Reliability Modeling for Large Scale Declustered Storage
SESSION: Research Poster Reception
EVENT TYPE: Poster
TIME: 5:15PM - 7:00PM
AUTHOR(S):Lin Xiao, Garth Gibson
ABSTRACT: Most reliability models built for RAID are based on non-declustered storage or with simple notion that reliability should be measured as the time until any data is lost. With declustered storage system, the amount of data loss varies as data loss happens. The data loss can be infrequent but huge or frequent but tiny. In this poster we improve traditional models for redundant disk storage reliability in two ways: the inclusion in the model of parity declustering and expanded metrics for the amount of data lost annually rather than simply the expected number of loss events. With a renewal reward stochastic Markov model, we also show that many factors, such as failure detection time, I/O bandwidth for repair, and chunk size will influence the reliability for a declustered storage system and triplication is not always better than RAID6.