Research on Very Large Secondary Storage System based on Multiple Disk-Array and its Functional Management Mechanism

在本研究中，我们考虑了一个由多个磁盘阵列通过适当的网络连接而成的超大型二级存储系统，即阵列式磁盘阵列系统，在第一年，我们已经提出了新的磁盘阵列体系结构，如虚拟条带化，热块集群和热镜像，并基于这些技术开发了一个良好的磁盘阵列模拟器。由于阵列磁盘阵列这种需要超大容量存储系统的应用需要处理大量的数据，因此我们将大事务处理作为阵列磁盘阵列的目标应用，在商用关系数据库上检索运行在TPC-C上的跟踪数据，并通过分析跟踪数据来评估系统的性能。本文以大型数据仓库为目标应用，检索了运行在商业关系数据库上的TPC-D跟踪数据，并利用这些跟踪数据分析了超大型数据仓库的访问局部性。综合考虑各种索引机制，特别是多列索引及其元数据管理，提出了一种基于应用估计的数据特征的自适应性能调优机制。提出了利用元数据信息对阵列磁盘进行预取的机制。通过分析大型数据库应用程序，如事务应用程序（TPC-C）和数据仓库应用程序（TPC-D），这些方法的性能进行了评估。因此，我们展示了我们方法的效率。

In this research, we consider a very large secondary storage system configuration which consists of multiple disk arrays connected by appropriate network, that is, an arrayed disk-array system.At the first year, We have already proposed novel disk array architecture, such as virtual striping, hot block clustering and hot mirroring and developed a fine disk array simulator based on these techniques. Since the applications which require a very large secondary storage system, the arrayed disk-array, process a large amount of data, we took large transaction processing as a target application of our arrayed disk-array and retrieve trace data of TPC-C running on commercial RDB.Then, we estimate the performance of our system analyzing the trace data.At the second year, we took large data houses as a target application of our system and retrieve trace data of TPC-D running on commercial RDB.We analyzed the access locality of very large dataware houses by using the above trace data. Considering various index mechanisms, especially multicolumn index and its metadata management, we proposed the self-contained performance tuning mechanism based on data characteristics which estimated by applications. Moreover, the prefetch mechanism on the arryaed disk-array by utilizing metadata information was proposed. The performance evaluation of these proposed method was done by analyzing large database applications, such as transaction application (TPC-C) and dataware house application (TPC-D). Consequently, we showed the efficiency of our methods.