Modeling replication and erasure coding in large scale distributed storage systems based on CEPH

Abstract

The efficiency of storage systems is a key factor to ensure sustainability in data centers devoted to provide cloud services. A proper management of storage infrastructures can ensure the best trade off between costs, reliability and quality of service, enabling the provider to be competitive in the market. Heterogeneity of nodes, and the need for frequent expansion and reconfiguration of the subsystems fostered the development of efficient approaches that replace traditional data replication, by exploiting more advanced techniques, such the ones that leverage erasure codes. In this paper we use an ad-hoc discrete event simulation approach to study the performances of replication and erasure coding with different parametric configurations, aiming at the minimization of overheads while obtaining the desired reliability. The approach is demonstrated with a practical application to the erasure coding plugins of the increasingly popular CEPH distributed file system.