Cooperative optimization for improving performance in NAND flash-based storage systems

Abstract

NAND flash memory has become the essential storage medium for various new-fashioned storage devices, such as smartphones, tablet PCs, Solid State Drives (SSDs), and boasts many useful features. However, NAND flash memory has some idiosyncrasies, such as out-of-place updates and bulk erase operations. It is a key issue to improve the performance of write operations in NAND flash-based storage systems, simply called flash storage systems. In addition to improving I/O performance, in mobile devices, energy saving is another major concern due to limited battery capacity. The performance and energy efficiency of flash storage systems are strongly influenced by the system software, such as operating systems and Flash Translation Layers (FTLs). This dissertation proposes three kinds of novel software-based performance and energy-efficiency improvement techniques for flash storage systems. First, there have been several studies on reducing the garbage collection overheads of FTL by rearranging write patterns in a sequential form in the operating system and device buffer. However, this one-sided approach cannot deliver completely sequential write patterns to FTL with random write requests, and FTL still suffers from large garbage collection overheads. If the operating system and device buffer can be made aware of the write patterns that FTL desires, the overheads can be alleviated even under random write patterns. This dissertation presents a system-wide cooperative optimization scheme, where the operating system and device buffer preferentially issue the write requests that can reduce the FTL overheads by inquiring into the state of FTL with the newly-defined cooperation interface. Second, the internal device cache of flash storage devices, such as SSDs, also has a significant impact on I/O performance. The device cache is used for two main purposes: to absorb frequent I/O requests and to store FTL mapping information. In the device cache, we observed that the optimal...