Abstract
Primary Backup Replication (PBR) suffers from the bottleneck problem at the primary, single point of failure, poor scalability, and imbalanced utilization of computing resources. These shortcomings are mainly due to the exclusive role of the primary in accepting and executing client's requests. In this paper, we tackle this problem and propose a novel approach that distributes the primary role among several nodes, where each node owns a subset (partition) of replicated data objects. In our approach, when a client creates an object, the created object is propagated to all nodes; however, the node that executes object's creation call becomes the owner (the primary) of this object. As a result, several nodes own different disjoint subsets of replicated objects instead of having only a single primary node owning the whole object store. The expression "owning an object" means the exclusive right of the owner node to lock/unlock the object. Distribution of data ownership among several nodes achieves scalability and improves load balancing. Furthermore, under our approach a replication system can be administrated seamlessly in case of node failure and network partitioning. Our approach generalizes and subsumes the traditional PBR approach.