Delay and Loss Rate Analysis of the Log Commitment Process in Raft

Consensus algorithms are crucial in attaining consistency among data saved at different locations across distributed and decentralized computing systems. For consensus algorithms to work properly on the Internet, the designs must provide sustainable performance over different adverse networking conditions, including packet loss events, the extra delay incurred during deliveries, etc. In this paper, we shall analyze the performance of the log replication process in the Raft consensus algorithm through a generalized network emulation conceptual switch model. Though there are other approaches, e.g., testbed experiments and discrete event simulation (DES), to validate the algorithmic performance of the log replication process, many among them usually associate different packet loss rates with the overall throughput performance. Our proposed model can associate the loss rates with the expected wait durations until commitments or confirmations (leader elections or data storage) in Raft. Furthermore, we analyze the bounds on the mean and variance of these successive log replication confirmation durations. Compared to DES simulations, our analysis results are closely matched, and the efficiency is improved, e.g., on the same computing machine, it took about 11,148 seconds to run a batch of DES simulations with a 0.9 packet loss rate, and it took only 44 seconds to generate all bounding results in the range of [0, 0.9] through the derived expressions. In a distributed system with one leader and 30 followers using the regular Raft algorithm, through our analysis, increasing packet loss rate from 0.001 to 0.1 can lead to a 39.4% increase in the average log replication confirmation time. This indicates that the network packet loss rate significantly impacts the performance of consensus algorithms such as Raft. Our proposed model can provide performance measures for analyzing novel consensus protocols on the Internet in future.