Efficient and scalable replication of services over wide-area networks

Abstract

Service replication ensures reliability and availability, but accomplishing it requires solving the total-order problem of guaranteeing that all replicas receive service requests in the same order. The problem, however, cannot be solved for a specific combination of three factors, namely, when (i) the message transmission delays cannot be reliably bounded, as often the case over wide-area networks such as the Internet, (ii) replicas can fail, e.g., by crashing, the very events that have to be tolerated through replication, and finally (iii) the solution has to be deterministic as distributed algorithms generally are. Therefore, total-order protocols are developed by avoiding one or more of these three factors by resorting to realistic assumptions based on system contexts. Nevertheless, they tend to be complex in structure and impose time overhead with potentials to slow down the performance of replicated services themselves. This thesis work develops an efficient total-order protocol by leveraging the emergence of cluster computing. It assumes that a server replica is not a stand-alone computer but is a part of a cluster from which it can enlist the cooperation of some of its peers for solving the total-order problem locally. The local solution is then globalised with replicas spread over a wide-area network. This two-staged solution is highly scalable and is experimentally demonstrated to have a smaller performance overhead than a single-stage solution applied directly over a wide-area network. The local solution is derived from an existing, multi-coordinator protocol, Mencius, which is known to have the best performance. Through a careful analysis, the derivation modifies some aspects of Mencius for further performance improvements while retaining the best aspects.