Performance evaluation of the quorum consensus replication method

The goal of data replication in distributed database systems is to increase data availability in the presence of failures. Using the quorum consensus method, up to [(n+1)/2] site failures can be tolerated, in an n-site system without loss of data accessability. Quorum consensus can however be very expensive to use, especially in large-scale systems. This is because multiple sites in the system must be accessed to perform the read or write operations. This paper describes an actual implementation of the quorum consensus method and gives an experimental evaluation of its performance. The implementation was done in the context of the Purdue Raid, which is a LAN-based distributed database system with extensive experimentation infrastructure. We focus our description on the particularities of managing version numbers and choosing object weights and thresholds. We also present a performance evaluation study, where the message traffic overhead, throughput and response time, and availability of the quorum consensus method is studied and compared to the read-one-write-all method