Binding clients to replicated servers: initial and Continuous Binding

Much effort has been dedicated to the development of application-layer and network-layer mechanisms for selecting the "best" server among a set of replicated servers. In this work, we highlight a different problem that arises in the replicated server context; namely, when to to invoke and use a server selection mechanism. We use the term "binding" to refer to the function that queries a server selection mechanism and makes decisions about how to use the results. We consider two extreme cases of binding frequency, Initial Binding at the start of a connection, and Continuous Binding for each packet. We propose and evaluate a Continuous Binding scheme, and then compare these two extremes to one another. Our results indicate that a single client can improve performance by using Continuous Binding for a long-lived connection. However, when multiple clients use Continuous Binding, performance degrades. These results imply that network-layer server selection mechanisms may not be useful in the data plane for connection-based applications, which are the dominant form of communication in the current Internet.