A realistic evaluation of consistency algorithms for replicated files

Abstract

Data are often replicated in distributed systems to protect them against site failures and network malfunctions. When this is the case, an access policy must be chosen to insure that a consistent view of the data is always presented. Voting protocols guarantee consistency of replicated data in the presence of any scenario involving non-Byzantine site failures and network partitions. While Static Majority Consensus Voting protocols use static quorums, Dynamic Voting protocols, like Dynamic Voting and Lexicographic Dynamic Voting, dynamically adjust quorums to changes in the status of the network of sites holding the copies. The availabilities of replicated data managed by these three protocols are compared using a simulation model with realistic parameters. Dynamic Voting is found to perform better than Majority Consensus Voting for all files having more than three copies while Lexicographic Dynamic Voting performs much better than the two other protocols for all eleven configurations under study.