基于PCM的分布式文件系统写缓存机制  

Write Cache Mechanism for Distributed File System Based on Phase Change Memory

作　　者：李帅[1] 蔡涛[1] 肖铁军[1] 陆游游[2] 

机构地区：[1]江苏大学计算机科学与通信工程学院,江苏镇江212013 [2]清华大学计算机科学与技术系,北京100084

出　　处：《小型微型计算机系统》2014年第2期217-221,共5页Journal of Chinese Computer Systems

基　　金：广东省自然科学基金项目(S2011010006118)资助;江苏省高校自然科学基金项目(09KJB520001)资助;江苏大学高级人才启动基金项目(09JDG038)资助

摘　　要：分布式文件系统在写操作结束前所需的元数据和数据写回需要大量的时间开销,现有的写缓存机制无法解决易丢失更新、使用寿命短和无法适应元数据访问特点等问题.提出基于PCM的写缓存机制,在主机中使用PCM构建写缓存,并针对分布式文件系统写操作和PCM的特性,设计了基于PCM缓存的写流程和自适应缓存写回策略,使得应用在将元数据和数据写入PCM缓存后就能安全返回,并使用自适应的写回线程池将PCM缓存中元数据和数据写回,提高了分布式文件系统的写速度,避免了元数据服务器和磁盘成为影响分布式文件系统I/O性能的瓶颈.在分布式文件系统pNFS的基础上,实现了原型系统,使用标准测试工具filebench和postmark和多个通用测试数据集进行了测试,验证了基于PCM的写缓存机制能使得pNFS提高7%-1倍的I/O性能和1倍以上的事务处理速度.Distributed file system needs to commit metadata and data before write operation is completed, it needs a lot of time con- sumption. The updates lost, short lifetime and inappropriate management strategy are the main problem in current caching mecha- nism. This paper proposes write cache mechanism based on Phase Change Memory and use Phase Change Memory to build write cache on the host. By analyzing the characters of write operation in distributed file system and Phase Change Memory, designing new write process for distributed file system and adaptive cache write back strategy. The application can return safely after writing metadata and data into cache, and they will be committed by adaptive writing thread pool. Which can improve the write performance of distrib- uted file system, and avoid metadata server and hard disk become the performance bottleneck of distributed file system. We realize the prototype on the pNFS, using filebench and postmark to evaluate. The results show that write cache mechanism based on Phase Change Memory can improve 7% -1 times I/O performance and more than one times transaction processing speed of distributed file system.

关 键 词：分布式文件系统 PCM 缓存 

分 类 号：TP316[自动化与计算机技术—计算机软件与理论]