Conngurable Highly Available Distributed Services

A service provided by a computing system is characterised as fault-tolerant [4] when it continues to be provided according to its speci cations despite failures of system components (software or hardware) that participate in the service provision. With the ever increasing introduction of computing systems in many aspects of today's life, fault-tolerance of critical services becomes of great importance. There are two main parameters related to the faulttolerant behaviour of a service: reliability de ned as the eventual correctness of the service, and availability the probability of a service being correct at a speci c moment in time. Techniques for consistent distributed recovery from failures are employed to achieve reliability. Redundancy is used to improve availability: replicated servers are employed for the provision of the service. In the literature [14, 11, 9, 2], the interest of researchers focuses on the problem of maintaining replica consistency, when replication is employed to achieve high availability. An additional problem raised is that of dynamic system recon guration. Dynamic con guration management is required in order to replace failed replicas, upgrade the server implementation, or change the availability characteristics of the service. According to Cristian [5], the availability characteristics of a service are described in terms of an availability policy for the service. Two dimensions are distinguished, namely replication policy (number of replicas required) and synchronisation policy (how close the states of replicas are synchronised). Here, we focus only on replication policy as a recon guration problem. For simplicity, we assume close [4] synchronisation policy throughout the paper. In this context, the requirements for providing highly available services are analysed, and an architecture and partial implementation for a replicated server group is presented. The architecture facilitates the dynamic con guration management of the replicated server group, while maintaining replica consistency. 2 System model