信息科学技术学院计算机科学系学术讲座（二）

  
题  目：HIERARCHICAL RAID: Organization, Operation, Performance, Reliability, and Performability

讲座摘要：We use erasure coding (multiple parities) associated with Redundant Array of Independent Disks (RAID) to deal with both disk and controller failures in a 2-level Hierarchical RAID (HRAID). There are N nodes with redundancy K and M disks per node with RAID level L+4 array tolerating L=1,2,3 disk failures. HRAID K/L can tolerate K node failures and L disk failures per node. We can reconstruct a missing block on demand by accessing the corresponding M-L blocks at the RAID level,  or N-K blocks across the nodes, if the former is not possible. Rebuild (systematic reconstruction) is carried out by parity sparing both at the disk and node level. If disks are the only source of failure, (K+1)(L+1) failed disks can lead to data loss, but this number is much larger in practice. We report on a multistep recovery method which shows that the number of steps grows linearly with the number of failed disks. We describe the HRAID organization and operation, which requires locking to ensure correctness in updating check blocks. We also apply an approximate reliability analysis method, which shows that HRAID1/2 is more reliable than HRAID2/1. We discuss a simple analysis to assess the effect of updating parities on the maximum throughput attainable and the increased delayed in accessing disks. We then report on simulation studies to obtain the distribution of the number of disk failures leading to data loss, the MTTDL (Mean Time to Data Loss), and performability, which is the number of I/O accesses processed by HRAID. We conclude the discussion with a design study.
报告人：中科院外国专家特聘研究员、新泽西理工大学教授 Alex Thomasian
报告人简介：Alexander Thomasian is currently an outstanding visiting scientist at SIAT (Shenzhen Institutes of Advanced Technology). He received his PhD in Computer Science from UCLA (University of California in Los Angeles) in 1977. He was a faculty at Case Western Reserve University, University of Southern California, and more recently, after taking an early retirement from IBM Research in New York at New Jersey Institute of Technology. He graduated 10 PhD and a dozen MS students. He became an IEEE Fellow in January 2000 based on his early research in transaction processing in databases and performance analysis of computer systems. He holds 4 patents two of which are on high-dimensional indexing structures, another area in which he conducted research both at IBM and NJIT. In the last ten years his research has dealt with storage systems. He has been an editor of IEEE Transactions on Parallel and Distributed Systems and on the program committees of numerous conferences.  He has over 150 publications, with over 60 in journals.
时 间：2011年6月27日（周一）上午10：00
地 点：教学大楼A215室
 
热烈欢迎广大感兴趣的师生参加!
 
信息科学技术学院
2011年6月22日

作者：吉念宁