Data allocation for multi-class distributed storage systems

Abstract

Distributed storage systems (DSSs) are vastly used for reliably storing large amounts of data generated by current and future wireless networks, e.g. social mobile networks or Internet of things. Depending on the features of the data source, various data files may require different levels of quality of service (QoS), e.g. in terms of the probability of successful recovery or data recovery delay. This means that data files can be divided into different classes in terms of their QoS requirements. To address the requirements of each class of data, efficient data (storage) allocation methods, meaning how data is spread over the storage nodes, should be devised. In this paper, we study the optimal data allocation for maximizing the weighted sum of the probability of successful recovery of the data of different classes. Finding such optimal allocations is intractable in general. Therefore, we focus on finding the optimal minimal spreading allocation (MSA) where the data of each class is spread minimally over the storage nodes. MSA possesses several benefits including minimum expected recovery delay and maximum average service rate. Simulation results show that our proposed MSA is indeed the optimal storage allocation in many cases.