Research On Key Problems Of Flash-based DBMS

Flash memory is a novel solid-state storage medium and first appeared in 1980s.Flash memory has many advantages over traditional magnetic disk such as high-speed,non-volatility,low power consumption,shock resistence, and portability,so it is a preferred storage medium to overcome the limitations of magnetic disk.In recent years,flash memory is widely used in various embedded systems and portable devices.With the rapid increase of capacity and decrease of price,flash memory become a new important secondary storage device,and has been applied in enterprise computing environment.New challenges are posed by the increasingly various and complex applications to the data management on flash memory.Managing data with database technique,i.e.,establishing flash-based DBMS,is a preferred approach to meet the challenges of flash data management.Due to the distinct characteristics of flash memory,directly transplanting traditional magnetic disk based database technique to flash memory could not fully play the performance advantages of flash memory.Therefore,starting with the characteristics of flash memory,researching on the problems in flash-based DBMS according to the data access patterns of database,has important theoretic significance and value.This dissertation summarized the existing researches on flash-based DBMS and made deeply study on several key problems,including storage management,index and transactional recovery.Firstly,the dissertation introduced the characteristics and applications of flash memory.There are mainly two types of flash memory,NOR flash and NAND flash.The characteristics of NOR flash and NAND flash were also introduced,respectively.Moreover,a comparison was made between the characteristics and application areas of NOR flash and NAND flash.Secondly,index is one of the key techniques to improve the performance of database.However,in the existing methods,it is very costly to reconstruct the index node translation table after system failure.According to the characteristics of NOR and NAND flash memory,the dissertation proposed a reliable B+-tree implementation over compound flash storage architecture, using both snapshot and logging approaches to reconstruct the index node translation table efficiently after system failure.Thirdly,storage management is the basis of flash-based DBMS.In this dessertation,an out-page logging approach was proposed for storage management and could improve data update performance significantly. Furthermore,by combining the two approaches for flash data update,out-page logging approach and out-place update approach,an adaptive storage management was further proposed.This method could give consideration to both data update performance and data read performance,adapting to varying workloads.Finally,transactional recovery is an important component of flash-based DBMS.In this dessertation,according to the disadvantages of existing flash transactional recovery methods,such as high system overhead and high commit cost,an out-page-logging based transactional recovery method was proposed.The proposed method could reduce transaction commit cost, keeping good recovery performance.The contributions of the dissertation are listed as follows:(1) A reliable B+-tree implementation was proposed over compound flash storage architecture.By combining snapshot approach with logging approach,the proposed method could effciently reconstruct the index node translation table after system failure.(2) An adaptive storage management was proposed for flash memory. By combining out-page logging approach with out-place update approach,the proposed method could enhance data update performance significantly,giving consideration to data read performance,adapting to varying workloads.(3) A transactional recovery method based on out-page-logging approach was proposed.The proposed method could reduce transaction commit cost, providing good recovery performance.