An interesting approach to maintaining transmission control protocol (TCP) connections when a server crashes is described in this paper. TCP is usually implemented in the operating system, and the protocol provides reliable communications between a client and a server. TCP is designed to make it difficult for a third party to masquerade as the client or server; yet, this paper provides a novel mechanism to do just that. The technique allows a group of servers to masquerade as a failed server, maintaining and continuing all of its connections.

The implementation relies upon kernel modifications to track TCP packets and session state. This information is replicated on backup servers, which can then use their own modified kernel to masquerade as the original server. One of the interesting results is that this technique requires the ability to replay program state so that the connection, as well as the content, remains consistent.

This is a long paper that assumes detailed knowledge of TCP. The authors provide measurements and insight into the performance implications of the proposed technique. The technique explores a novel approach to server recovery that would be commercially interesting for many large-scale applications.