DRESS codes for the storage cloud: Simple randomized constructions

Abstract

We introduce an efficient family of exact regenerating codes for data storage in large-scale distributed systems. We refer to these new codes as Distributed Replication-based Exact Simple Storage (DRESS) codes. A key property of DRESS codes is their very efficient distributed and uncoded repair and growth processes that have minimum bandwidth, reads and computational overheads. This property is essential for large-scale systems with high reliability and availability requirements. DRESS codes will first encode the file using a Maximum Distance Separable (MDS) code, then place multiple replicas of the coded packets on different nodes in the system. We propose a simple and flexible randomized scheme for placing those replicas based on the balls-and-bins model. Our construction showcases the power of the probabilistic approach in constructing regenerating codes that can be efficiently repaired and grown.