The effect of workload on the performance and availability of voting algorithms

Abstract

Voting algorithms are a popular way to provide data consistency in replicated data systems. By maintaining multiple copies of data on distinct servers, they can increase the data's availability, as perceived by a user. Many models have been made to study the degree to which replication increases the availability of data, and some have been made to study the cost incurred in maintaining consistency. However, little work has been done to evaluate the time it takes to serve a request, accounting for server and network failures, or to determine the effect of workload on these measures. The effect of workload can be significant, since failures of system components are not important unless they are needed to deliver a service, and requests can force updates on data that would otherwise be outdated. In this paper, we use stochastic activity networks (SANs), a variant of stochastic Petri nets, to construct two variant models of a replicated file system, one using a static voting algorithm while the other using a dynamic voting algorithm to maintain data consistency. A Markov process representation is automatically constructed from each SAN model and is solved numerically. Specifically, we determine the availability and mean time to respond to write requests as a function of the number of replicated copies and workload offered to the system. The results illustrate that it is indeed possible to determine such measures analytically and that workload, as well as the number of copies, is an important determinant of availability and response time.