Research On Buffer Management For Flash-based DBMS

With the advance of flash memory industry, flash disk has become widely used in computer system as a new type of storage device, and even takes place of magnetic disk in the future. Due to the uncommon characteristics of the flash memory, the traditional strategies which are designed for disk-based database may produce a deteriorative performance if applied directly. Now the improvement of buffer manager, store structure and data manager of traditional database model has become hot points of flash database. In this paper, we focuses on the buffer management of flash database in OLTP application environment.In this paper, we first compared the performance of MYSQL database which based on flash-database and magnetic-database respectively in details under the OLTP environment, which simulating by TPC-C standard. And then analyzed the flash database characteristics by the results. From the experiment results, we can see that flash memory does bring the advantages of traditional disk-based database, but also found the bottlenecks which produced by the uncommon characteristics of flash memory. To solve these bottlenecks, in this paper, we proposed a new buffer management named OF-Buffer which contains two improvements. First we proposed a new buffer page replacement named PWLRU. Based on the original LRU logical structure, PWLRU reduced the cost of buffer r/w operation through delaying writing back dirty pages. Second we also proposed RWO strategy for write-back action. RWO refractor random pages set to a page sequence through adding some clean pages, and then wrote back page sequence by sequential write operation.We compared the proposed PWLRU algorithm with the conventional LRU and CFLRU in terms of TPC-C benchmarks on the MYSQL database. The results indicates that PWLRU are effective in reducing energy costs with improved R/W performance, and achieving comparable hit ratio to the conventional LRU algorithm. Therefore, PWLRU algorithm is suitable for page replacement in flash-based DBMS.