Research And Implementation Of A File System Metadata Storage Technology Based On NVM

In recent years,the technology of the new storage medium NVM(Non-Volatile Memory)has continued to mature,which has the features of low access latency,nonvolatility,and byte-addressability.NVM is very likely to become the next generation mainstream storage medium.Users generally access storage medium through the file system.A file system suitable for NVM will be particularly important.Metadata is the key data of the file system.It records the inode data and dentry information of the file,and is usually managed by a key-value storage system.However,through testing,it is found that the performance of directly using existing key-value storage systems to store file system metadata in NVM is poor.Aiming at the problems of NVM-based file system metadata key-value storage,a new key-value storage system Meta DB(Metadata Database)based on NVM to store file system metadata is proposed.In view of the different characteristics of dentry data and inode data in the file system metadata,Meta DB separates the two parts of data and designs the DTable(Dentry Table)to store the dentry data and ITable(Inode Table)to store the inode data.According to the characteristics of dentry,DTable adopts a combination of a Hash structure based on the directory path and a B+Tree structure in the directory for storage.DTable can use the high performance of Hash to quickly locate the directory,and at the same time use the B+Tree to realize the range search in the directory.ITable uses a KV separated Hash structure for storage,which not only ensures the high performance of the Hash structure,but also reduces the write amplification during hash resizing.Finally,a file system HFS(Hybrid File System)based on Meta DB is designed.HFS uses Meta DB to store the metadata of the file system and merges the data of small files into the inode data,which can avoid space fragmentation and performance degradation caused by small files.And the data of dentry operations and inode operations are completely separated,without interfering with each other,and reduce the access overhead.The system of Meta DB and HFS are implemented,and tested on real NVM storage medium.At the key-value storage system level,the write performance of Meta DB is 1.1 to3 times that of Level DB,and the read performance is 6 to 10 times that of Level DB.At the file system level,the overall performance of HFS is increased by 7% to 30% compared to EXT4 and XFS,and 1.4 times to 2 times that of Table FS.