Boosting random write performance for enterprise flash storage systems

Abstract

NAND flash memory has been successfully employed in mobile devices like PDAs and laptops. With recent advances in capacity, bandwidth, and durability, NAND flash memory based Solid State Disk (SSD) is starting to replace hard disk drive (HDD) in desktop systems. Integrating SSD into enterprise storage systems, however, is much more challenging. One of the major challenges is that server applications normally demand an exceptional random I/O performance, whereas current SSD performs poorly in random writes. To fundamentally boost random write performance, in this paper we propose a new write cache management scheme called EPO (element-level parallel optimization), which reorders write requests so that element-level parallelism within SSD can be effectively exploited. We evaluate EPO using a validated disk simulator with realistic server-class traces. Experimental results show that EPO noticeably outperforms traditional LRU algorithm and a state-of-the-art flash write buffer management scheme BPLRU (block padding least recently used).