Client Group Membership as an Architectural Approach for Dependability in Large Scale Systems

Dependability in large scale systems combines the inherent complexity of fault tolerant systems with the scalability problems that appear in large environments. In this paper we propose an architectural approach to design highly available systems in the scenario where a reduced set of servers cooperate to provide replicated services to an unbound and dynamically changing set of clients. While the central group of servers is supposed to be totally connected and fault free during long periods of time, clients are supposed to access them with a highly dynamic pattern of connections, disconnections and failures. This particular scenario often appears in large scale dependable environments, where fault tolerant services hosted at cluster-like networks are accessed from over the world through much less resilient communications and computers. Our approach will consist in extending traditional group membership protocols with information about client membership, so that the server software can react upon client activity.