An In-depth Analysis on File System Performance for NVM-based Storage System

• Main Authors: Pei Yin Eng, 翁佩銀
• Other Authors: 楊佳玲
• Format: Others
• Language: en_US
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/25655790247038625665

碩士 === 國立臺灣大學 === 資訊工程學研究所 === 105 === Nowadays, the access latency of byte-addressable NVM decreases and the access latency to the traditional design of software layers becomes an overhead. Thus, byte-addressable NVM technologies require significant changes to the way systems interact with storage devices. In this paper, we analyze the legacy storage stack, focusing on the file systems designed for the NVM, aiming to determine the potential issues of the file systems designed for the NVM-based storage system. We target on NVM in DIMMs which is attached to the processor’s memory bus in a manner similar to DRAM and select some NVM-aware file systems due to their unique designs for NVM as well as due to their direct access (DAX) support, which bypasses DRAM page cache. The selected file systems are NOVA [12], a file system that highly optimized for NVM, and Ext4- DAX, a widely used and stable block-based file system that adjusts operating system for NVM, and SIMFS [11] a new file system that embeds address space of a file into process address space, as our studied NVM-aware file systems. We also focus on four features of the file system, which are metadata (inode) indexing, file indexing, NVM addressing and consistency mechanism. We categorize our analysis into three sections. First, we analyze the operations of a single I/O request of the NVM-aware file systems and their corresponding overhead in each of the selected file system. Next, we study the sequential and random I/O requests in a single thread to find out the difference in performance between single I/O request and multiple I/O requests among the NVM-aware file systems, as well as the performance in multi-threads to study the performance of concurrency support of the file system. Finally, we study the performance of NVM- aware file system in workload scenario using four workloads from Filebench as our evaluated macro-benchmark. In our analysis, we find some design concerns that needed to be reconsidered when designing NVM-aware and NVM-optimized file system. We suggest that the design of NVM-aware file system should avoid the overhead of repeatedly performing file indexing, NVM addressing and data access in large data access by using a larger granularity of data when performing these actions. Besides, the legacy driver-based NVM address calculation and the legacy address translation are found unnecessary for the NVM and suggestions are provided. We also provide some suggestions for consistency and concurrency support for NVM-aware file systems.