Lustre文件系统元数据服务恢复机制研究  

Research on Metadata Service Recovery Mechanism for Luster File System

作　　者：钱迎进[1] 伊瑞海[1] 肖侬[2] 金士尧[2] 

机构地区：[1]中国卫星海上测控部,江阴214431 [2]国防科学技术大学并行与分布式处理国家重点实验室,长沙410073

出　　处：《高性能计算技术》2013年第6期10-16,共7页

基　　金：本文获得国家自然科学基金（60736013）资助.

摘　　要：摘要在分布式文件系统中，性能和可恢复性是两个相互冲突的需求。为了能够向用户提供高可用的连续数据服务，它要求系统在检测各种错误和失效时能够自动进行恢复。Lustre是HPC业界中领先的一个开源并行文件系统。它在设计上着重考虑了性能，采用了较为复杂的基于事务的有状态分布式恢复协议，可支持多种失效情况的透明恢复，并提供透明的故障接管（failover）功能。本文着重对Lustre元数据服务的恢复机制进行了研究，为分布式文件系统恢复机制的设计提供了一个范例。In the distributed file system, there are two conflict requirements between performance and recoverability. To provide users with high available and continuous data service, it needs that the system can recover automatically when detect various of wrongs and failures. Lustre is a leading parallel file system in HPC industry. It mainly considers the performance during design and uses complex transaction-based stateful distributed recovery protocol to support recovery for most of failures and provide failover functionality transparently. This paper focuses on the research on recovery mechanism of Lustre metadata service and provides a good paradigm for designing recovery mechanism of distributed file systems.

关 键 词：LUSTRE 高性能计算 分布式文件系统 事务 元数据更新 恢复 高可用性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]