Tree-structured Data Regeneration in Distributed Storage Systems with Regenerating Codes,” in INFOCOM, 2010. Centre de recherche INRIA Rennes – Bretagne Atlantique IRISA, Campus universitaire de Beaulieu - 35042 Rennes Cedex (France) Centre de recherche I

✤ Large-scale data storage service ✤ Data are stored in a distributed fashion, i.e., in a number of storage nodes in the network ✤ Disk failures are common ✤ There are so many disks in large distributed storage systems, e.g., data centers ✤ The permanent data loss then becomes very likely ✤ We address this issue in this presentation 2 Redundancy ✤ To compensate for disk failures, redundancy is introduced in the distributed storage system ✤ Replicas ✤ Erasure codes ✤ Encode data on n disks into n+m disks so that m disk failures can be tolerated. ✤ RAID is a centralized fashion of erasure codes. 3 Data Regeneration ✤ Redundancy is not enough. ✤ Using erasure codes, data become not available if more than m disk failures occur ✤ Newcomer and providers ✤ To repair the lost redundancy, a newcomer receives redundant data from a certain number of providers 4 Data Regeneration ✤ Redundancy is not enough. ✤ Using erasure codes, data become not available if more than m disk failures occur ✤ Newcomer and providers ✤ To repair the lost redundancy, a newcomer receives redundant data from a certain number of providers 4 Network Coding [Dimakis07] ✤ This paper addressed how to regenerate the lost data by transmitting data as little as possible ✤ More providers, less regeneration traffic 5 Our Observation ✤ An efficient regeneration should not only consider how to achieve the optimal regeneration traffi