Efficient quorum operations in replicated databases

Accessing and updating replicated data using quorum based methods achieve the highest degrees of data and system availability. A quorum of replicas must be used to access/update the data. Unfortunately the size of the quorums increases as the number of replicas increases, imposing limitations on the scalability of replicated systems that use the quorum method. We present and evaluate heuristics for optimizing the selection of quorums by: minimizing message exchange, load balancing the replicas, and minimizing the quorum formation delays. We show that even though minimum size quorums incur the least number of messages, they cause serious congestion problems, and are not optimal quorums in all cases. We also show that larger but load balanced quorums achieve better performance. In addition, we show a particularly interesting and important observation on the effect of quorum selection on the concurrency conflict behavior of transactions. Theoretically, it is known that the choice of quorums may affect the transaction conflict rate, but not the transaction abort rate. Practically, however, we found that the choice of quorums may affect both conflict and abort rates. We provide a detailed explanation of this effect, and offer heuristic based guidelines for the efficient selection of quorums.