Reinforcement Learning based Background Segment Cleaning for Log-structured File System on Mobile Devices

Chao Wu, Cheng Ji, Chun Jason Xue

Abstract: With the adoption of Log-structured file system in mobile devices, the impact of background segment cleaning on system performance and storage lifetime becomes notable. Aggressive background segment cleaning solution generates excessive block migrations and impairs the endurance of NAND storage device, while a lazy solution cannot reclaim enough segments for subsequent I/O requests thus leading to the occurrence of foreground segment cleaning and prolonging I/O latency. In this paper, a reinforcement learning based approach is proposed to balance the trade-off. Through learning the behaviors of I/O workloads and the statuses of logical address space, the proposed approach can adaptively reduce the frequency of foreground segment cleaning by 68.57% on average, and decrease the number of block migrations by 71.10% over existing approaches.