On the Latency of Erasure-Coded Cloud Storage Systems

Distributed (Cloud) Storage Systems (DSS) exhibit heterogeneity in several dimensions such as the volume (size) of data, frequency of data access and the desired degree of reliability. Ultimately, the complex interplay between these dimensions impacts the latency performance of cloud storage systems. To this end, we propose and analyze a heterogeneous distributed storage model in which $n$ storage servers (disks) store the data of $R$ distinct classes. Data of class $i$ is encoded using a $(n,k_{i})$ erasure code and the (random) data retrieval requests can also vary from class to class. We present a queuing theoretic analysis of the proposed model and establish upper and lower bounds on the average latency for each data class under various scheduling policies for data retrieval. Using simulations, we verify the accuracy of the proposed bounds and present qualitative insights which reveal the impact of heterogeneity and scheduling policies on the mean latency of different data classes. Lastly, we conclude with a discussion on per-class fairness in heterogeneous DSS.