Research On Key Technologies In Flash-based DBMS

Flash memory is an electrically erasable programmable read-only memory. Theflash has dominated the storage on mobile devices such as cell phones, MP3andsensors because of its small size, light weight and fast speed. With the incrementof capacity, decrement of price, flash-based storage devices are increasingly used inenterprise large-scale data storage systems, including database system. Flash mem-ory has much higher access speed than disk, especially on its random access speed,so the adoption of flash memory can signifcantly alleviate the storage bottleneckin database system.Compared with the conventional magnetic disk, flash memory has its own char-acteristics. As an electronic device, a flash memory does not need mechanical seeklatency, so the random access speed is similar to sequential access. In addition, dueto the erase-before-overwrite limitation, the read speed of flash memory is severaltimes faster than write speed, which is known as asymmetric I/O feature. However,the current database management system optimizes its performance according todisk characteristics, so it is suboptimal to deploy the database systems directly to flash memory. Consequently it is necessary to adjust the database due to the featureof flash memory.The optimization on flash memory usually needs to avoid write operation, par-ticularly the random write. For example, flash-based index structure usually usedthe log structure to transfer the random updates to sequential write. In this case,how to design a proper log organization is an important issue. Similar problemexists in bufer management. How to reduce the write operation to flash and im-prove the overall I/O performance is the key issue for flash-based bufer design. Inaddition, the current flash disk hybrid storage system design is paid more and moreattention. This paper focuses on the optimization on these key technologies in flashmemory database management system, include the following aspects.(1) We propose a flash-based operation-aware bufer management strategy,named FOR. Based on the operation-aware cost model for flash memory, in FORstrategy, both the state of a page and the operation on this page are taken into con-sideration. Compared with the approaches that only consider page state, FOR canachieve better performance for flash-based bufer management. In this paper, an ap-proximation of FOR, FOR+algorithm is also proposed to accelerate the executionof the FOR method.(2) We propose a flash-based high-dimensional indexing structure named LCRtree. LCR tree designs a special log structure for the traditional high dimensionalindex. The log area can transfer the random updates into sequential ones whichare more suitable for flash write. Furthermore, LCR tree organizes the logs incompressed log manner; so that compared with traditional methods, LCR tree canreduce the number of read operations on the logs and achieve better performance.(3) We propose approaches for hybrid storage system in this paper, that is,probability-based hybrid storage strategy and hot-aware strategy respectively. Inprobability-based approach, two probability parameters are incurred to control thedata migration which can reduce unnecessary data movement between system mem-ory hierarchies, so that the overall performance of the hybrid storage method isenhanced. The hot-aware approach identifes the storage position of a page based on the hotness of the page. A page with higher hotness will be placed in highermemory hierarchy. Compared with the existing methods, this method can achievebetter performance in a lower cost.