gradually increasing,which puts forward greater demand for access speed of the database system.Due to the limitation of access speed,the disk-based traditional database cannot meet the increasing data processing demand,so the in-memory database system has become a good solution and has been widely used in today’s data centers,such as e-commerce websites,games apps and social networking sites,etc.In-memory database systems store data in memory,enabling faster response time by avoiding the latency of disk access.The read and write speed of memory media will have a significant impact on the performance of an in-memory database system.However,traditional Dynamic Random Access Memory(DRAM)faces the bottleneck of limited bandwidth due to scalability limitation in pin count.New stacked DRAM memories can provide orders of magnitude higher bandwidth to alleviate this problem.However,the capacity of the new stacked DRAM memory is limited due to the limited number of stacks.Thus,hybrid memory architecture composed of traditional DRAM memory and new stacked DRAM memory has emerged,which provides a new idea to address both capacity and bandwidth issues.This thesis focuses on building an in-memory database data migration approach for the hybrid memory composed of traditional DRAM and new stacked DRAM memory,to optimize the running performance of in-memory database system on hybrid memory architecture.The main work is summarized as follows:(1)The data access characteristic of in-memory database system is analyzed through preliminary experiment.In order to mine the data access behavior characteristic of the in-memory database system,this thesis combines it with the features of the hybrid memory structure and carries out a preliminary experimental study.The experimental results show that in the in-memory database system,there is a phenomenon that a small number of data objects bear most of the workload requests,that is,there are some hot data frequently accessed by clients in the in-memory database.(2)ADM,a data migration approach of in-memory database for hybrid memory,is proposed.According to the data access characteristic obtained from the above preliminary experiment,a data migration approach of in-memory database for hybrid memory is proposed to accelerate the performance of in-memory database systems based on hybrid memory architecture.By adding the functionality of data migration to in-memory database system,store a small amount of frequently accessed hot data in the stacked DRAM memory media to utilize its high bandwidth advantage,while storing most of the remaining data in traditional DRAM memory media to take advantage of its capacity advantage.The experimental results show that,compared with the in-memory database system without migration support on the hybrid memory architecture,the in-memory database system based on the ADM method can reduce the access latency by about 20%,read energy by 26% and write energy by 18% on average.(3)ADM-OPT,an optimized data migration approach based on adaptive threshold,is proposed.Further research is conducted on the performance of the ADM method for various workloads under different data migration thresholds,and it is found that different types of workloads are suitable for different thresholds.Therefore,a data migration optimization approach based on adaptive threshold,called ADM-OPT,is proposed.The experimental results show that the ADM-OPT method further reduces the migration overhead by about 58% compared with the ADM method.(4)Application-level implementation interfaces and data management mechanism are designed.Unlike most studies from the bottom-level perspective,the methods designed in this thesis are mainly oriented to the application-level,without modification to the underlying hardware.In order to realize the methods proposed in this thesis at the application-level,the implementation interfaces and data management mechanism of hybrid memory system at the application-level are designed.By directly modifying and expanding the source code of in-memory database system,the experimental verification is conducted.