Group Communication as a Base for a Load-balancing Replicated Data Service Group Communication as a Base for a Load-balancing Replicated Data Service

Replication and load-balancing are two fundamental techniques for improving availability and performance of distributed systems. However, correct and eecient realization of these techniques is intricate when the distributed environment may partition and merge because of processor and communication failures. In this thesis, we show how a view-synchronous group communication service recently speciied by Fekete, Lynch, and Shvartsman can be used to support a sequentially consistent replicated data service that load balances queries and tolerates partitioning and merging of the underlying network. Our work is done in the framework of the I/O automaton model of Lynch and Tuttle. First, we present an I/O automaton speciication that deenes the allowed behavior of a sequentially consistent replicated service. Then, we successfully reene this speciica-tion to arrive at an I/O automaton that models a distributed implementation of this service. In this implementation, update requests are processed in the same order at all servers of the system, thus guaranteeing mutual consistency of all replicas; query requests are processed at single servers determined by a load-balancing strategy which equalizes the number of queries assigned to each member of the same group. Third, we give a rigorous hierarchical proof that the implementation automaton implements the speciication automaton in the sense of trace inclusion. This proof establishes partial correctness of the implementation. Finally, we prove a liveness-related claim that servers are always able to process the queries assigned to them; that is, the servers are never blocked by missing update information. Acknowledgments I am grateful to Dr. Nancy Lynch for her guidance and many insightful discussions throughout this project. I am also grateful to Dr. Alan Fekete who, together with Dr. Nancy Lynch, suggested this project and helped me make the rst steps. I owe many thanks to Dr. Alex Shvartsman for introducing me to Distributed Computing back when I was an undergraduate and for encouraging me to pursue this eld during my graduate studies.