Research On Write Optimizations For File Systems Based On Persistent Memory

With the rapid development of the Internet of things,mobile edge computing,and big data,the amount of data that file systems need to process increases drastically.However,the traditional block-based file systems cannot meet the demands of big data storage,such as high performance,high reliability,and scalability.Fortunately,the emerging persistent memories(PMs)bring new opportunities for big data storage since PM shows the advantages of nonvolatility,byte-addressability,low latency,and high density.Nevertheless,PMs generally have limited write endurance,which proposes new challenges to the design of file systems.Therefore,facing the challenges and opportunities of PMs,this thesis focuses on optimizing the writes in file systems.Firstly,a virtual superpage mechanism has been proposed to organize file data to reduce the overheads of write,locating file data,and space management for persistent memory file systems.Secondly,a wear-leveling mechanism for inode management of persistent memory file systems has been proposed to achieve high-precision wear-leveling with low overhead.Finally,a PM-based write back mechanism for the hybrid storage file system has been proposed to improve file system performance.The key contributions of this thesis are summarized below.(1)We propose a Virtual Superpage Mechanism(VSM)to reduce the write amplification and improve the space utilization of persistent memory file systems using superpages.VSM employs the Multi-Grained Copy-on-Write(MCo W)mechanism to reduce the over-write amplification by modifying the file data index structure and page table.Moreover,to improve the space utilization and reduce the cost of data migration,we propose Zero-copy File Data Migration(ZFDM)mechanisms to avoid the copy cost of data migration by modifying the page table.(2)The inodes in existing persistent memory file systems are frequently updated in fine granularity and stored in fixed locations throughout their lifetime.However,existing persistent memory file systems do not consider the damage of inodes to the write endurance of PM,which results in fast wear out of inode table.In this thesis,we present a Lightweight and Multi-grained Wear-leveling Mechanism(LMWM)to achieve wear-leveling of the inode table in PM.LMWM adopts 64bytes-sized and page-sized(e.g.,4KB)data migration granularity for alleviating the overhead of data migration and improving wear accuracy.(3)The block-based file systems usually employ the DRAM-based buffer caching mechanism to improve performance.Unfortunately,this scheme can lead to data loss when power failure or system crash occurs.To this end,the synchronous operations are provided to write the cache data back to the block device immediately,which seriously reduces the performance.Therefore,we propose a hybrid storage file system architecture based on small-capacity PM to solve this problem.Moreover,an efficient PM assisted Write-back Mechanism(PMW)has been proposed to alleviate the overhead of synchronization operations for improving the performance of the file system.We implement the proposed techniques in the Linux kernel and evaluate them on real platforms.The experimental results show that the proposed VSM can effectively improve the performance of persistent memory file systems.The proposed LMWM can achieve low-cost wear-leveling of the inode table and improve the lifetime of the underlying PM.Moreover,the proposed PMW can effectively reduce I/O operations and improve the performance of file systems.