Access Characteristic Guided Partition for Nand Flash-Based High-Density SSDs

Abstract

NAND flash-based solid-state drives (SSDs) are a kind of widely adopted storage. However, state-of-the-art works presented that the SSD always suffers from significant read performance degradation. One of the most critical reasons is access interference between read and write operations. This is because the read and write latency gaps are more pronounced for the latest NAND flash in SSDs. In this paper, an interference reduction scheme is proposed to improve performance. This is motivated by the observation from several server workloads, where read and write operations can be easily separated based on access characteristics. Considering that SSDs are always organized with many parallel units, the basic idea of this work is to partition the parallel units of the SSD into different areas and place data in the corresponding area according to access characteristics. Then, the interference can be optimized by issuing read and write requests to the different areas. To realize the above design, several approaches are proposed: First, an access characteristic-based data placement and migration method is proposed for read and write request separation. Second, to further adapt the parallel requirement for different workloads, a workload-based partitioning scheme is proposed to determine the number of parallel units for read and write areas. Finally, based on partitioned SSD, a hot-data driven wear-leveling method is further proposed to balance the wearing of parallel units in read and write areas. Experimental results show that partitioned SSD can significantly improve the read performance and wear-leveling of partitioned SSD can guarantee performance and lifetime.