Tradeoffs in implementing primary-backup protocols

One way to implement a fault-tolerant service is by using multiple servers that fail independently. The state of the service is replicated and distributed among these servers, and updates are coordinated so that even when a subset of the servers fail, the service remains available. A common approach to structuring such replicated services is to designate one server as the primary and all the others as backups. Clients make requests by sending messages only to the primary. If the primary fails, then a failover occurs and one of the backups takes over. This service architecture is commonly called the primary-backup or the primary-copy approach. In most such primary-backup protocols, when the primary receives a client request, it informs the backups about the request, and then responds to the client. Informally, this primary-backup protocol is non-blocking if the primary does not wait for an acknowledgement from the backups before it sends the response; otherwise, it is blocking. Most of the existing protocols are blocking as non-blocking protocols cannot be constructed for some kinds of failures. However, it is shown that non-blocking protocols can be constructed for most of the process and communication failures that are expected to occur in the primary-backup systems of the future. Since non-blocking protocols can theoretically achieve the smallest possible response time, this paper analyzes these protocols under various system parameters. Two kinds of non-blocking protocols are analyzed: one in which the processes use point-to-point communication to exchange messages, and the other in which processes use hardware broadcasts.