Research And Implementation Of Consistency Mechanism Of In-memory File System Based On RDAM And NVM

With the increasing scale of data in the Internet,people have put forward more requirements for the performance of the storage system in the data center,especially in data storage and network transmission.NVM and RDMA bring new opportunities for the design of the storage systems in data center.Due to the design of internal cache in modern processors and the lack of guarantee for RDMA in NVM persistently,there is still several problem of data consistency when using persistent memory.This brings challenges to the design of consistency mechanism based on RDMA and NVM.After investigating and testing the way RDMA persistence writes to NVM,this paper summarizes the persistence writing methods used in different scenarios.Combined with these characteristics,an efficient data consistency update mechanism which can ensure the consistency of the file system is designed.The main work of this paper is as follows:(1)this paper investigates the persistence of writing to NVM by RDMA,which is RDMA read after writingand RPC,respectively.After experimental analysis,it is found that in the case of small granularity writing,the advantage of RDMA primitive is no longer obvious,and the delay of RPC is lower than that of RDMA read after writing.In the case of large granularity,the advantages of RDMA primitives still exist.Therefore,in the scenario of small granularity writing and low latency,RPC should be chosen.(2)based on the analysis of the sequence of RDMA writes,it is found that after using send/recv or write,RDMA itself can guarantee the sequence between the two operations.RPC is implemented in send/recv,so the sequence of operations can be guaranteed between RPC and RDMA read after write.In terms of atomic operation,RPC has better performance than RDMA's atomic write operation.Therefore,in the design of the consistency mechanism,RPC or RDMA should be used to write and read after writing to ensure the sequence of operations,and RPC should be used as a way to modify atomic operations.(3)A data consistency update mechanism is designed,in which a log management module is added to the prototype file system to deal with the persistent submission of logs,and a log buffer is added to the storage structure to record logs.A backup buffer is used to store the written data.By separating log buffer and backup buffer,logs can be lightweight so that logs can be recorded as RPC,and data can be written as RPC and RDMA read after write.(4)the two-stage update strategy and ring counter of the log are designed.The twophase update strategy of the log is divided into data update and metadata update.In the aspect of data update,the write mode of RPC or RDMA is selected according to the write size,the consistency mechanism of redo log and copy-on-write is selected according to the write size,redo log is selected in terms of metadata,and asynchronous persistence of metadata is adopted to improve performance.In the log buffer,according to the characteristics of NVM,efficient ring log buffer and ring counter are designed.When modifying the counter,we use the way of RPC to update the counter on the server side.After testing,the consistency update mechanism designed in this paper can not only ensure the consistency of metadata and data,but also achieve efficient data update performance.