Flash-enabled database storage

Abstract

Flash memory has emerged as a high-performing and viable storage alternative to magnetic disks for data-intensive applications. In our work we study how the storage layer of a database system can benefit from the presence of a flash disk. Due to the varying price and I/O characteristics of flash disks the optimal design decisions vary widely across different setups. We study how the system can take advantage of the random read efficiency of inexpensive flash disks by using the latter at the same level of memory hierarchy as magnetic disks in a hybrid setup; our algorithms provide efficient and adaptive data placement that leads to substantial performance improvement. We propose techniques to accurately predict the main memory cache behavior for systems consisting of heterogeneous storage media and selectively allocate memory buffers to devices; thereby, the I/O cost of the system drops significantly, even offsetting wrong data placement decisions. We also explore the design space for a system that uses flash memory as a cache to the underlying storage and propose techniques towards high performance. The experimental results, we believe, exhibit both the potential and necessity of our techniques in future database systems.