On Intelligent Storage Stack

Abstract

NAND flash storage devices are dominating the storage market. With faster accessing speed, higher shock resistance, and lower cost per unit, NAND flash devices are adopted in personal computers, mobile devices, cloud storage, etc. However, traditional I/O stack designed for hard-disk drives (HDD) cannot adapt to the features of NAND flash storage, which might become the bottleneck of system performance. For example, current I/O merging operation in the I/O scheduler layer is proposed for the maximization of the system I/O throughput on HDDs, which could impact the I/O performance on SSDs. In addition, the I/O stack on mobile devices is inherited from traditional computers, which cannot adapt to the features of mobile platforms. In the file system layer, the segment cleaning in the log-structured file system, e.g., F2FS, might either reclaim insufficient logical space for subsequent I/O or generate excessive page migrations and impair the lifespan of the built-in storage device on mobile devices. In the storage device layer, the build-in SRAM management cannot distinguish the FG I/O from BG I/O, which leads to the prolonged FG I/O latency and impacts user experience on mobile devices. These factors motivate us to revisit the I/O stack design on both SSDs and mobile storage.

This thesis focuses on the intelligent I/O stack optimization for SSDs and mobile storage. It consists of three parts: 1) an adaptive I/O merging approach in the I/O scheduler layer; 2) a multi-objective segment cleaning approach in the file system layer; 3) a dynamic cache management approach in the storage device layer. The first two approaches are designed for the host OS I/O, and the third approach is designed for the UFS-based mobile storage.

In the first part, the observation that the current I/O merging scheme could incur sluggish I/O performance is obtained through experiments. By studying the impact of I/O merging on various I/O patterns, a reinforcement learning (RL)-based approach is proposed to make merging decisions adaptive to different I/O patterns. In this manner, the system can deliver an optimal system I/O throughput with minimized I/O performance variation. In the second part, the drawback of the current background segment cleaning scheme in the log-structured file system (LFS) on mobile devices is discussed. Through learning the impact of the triggering frequency of background segment cleaning on the QoS performance and storage lifetime, a pruned multi-objective deep reinforcement learning-assisted approach (MOBC) is proposed. MOBC triggers background segment cleaning adaptive to I/O workload patterns and logical space status, thus managing to reach Pareto-optimal QoS performance and storage lifetime on mobile devices. In addition, a novel Multi-path Multi-layer Perceptron (MMLP)-based structured pruning approach is proposed to save the time and space cost of MOBC, thus making it more efficient on mobile devices. In the third part, the SRAM-contention issue is addressed on the UFS-based mobile storage. Experimental results show that background I/O could rapidly occupy the in-device SRAM resource in mobile storage. This SRAM-contention between background and foreground I/O could degrade the foreground I/O performance and impair user experience significantly. Prior works neither consider the foreground/background information in SRAM management nor consider the UFS-specific features. As a result, they fail to address the SRAM-contention issue. Motivated by these observations, a novel cache management approach is proposed. Through comprehensively considering the foreground/background information and UFS features, the proposed approach can mitigate the contention in the SRAM between foreground and background I/O, meanwhile handle foreground I/O with maximized parallelism, thus improving foreground I/O performance significantly. In addition, a cache relocation scheme is proposed to boost the foreground I/O with a high cache hit rate, thus improve foreground I/O performance further. In this manner, the SRAM-contention issue is well addressed.

On the basis of these techniques, the system I/O performance and storage lifetime are both expected to be improved significantly for SSDs and mobile storage devices.

Date of Award	20 Aug 2020
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Chun Jason XUE (Supervisor)