Implementation and experimental evaluation of live replacement and reconfiguration

Abstract

State machine replication is a common applied technique for building fault-tolerant services. The technique uses a collection of replicas to mask failures. All replicas are provided the same sequence of operations (requests), resulting in that they end up in a consistent state. A consensus protocol such as Paxos is normally used to the order request issued by multiple clients to a Replicated State Machine (RSM). RSMs achieve high availability by replicating state across several machines. Such an approach enables access to state stored in system even in the presence of failures. An RSM is prohibited from processing new requests if more than half of its replicas fail. It is therefore important that replica failures are handled and repaired as soon as possible for keeping the availability and fault-tolerance of the RSM high. This thesis presents an description and implementation of two existing methods for immediate failure handling for Paxos-based RSMs: Live Replacement and Reconfiguration. Both failure handling methods have been implemented as part of the Goxos framework. An experimental evaluation and comparison of the two methods is also presented.