NAKANO Miyuki Univ. of Tokyo, Institute of In-Research Associate dustrial Science, 生産技術研究所, 助手 (30227863)
KANEKO Storu Fuijitsu Ltd., File system Director Division, 富士通・ファイルシステム事業部, 部長
TAKAGI Mikio Univ. of Tokyo, Institute of In-Professor dustrial Science, 生産技術研究所, 教授 (30013113)
|Budget Amount *help
¥11,000,000 (Direct Cost : ¥11,000,000)
Fiscal Year 1996 : ¥6,900,000 (Direct Cost : ¥6,900,000)
Fiscal Year 1995 : ¥4,100,000 (Direct Cost : ¥4,100,000)
In this research, we propose the high performance disk array, which is expected for information sever supporting the multi-media information infrastructure. The disk array architecture has large potential which can achieve not only high reliability by redundant derives but also high performance by adopting parallel access to multi drives. However, there is a problem on commercial disk array whose performance is limited by the penalty of block update in comparison with the performance of conventional disk systems. That is, we cannot make the best use of the potential of disk array.
At the first year of this research, we proposed a new volume management method of disk array called "hot mirroring". This method combines the feature of RAID-1, in which the response time is fast and the volume capacity is low, and RAID-5, in which the response time is low and the volume capacity is high, by considering the access locality. The feature of the proposed method are the following : introducing the
hierarchical array structure into the one disk derive, making the virtual micro derive by collecting the rational frequently accessed block in physically local area with hot block clustering, which has already been proposed in our lab, and, in addition, guaranteeing the stable response time for frequently referenced blocks in the case of existing a fault derive by placing the pair block of RAID-1 and the striping clock of RAID-5 orthogonally.
At the second year, we evaluated the proposed method in detail by simulation and clarify the efficiency of our method quantitatively. We considered the implementation of the block migration mechanism between RAID-1 area and RAID-5 area and the load balancing algorithm among multi derives. Then, we modeled these mechanisms and compared our method in comparison with the conventional raid systems by simulation. Consequently, we showed that our method could achieve the high performance. Although the essence of our method is extracting the potential parallelization of multi disks, it takes very long time to simulate our method. So, the evaluation of our method by multi media applications is the next subject and we are going to evolve our proposed method to these multi media applications. Less