Highly Scalable NVM Storage System For Multi-core Computer System

NVM,a new type of byte addressable nonvolatile memory,has become an important means to solve the problem of storage wall in computer system because of its close to memory read-write speed,internal multi-channel,low latency,support for byte addressing and embedded processing capability.But the existing storage system I/O system software stack is designed for the slow read-write speed of Solid State Drive(SSD)and DISK,it is difficult to effectively play the advantages of NVM storage medium with high read-write speed,there are problems such as extra time and space overhead,low concurrency,and it is difficult to adapt to the requirements of new computing technology such as multi-core processor.This makes the software stack management of I/O system an important factor affecting the performance of NVM storage system.This paper takes NVM storage system in multi-core environment as the research object,and takes how to improve the concurrency of I/O software stack of NVM storage system as the goal to research and design a highly scalable NVM storage system for multi-core computer system.(1)Based on the analysis of the existing NVM storage system,aiming at the challenges of multi-core processor to the storage system,the paper designs the structure of the highly scalable NVM storage system for multi-core computer system,including the multi-core NVM storage device driver module and the high concurrent file system metadata management strategy module,so as to adapt to the characteristics of multi-core environment,give full play to the advantages of NVM storage medium read-write speed,and lay the foundation for improving the scalability and concurrent read-write efficiency of NVM storage system.(2)Based on the analysis of the existing NVM storage device drivers,this paper studies and designs a multi-core oriented NVM storage device driver for the characteristics of multi access request conflicts and high I/O speed in multi-core environment.The structure is given,and designs a storage space management module integrating partition and stripe,which simulates multiple parallel data access channels in NVM storage device,reduces the management granularity of storage space and improves the efficiency of executing access requests.The meta-operation based access request decomposition module is designed to decompose the received access request and completes the corresponding access request,thus reducing the locking granularity and locking times of the NVM storage device driver in executing access request,and increase the concurrency of executing access request.On the basis of Intel’s open source NVM storage device driver,the system MCNVMsim is implemented,which is tested and analyzed by Filebench and Fio testing tools respectively.The experimental results show that MCNVMsim can improve the IOPS value and I/O performance by 22.7% and 16.9% compared with PMEM.(3)The problems of current metadata management method in NVM storage system in multi-core environment are analyzed,and the high concurrency file system metadata management strategy is designed.The structure of the system is given.The original file system metadata management function is divided into dynamic classification directory tree module and metadata index based on multi-skiplists module.The dynamic hierarchical directory tree module divides the traditional directory tree into centralized layer and distributed layer.The centralized layer is stored in DRAM,which only contains the first N layers of directory trees,and uses the hash based method to manage the metadata.The distributed layer is stored in NVM,which contains the metadata of the rest layers,and uses the skiplist to manage it.Furthermore,the metadata cache fast replacement based on fixed window algorithm is designed,the fixed size window is used to count the frequency of metadata access in the distribution layer,and the metadata with high frequency access is selected to be transferred into the centralized layer cache in DRAM,so as to reduce the extra time and space cost of cache management and improve the speed of accessing the metadata in the distribution layer.The metadata index based on multi-skiplists module is designed.In the NVM storage device driver,multiple skiplists is used to manage the metadata stored in the distribution layer of NVM,so as to improves the concurrency of accessing metadata,reduce the access conflict rate,shorten the I/O software stack for reading and writing metadata,and improve the concurrent access performance of metadata in multi-core environment.The system HCFMS is built,and uses filebench to test the throughput under different number of files and threads,and Fio to tests the read-write performance.Compared with P-EXT4 and P-NOVA file system,the experimental results show that HCFMS can improve IOPS by 67.25% and 21.25%,and I/O performance by 41.25% and 24.52%.(4)Based on the source code of PMEM and EXT4,the prototype of highly scalable NVM storage system for multi-core computer system MSNSS is implemented,and the throughput of MSNSS were tested using various loads in the Filebench.The experimental results show that,compared with PM-EXT4 prototype system,MSNSS system can can improve IOPS by 75.86% and 37.08%.