An Evaluation of Consensus Latency in Partitioning Networks

Consensus, or state machine replication, is critical for the deployment of distributed battlefield systems. Battlefield networks operate in environments with unpredictable wireless connectivity which lead to sparse networks and frequent partitioning, and this makes deploying centralized architectures where nodes require a connection to a remote server unsuitable. The Extended Virtual Synchrony (EVS) model provides membership views which enables a network to reach consensus even after experiencing a series of partitions and mergers. If a node wants to propose state transitions that require nodes that are not currently in its membership view, then the node needs to wait until it reconnects with those nodes. The time the node has to wait to reconnect to the other nodes introduces consensus delays in the network. In this work, we evaluate consensus latency by focusing on these queued state transition proposals due to both network partition characteristics and distributed application/mission design. The key findings of our results show that consensus delay is least affected by network partitioning when the network splits at a rate equal to or less than 1/4 the rate in which partitions merge. Our evaluation results provide application and mission designers guidelines on the tradeoffs between several network characteristics and desired consensus latency properties.