Scalable Agreement: Toward Ordering as a Service

Replicated state machines use agreement protocols such as Paxos to order client requests. These protocols are not scalable and can quickly become a performance bottleneck as the degree of fault-tolerance and the demand for throughput performance increase. We propose a scalable agreement protocol that can utilize additional resources to provide higher throughput, while guaranteeing linearizability for client requests. Our protocol can build on existing optimizations, as it can use protocols like Paxos as a building block. A preliminary performance evaluation shows a throughput increase of 50%−179% over a baseline strategy, even without adding any hardware; and with additional hardware we are able to achieve even higher performance gains.