| Database indexing technology is a decades research field.With the development of in-memory databases and artificial intelligence technology,researchers can have a new perspective to consider the design of index structures.B+ tree is a prevalent index structure in disk-based database systems.However,according to recent research,the B+ tree index will occupy more than 50% of memory space,significantly reducing the space available for storing data.Compared with the traditional B+ tree index,the machine learning indexes have a certain degree of optimization in memory space occupation and query performance.However,the machine learning indexes still faces some improvement.Firstly,the long construction time of the model will lead to the extension of the recovery time of the in-memory database.Secondly,the current machine learning index can not effectively use the CPU cache to speed up the query.In addition,the restrictions on the modification of the machine learning indexes make it challenging to apply in practice.This study proposed a new index algorithm to solve the problems encountered in the machine learning index.The algorithm combines the sampling method and cache block data structures,which can accelerate the construction time of the index and reduce the memory occupation of the index at the same time.In addition,the study focuses on the problems that the learned indexes require data in the continuous memory address and do not provide modification methods.This study proposes an appending algorithm to meet the requirement of real-time data appending in realistic application environments.Main contributions of this paper are as follows:· This paper proposes an index structure based on a cache block-based cache geometric model(CBPGM).CBPGM index applies sampling technology on the index construction stage,which improves the construction speed of the index.The data sets are aligned and partitioned according to the size of cache blocks,making full use of the performance acceleration brought by the CPU cache.Experiments show that the algorithm can effectively reduce the memory occupation of the index structure.· This paper uses the algorithm of CBPGM to extend Radix Spline and TS.This paper proposes an overlapping construction algorithm to solve the incompatibility problems encountered in the construction process.Experiments show that the CBPGM algorithm can significantly improve the performance of the two indexes.· This paper proposes an optimized version of the index structure DCBPGM index(dynamic DCBPGM index).To meet the needs for modification in actual scenarios,this paper improves the CBPGM index and proposes DCBPGM.The index can support insertion,modification,and multithread concurrent access. |
