Efficient TCP connection failover in web server clusters

Web clusters continue to be widely used by large enterprises and organizations to host online services. Providing services without interruption is critical to the revenue and perceived image of both hosts and content providers. Therefore, server node failure and recovery should be invisible to the clients. Most of the existing fault-tolerance schemes simply stop dispatching future client requests to the failed server. They do not recover those connections handled by the node at the time of failure, which makes the failure visible to some clients. Making the failure transparent requires both application-layer and transport-layer mechanisms. While atomic application-layer primary-backup failover schemes have been addressed at length in previous literature, a transport-layer scheme is necessary in order to make them invisible to the clients. In this paper we describe a transparent TCP connection failover mechanism. Besides transparency, our solution is also highly efficient, and does not need any dedicated hardware support.