Research On File System For Non-Volatile Memory

Storage has always been the bottleneck of the entire computer system due to its poor performance compared with computing resources.This problem gets even worse in the era of big data,the disk-based storage system with slow performance improvement can hardly meet the requirement of big data applications to access data quickly.Phase Change Memory(PCM),Spin-Transfer Torque Random Access Memory(STT-RAM)and Resistive Random Access Memory(RRAM)as the main representatives of the next generation of new Non-Volatile Memory(NVM),have some superior characteristics such as low access Latency,near-zero static energy consumption,high density and byte-addressability etc.are considered as the candidate to replace disk to improve the access performance of storage system.However,NVM also suffers from asymmetric read-write performance and limited write counts.In addition,as the main system software in the storage system,most of the existing file systems are designed according to the characteristics of disks,they do not consider the low access latency characteristics of NVM,and nor support fine-grained byte-level space management,even more lack the corresponding management strategies for NVM features.Therefore,it is necessary to design a dedicated NVM-based file system to improve the access performance,space management efficiency and extend the lifetime of NVM.To reduce the overhead of traditional disk-based storage systems and existing NVMbased file systems,this thesis designs a NVM-based lightweight persistent file system to simplify the traditional disk-based file system I/O stack,reduces the software overhead in the I/O path and improve the access performance of NVM storage system.The analysis reveals that replacing the disk directly with NVM,the software overhead of the traditional disk-based file systems will become the new bottleneck of the storage system.On the other hand,compared with the low access latency of NVM,the software overhead of virtual file system in existing NVM-based file systems still too large.To address these problems,a NVMLPFS library is designed and implemented according to the characteristics of NVM.It directly routes the standard I/O requests to file system,which shortens the I/O path and reduces the software overhead of complicated I/O stack.NVMLPFS also exploits the byteaddressability of NVM to optimize the undo logging technology and reduces the overhead of traditional log operations,and ensures the consistency during file system updates.Experimental results show that compared with F2FS-SSD,Ext4-RamDisk,RAMFS,PMFS and NOVA,NVMLPFS significantly improves the access performance of NVM storage system.To reduce the internal fragmentation of NVM-based file systems,this thesis proposes a page-based fragmentation data merging(PFDM)method for internal fragmentation management of NVM file systems.PFDM uses shared pages to merge fragmented data with less than one page of files,to reduce internal fragmentation of the NVM file system and to increase the spatial utilization of the NVM file system.The analysis finds that file systems typically use blocks or pages with fixed-size as the basic unit of space management,which will generate a large amount of internal fragmentation in the expensive,space-constrained NVM device and waste the valuable free space of NVM.PFDM utilizes the byte-addressability of NVM to store multiple file data or the tail data of file(referred as file fragmentation data)in the same page,which significantly reduces the file system internal fragmentation.PFDM also designs the efficiently fragmentation data table(FDT)and shared page table(SPT)respectively to manage the address space of file fragmentation data and the free space of the shared pages,so as to reduce the management overhead of the file fragmentation Data.Experimental results show that,compared with Ext3,Ext4 and original PMFS file systems,the PMFS file system integrated with the PFDM method significantly increases the space utilization of the NVM storage system with low storage overhead.To reduce the extra overhead in existing wear-leveling methods,this thesis proposes a file-aware wear-leveling(FAWL)method for PCM storage system to optimize the layout of file data in the device,reduce the write counts and extend the lifetime of PCM.The analysis of existing NVM-based wear leveling including flash memory and PCM reveals that the device-based wear leveling suffer from extra wear overheads and low efficiency,whereas the host-based wear leveling only simply utilize partial file system information,and does not perform wear leveling operations on the blocks occupied by cold data.FAWL designs the hot and cold file identification method and the PCM page management method by using the file attributes(file type,file access authority and file access frequency)and the PCM page write information respectively.With the hotness of file and the wear information of page,FAWL allocates the young pages to frequently updated file data,middle-aged pages to moderately updated file data and old pages to infrequently updated file data.FAWL also implements a correction mechanism which combines dynamic and static methods to distribute the file data to suitable PCM pages to prevent the infrequently updated file data to occupy the young pages of PCM in a long time due to the change of file hotness.Experimental results show that,compared with popular wear-leveling methods including random-swapping,start-gap and segment-swapping,FAWL wears more evenly and further extends the lifetime of PCM.