The Research On Performance Improvement Of NAND Flash-Based DBMS

With a proper software layer, known as the Flash Translation Layer, which makes linear NAND Flash appear to upper layer like a disk drive, conventional disk-based database algorithms and access methods will function adequately without any modification. On the other hand, due to a few limitations of flash memory, this approach is not likely to yield the best attainable performance. Because writing a sector is much slower than reading a sector and overwriting a sector must be preceded by erasing the erase unit containing the sector, the effective write bandwidth of flash memory is even worse than disk. NAND Flash has poor write performance, particularly when small-to-moderate sized writes are requested in a random, which is quite a common access pattern for database applications such as on-line transaction processing (OLTP). So in order to effectively utilize flash memory as data storage media, we have to alter some data structures and algorithms of database management system.In this paper we analyze NAND Flash's two flash translation layers:FTL and NFTL. We choose NFTL as the future flash translation layer for its moderate main memory cost. FTL is popular in small size NAND Flash and there is algorithm performance model on FTL, but there is a lack of algorithm performance model on NFTL. So we present NEWOM model as algorithm performance model on NFTL, which is improved from EWOM model on FTL. Then we analyze the sorting algorithm and join algorithm of NAND Flash based database system's inquiry processing, find out the factors of their I/O performance, and present the formula of their I/O performance. Then we analyze three page layouts and RARE-join algorithm on PAX page layout, RARE-join algorithm has better I/O performance than Grace-hash join, but PAX page layout's performance is worse than NSM page layout for multiple attributes'values are updated. NAND Flash's reading is much faster then writing, so hit rate can't reflect the memory buffer's I/O performance, we have to analyze the buffer management algorithms:LRU algorithm and CFLRU algorithm. Then we analyze the in-page logging approach (IPL), which has many shortcomings. So we present logging page in-erase block approach, which improves the IPL approach's I/O performance by reducing main memory cost and increase the utilization ratio of logging page, further more, we use PAX page layout as NAND Flash based database management system's page layout and logging page in-erase block approach can eliminate PAX page layout's shortcoming. We demenstrate that logging page in-erase block approach has better I/O performance than IPL approach by simulation experiment. In order to improve DBMS's performance and reduce its cost, we present log disk based logging page in-erase block approach, which utilizes disk to store log, utilizes NAND Flash to store data. At last, we simulate the log disk based logging page in-erase block approach and testify that log disk based logging page in-erase block approach has better I/O performance than logging page in-erase block approach and IPL approach.