The Design And Implementation Of RDMA-Based Persistent Memory File System

In recent years,the non-volatile memory(NVM),which has the advantages of byte addressing,nanosecond reading and writing delay,power loss and no loss of data,is a high performance storage device.The academic has done a lot of innovative work for this new storage medium,and designed many new file systems.But these file systems are basically based on single host file system.Single host file system has limited storage capacity and is difficult to expand storage capacity because of the natural limitation of storage slot.It is not suitable for large data applications.For this high-performance storage medium,how to expand the storage capacity without losing much of its read and write performance is a problem to consider.In this paper,a multiple node nonvolatile file system RMFS is designed based on RDMA technology.RMFS can dynamically expand the NVM storage nodes through the network,and achieve high-speed reading and writing of file data through RDMA technology.To this end,the following study is carried out in this paper:(1)A flexible and fast extensible storage architecture design.RMFS file system adopts master-slave structure,Master node is responsible for meta-data and Slave node management function,and Slave node is responsible for data storage.Slave nodes can join Master node to RMFS file system at any time,and realize flexible and rapid expansion of storage nodes.(2)RDMA file data reading and writing process with low delay and high bandwidth.RMFS achieves efficient RDMA file data reading and writing operation based on Infiniband low delay network.RDMA file data reading and writing can reduce the utilization of CPU and improve the performance of RMFS.(3)The fast search mechanism of file data block index based on processor hardware MMU.In this paper,a fast search scheme for file data block index based on hardware MMU is designed.RMFS creates a virtual address space mapping for the file,and stores the storage node and NVM address of the file data block in the page table item.We can quickly locate the location of file data block through the virtual address of file.(4)An efficient NVM free space allocation management mechanism based on bitmap list.With the rapid expansion of Slave data storage nodes in the RMFS file system,the management pressure of NVM free space is increasing,and the storage overhead of management information is also increasing rapidly.This paper designs a new efficient free space management scheme for bitmap linked lists.A bitmap list is used to manage the allocation and use of the NVM storage space of the file system.At the same time,it has bitmap saving space and quick allocation function of linked list.By the predistribution mechanism of the continuous physical space,the distribution process is accelerated and the large file data block is stored.And the reading and writing performance of the large block data is improved.(5)The data backup function.For a multiple node file system,it is very important to provide normal file data reading and writing operation after a node failure.The RMFS file system uses a file data block as a unit for redundancy storage,and a file data block will have two Slave storage nodes to store,one of which is used as a main storage node of file data block,the other as a backup storage node of file data block.The nodes in the RMFS fault occurs,will run automatic data backup operation,ensure high availability of data in the file.Finally,aiming at the design goal of RMFS file system,a test experiment is designed.The experimental results show that the file data block index search based on MMU can reach 3 times the speed of the software search,and the reading and writing performance can reach 5387MB/s and 5321MB/s.After starting the data disaster recovery,the reading and writing performance can reach to 5355MB/s and 2180MB/s respectively.