On the use of NAND flash memory in high-performance relational databases

Abstract

High-density NAND flash storage has become relatively inexpensive due to the popularity of various consumer electronics. Recently, several manufacturers have released IDE-compatible NAND flash-based drives in sizes up to 64 GB at reasonable (sub-$1000) prices. Because flash is significantly more durable than mechanical hard drives and requires considerably less energy, there is some speculation that large data centers will adopt these devices. As database workloads make up a substantial fraction of the processing done by data centers, it is interesting to ask how switching to flash-based storage will affect the performance of database systems. We evaluate this question using IDE-based flash drives from two major manufacturers. We measure their read and write performance and find that flash has excellent random read performance, acceptable sequential read performance, and quite poor write performance compared to conventional IDE disks. We then consider how standard database algorithms are affected by these performance characteristics and find that the fast random read capability dramatically improves the performance of secondary indexes and index-based join algorithms. We next investigate using logstructured filesystems to mitigate the poor write performance of flash and find an 8.2x improvement in random write performance, but at the cost of a 3.7x decrease in random read performance. Finally, we study techniques for exploiting the inherent parallelism of multiple-chip flash devices, and we find that adaptive coding strategies can yield a 2x performance improvement over static ones. We conclude that in many cases flash disk performance is still worse than on traditional drives and that current flash technology may not yet be mature enough for widespread database adoption if performance is a dominant factor. Finally, we briefly speculate how this landscape may change based on expected performance of next-generation flash memories.