User-perceived availability and response-time in voting-based replicated systems: a case study

In this paper, we develop a modeling method based on stochastic Petri nets (SPN) to allow user-perceived measures in voting-based replicated systems to be estimated. The merit of our approach is that user-arrival, maintenance, and node/link-failure or -repair processes are fully decoupled, thus allowing us to remove some unnecessary modeling assumptions and also to keep track of states in which the system is unavailable to users from the user's perspective. We apply our method to contrast user-perceived availability and performance measures under dynamic and static voting algorithms in a 3-node, fully-connected network and discover that (a) for user-perceived availability, the conditions under which static voting is better than dynamic voting, or vice versa, are largely determined by the user workload; (b) for user-perceived response time, static voting is always better than dynamic voting. We give some physical interpretation of the analysis result. Our method is generic in nature and can be applied to analyzing other voting algorithms or network structures for replicated data management.