SCMPFS:面向SCM的聚合文件系统

发布时间：2018-06-18 02:04

  本文选题：存储级内存 + 非易失存储器　； 参考：《计算机研究与发展》2015年S2期

【摘要】：由于NVM器件集成度和DIMM接口数量的限制,大容量SCM系统中必然包含DIMM接口和PCIe接口NVM存储设备.这2类NVM存储设备在访问接口、访问方式、I/O性能、存储容量和内嵌处理能力等方面存在很大差异,这使得现有的内存和外存均难以直接应用与统一管理2类NVM存储设备;同时不同文件、文件数据与元数据、文件内各部分数据之间的访问特性和管理要求也不相同,这给高效管理大容量SCM系统带来了极大的挑战.提出了面向SCM的聚合文件系统SCMPFS,统一管理2类NVM存储设备.首先给出了聚合文件系统的结构;接着给出了元数据分层管理算法,具有元数据属性和NVM存储设备的特性,重构元数据的组织结构,将元数据属性合理分布到2类NVM存储设备中,提高元数据查找速度和管理效率;设计了数据聚合管理算法,通过打破文件数据的单一组织结构、通过文件数据在2类NVM存储设备之间的合理分布,提高访问文件的响应速度和文件系统的I/O性能,有效利用2类NVM存储设备中的存储容量;最后使用ramfs和PMBD分别模拟DIMM接口和PCIe接口NVM存储设备,在FUSE的基础上实现了SCMPFS的原型,使用postmark和Linux内核源码进行了测试与分析,验证了SCMPFS相比PMBD上Ext3能提高30%~63%的I/O性能,接近与PMFS,同时避免DIMM接口NVM存储设备容量的限制;此外测试也表明软件和接口开销成为影响大容量SCM系统性能的重要因素之一.
[Abstract]:Due to the limitation of the integration of NVM devices and the number of DIMM interfaces, the NVM storage devices with DIMM interface and PCIe interface must be included in the large capacity SCM system. These two kinds of NVM storage devices have great differences in accessing interface, accessing mode and I / O performance, storage capacity and embedded processing ability, which makes it difficult for existing memory and external memory to directly apply and uniformly manage two classes of NVM storage devices. At the same time, the access characteristics and management requirements of different files, file data and metadata as well as the data in each part of the file are different, which brings great challenges to the efficient management of large capacity SCM systems. SCMPFS, an aggregate file system for SCM, is proposed, which manages two classes of NVM storage devices uniformly. Firstly, the structure of the aggregate file system is given, and then the hierarchical management algorithm of metadata is given, which has the properties of metadata and the characteristics of NVM storage device, and reconstructs the organization structure of metadata. In order to improve the speed of metadata lookup and management efficiency, a data aggregation management algorithm is designed to break the single organization structure of file data. Through the reasonable distribution of file data among two NVM storage devices, the response speed of accessing files and the I / O performance of file system are improved, and the storage capacity of two NVM storage devices is utilized effectively. Finally, ramfs and PMBD are used to simulate the DIMM interface and PCIe interface NVM storage device respectively. The prototype of SCMPFS is implemented on the basis of fuse. The postmark and postmark kernel source code is used to test and analyze. It is verified that the I / O performance of SCMPFs can be increased by 30% compared with that on PMBD. It is close to PMFSs and avoids the limitation of storage capacity of DIMM interface NVM storage devices. Besides, the test also shows that software and interface overhead become one of the important factors that affect the performance of large capacity SCM systems.
【作者单位】： 江苏大学计算机科学与通信工程学院;
【基金】：国家自然科学基金项目(61300228) 江苏省自然科学基金项目(BK20140570) 江苏省科技支撑项目(BE2013103) 浙江省自然科学基金项目(LY13F020012) 深圳市科技计划项目(JCYJ20130401095947222)
【分类号】：TP333
，

本文编号：2033516