Research On Learned Index Based On NVM

In the big data era,the in-memory database based on the current computer architecture faces severe challenges in power consumption,capacity and performance.The emergence of Non-Volatile Memory(NVM)has changed the computer architecture,breaking through the bottleneck of energy consumption and capacity.DRAM possesses byte addressability and fast memory access,while NVM has other advantages over non-volatile,high storage density,and low static power consumption,which makes NVM a promising alternative to DRAM.With NVM as part of the storage system,an efficient index structure is vital to improve database performance.The learning index provides a new solution to improve database performance.The learning index introduces machine learning into the database,and research has shown that the performance of the learning index is significantly better than the traditional B+ Tree.This paper focuses on the characteristics of NVM and combines the related technologies of learning index to propose a learning index for non-volatile memory.We designs the index structure and memory architecture,and optimizes the operation algorithm of the index.This paper designs the index structure and memory architecture,and optimizes the operation algorithm of the index,so that the learned index can play a magical performance in NVM and hybrid architecture.The main research work is as follows:(1)A learning index LION for non-volatile memoryThe core idea of LION is to use the learning index instead of B+ Tree to improve the search speed on NVM and reduce the space overhead.It adopts the strategy of out-place insertion,designs the buffer,reduces the retraining of the learning model,and proposes the operation of the index.Compared with NVM-based indexes,LION has better read performance and better write performance.(2)The optimization method of LION under the hybrid memory architectureUnder the current technical conditions,the performance of NVM still has a certain gap compared with DRAM,and it cannot completely replace DRAM as the only main memory.In response to this problem,this paper proposes LION for the combination of non-volatile memory and DRAM,using DRAM to improve index read and write performance.The core idea of LION,which is based on the hybrid architecture of NVM and DRAM,is to put easily reconfigurable data structures into DRAM and use batch updates to reduce the write cost of NVM,and design a write-optimized log to ensure crash consistency.Compared with the current index that also adopts a hybrid architecture,LION has better read and write performance...