Most of the Blockchain permissioned systems employ Byzantine fault-tolerance (BFT) consensus protocols to ensure that honest validators agree on the order for appending entries to their ledgers. In this paper, we study the performance and the scalability of prominent consensus protocols, namely PBFT, Tendermint, HotStuff, and Streamlet, both analytically via load formulas and practically via implementation and evaluation. Under identical conditions, we identify the bottlenecks of these consensus protocols and show that these protocols do not scale well as the number of validators increases. Our investigation points to the communication complexity as the culprit. Even when there is enough network bandwidth, the CPU cost of serialization and deserialization of the messages limits the throughput and increases the latency of the protocols. To alleviate the bottlenecks, the most useful techniques include reducing the communication complexity, rotating the hotspot of communications, and pipelining across consensus instances.