| Budget Amount *help |
¥1,400,000 (Direct Cost: ¥1,400,000)
Fiscal Year 2001: ¥1,400,000 (Direct Cost: ¥1,400,000)
|
|---|
| Research Abstract |
Storage-centric configurations in which magnetic hard disks are directly connected to a network have recently attracted considerable attention to implement effective and high performance large-scale data storage. However, centralized control for managing data placement and failure handling has problems in its performance and reliability. If each disk can automatically manage those matters, management cost will be significantly reduced, and high scalability will be guaranteed.
This research studied the following items to realize the autonomous control for data placement and failure handling.
・ We proved that the cluster configuration of autonomous disks adopting distributed directory, rule description, and stream interface realized high flexibility. We developed an experimental autonomous cluster with network connected PCs to demonstrated the effect of rule description. The experimental system also showed that the performance of the autonomous disks are sufficient compared with current NF
… More
S implementation. We then proposed a method for modifying configuration of a cluster dynamically, and evaluated the effect of the reconfiguration mechanism. We proved that the cluster configuration of autonomous disks adopting distributed directory, rule description, and stream interface realized high flexibility. We developed an experimental autonomous cluster with network connected PCs to demonstrated the effect of rule description. The experimental system also showed that the performance of the autonomous disks are sufficient compared with current NFS implementation. We then proposed a method for modifying configuration of a cluster dynamically, and evaluated the effect of the reconfiguration mechanism.
・ As the distributed directory in autonomous disks, we adopted Fat-Btree structure that we had proposed previously. The research studied the treatment of range queries on it, its concurrency control, and its recovery method, and demonstrated the feasibility of Fat-Btree for the distributed directory.
・ It is very important to balance load among components of a cluster. Fat-Btree has a functionality of balancing load and amount of data. However, we had to consider efficient load evaluation methods and control mechanism of handling load skews. We proposed a traverse-cost based load-evaluation method and parallel skew management mechanism using a process tree structure. We then show the effectiveness of those method by simulation.
・ We also considered applications on a autonomous disk cluster. We proposed a method to store XML into distributed disks, a method to implement Web server using the autonomous disks, and a disk retry algorithm for treating continuous media data by combining parity calculation. We then evaluated those method arid algorithm and demonstrate there effectiveness. Less
|
