Extending Byzantine Fault Tolerance to Replicated Clients
Ian Wehrman · Open Scholarship Institutional Repository (Washington University in St. Louis) · 2006
Byzantine agreement protocols for replicated deterministic state machines guarantee that externally re-quested operations continue to execute correctly even if a bounded number of replicas fail in arbitrary ways. The state machines are passive, with clients responsible for any active ongoing application behavior. How-ever, the clients are unreplicated and outside the fault-tolerance boundary. Consequently, agreement protocols for replicated state machines do not guarantee continued correct execution of long-running client applications. Building on the Castro and Liskov Byzantine Fault Tolerance protocol for unreplicated clients (CLBFT), we present a practical algorithm for Byzantine fault-tolerant execution of long-running distributed applica-tions in which replicated deterministic clients invoke operations on replicated deterministic servers. The algorithm scales well to large replica groups, with roughly double the latency and message count when com-pared to CLBFT, which supports only unreplicated clients. The algorithm supports both synchronous and asynchronous clients, provides fault isolation between client and server groups with respect to both correctness and performance, and uses a novel architecture that accommodates externally requested software upgrades for long-running evolvable client applications.