计算正迅速朝着高性能、大数据量的方向发展，许多大规模科学和工程应用依赖多个分布在不同地理位置、拥有不同存储介质、容量规模各不相同的异构存储系统协作构成的存储池提供足够的数据存储支持。广域分布式存储系统GDSS-Alpha（Global Distributed Storage System Alpha）系统正是在这种背景下诞生的。
在分布式存储系统中，元数据及其管理处于相当重要的地位。GDSS-Alpha系统的元数据管理支持元数据的准确定位和完整表达，并通过建立缓存和副本机制来保证系统的性能和可靠性。
GDSS-Alpha元数据缓存模型采用弱一致性管理策略，并提出一种基于H值的元数据缓存替换算法，该模型的实现主要由对象传输器、替换评估器和缓存管理器三个模块构成。元数据缓存模型提高了元数据访问效率，并减轻了元数据服务器的负载。在元数据缓存基础上，系统采用复制技术创建和维护多个元数据副本，进一步提高了元数据的访问性能，并增强了系统的可靠性。为了能够准确地把握元数据副本的创建时机，系统提出基于访问频度阈值的副本创建策略，根据用户访问特征放置元数据副本，利用树结构进行元数据服务器的动态管理。
元数据管理系统采用JAVA语言编写，具有平台无关特性。测试结果表明，缓存加副本的元数据管理方案能够有效地提高用户的访问效率，元数据响应时间比原来减少了76%，减轻了元数据服务器的负载，并增加了元数据服务器的可靠性。
关键字： 分布式存储，元数据缓存，弱一致性，元数据副本
With the development of high performance computing, more and more large-scale scientific and engineering applications produce huge amount of data that can not be stored in a single storage system. To meet the storage demands of these applications, Global Distributed Storage System Alpha (GDSS-Alpha) is developing to group heterogeneous storage systems which are distributed in different places into a cooperative and virtual storage pool to provide huge storage capability.
To achieve this goal, metadata management in GDSS-Alpha is one of the key problems which need comprehensive exploration. Cache and replication methods are proposed to provide powerful, flexible, reliable and scalable metadata management mechanism.
Metadata cache, which is located in the in storage service point (SSP), is utilized to improve metadata accessing performance. Due to the performance consideration and practical demand, weak consistency model is adopted to maintain the metadata consistency. A H-value based cache replacement algorithm is designed to improve cache hitting rate to ensure certain performance gain. In practical system, metadata cache management consists of three modules, object transporter, replacement evaluator and cache manager. Replication is exploited to eliminate performance bottleneck and single failure point caused by single metadata server. Multiple metadata servers hosting metadata replicas construct a dynamic tree structure. The replica can be created on demand based on access frequency and distributed along the path in the former tree structure.
A practical metadata management system, which is platform independent because of being written in Java, is in the b testing phase. The performance test shows that caching and replicating methods adopted in this system are very efficient in reducing the system responding time, balancing the loads among different servers, and improve the system reliability.
Keywords: Distributed storage, Metadata cache, Weak consistency, Metadata replica
1 绪论
2 GDSS-Alpha元数据管理
设计文档搜索：
