Scheduling flash request in SSD to improve response time

Abstract

Modern storage devices are becoming more complex and smarter compared to the past. As a consequence the central role of IO scheduling is shifting from operating systems to storage devices. Many scheduling algorithms for magnetic disk drives have been proposed, but for NAND-flash-based solid state drives (SSD), dissimilar approaches are necessary due to different physical characteristics. SSD has no moving parts and has multiple flash chips operating in parallel, where IO requests are processed in a unit of page. Its write latency is tens of times as great as the read latency, and SSD requires erase operation to overwrite in a memory cell. Considering the characteristics, scheduling algorithms, such as earliest finish time first (EFTF) algorithm and simple intuitive ones, are explored and their implementations are provided. By employing a simulation-based method using various parameters, the performance of the algorithms are analyzed. The simulation result shows that utilizing the EFTF algorithm improves the response time of SSD by average 16.85\% and maximum 34.63\% depending on the workloads intensity compared to first-in-first-out policy. Throughout our work we also have verified that small cache hit rate, large scheduling queue size, and high intensity of IO contribute to the effectiveness of the scheduling algorithms. In addition, by employing time-out strategy to prevent starvation, we have shown that there is a tradeoff between maximum response time and average response time.