Has anybody tried to run PoA networks with 100 or more validators? Given a fixed block size, how did the number of validators effect transaction throughput and latency?
This paper addresses detail comparison of Aura, Clique and PBFT.
The key difference in terms of latency according to this paper is :
In Aura, each block proposal requires two message rounds: in the first round the leader sends the proposed block to all the other authorities, in the second round each authority sends the received block to all the other authorities. A block is committed after a majority of authorities have proposed their blocks, hence the latency in terms of message rounds in Aura is 2( N/ 2 + 1), where N is the number of authorities. In Clique, a block proposal consists of a single round, where the leader sends the new block to all the other authorities. The block is committed straight away, hence the latency in terms of message rounds in Clique is 1. Such a huge difference between Aura and Clique is due to their different strategies to cope with malicious authorities aiming at creating forks: Aura waits that enough other blocks have been proposed before committing, Clique commits immediately and copes with possible forks after they occur. Clique seems to outperform PBFT too, which takes three message rounds to commit a block.
So the number of authorities has a linear relationship with the latency (in terms of message rounds)in case of Aura. If the number of authorities increases the latency (in terms of message rounds) will increase.
On the other hand latency(in terms of message rounds) remains unaffected by the number of authorities.