P-HotStuff: Parallel BFT algorithm with throughput insensitive to propagation delay

Abstract

In this work, we present P-HotStuff, a novel variant of HotStuff consensus algorithm with multiple parallel operations, which can effectively solve the bottleneck of the Byzantine Fault Tolerance (BFT) algorithms that employ the leader-based consensus model, where the throughput is sensitive to Propagation Delay, resulting in the bandwidth of each node is frequently idle. The parallel operations consist of three parts. First, the Broadcast layer is decoupled from the Agreement layer and they run in parallel, where the Broadcast is for preparing the inputs for each consensus, and the Agreement is for determining the inputs. Secondly, instead of only the leader, all the nodes can prepare the inputs in parallel. Lastly, the node can prepare each input in parallel, which means that it can directly prepare its next input without waiting for the completion of its preceding preparation. We have conducted the experiments and compared our P-HotStuff with HotStuff and the latest work Motorway. The experimental results show that P-HotStuff can achieve an average throughput that is about 20 times that of HotStuff and 50% higher than that of Motorway under the condition of about 60 nodes, 256 Bytes payload, batch size of 400 and 100 Mbps bandwidth in a Wide Area Network spanning multiple states with an average propagation delay of 260 ms.