Analysis of Replicated Data with Repair Dependency

Abstract

Pessimistic control algorithms for replicated data permit only one partition to perform update operations at any time so as to ensure mutual exclusion of the replicated data object. Existing availability modelling and analyses of pessimistic control algorithms for replicated data management are constrained to either site-failure-only or link-failure-only models, but not both, because of the large state space which needs to be considered. Moreover, the assumption of having an independent repairman for each link and each site has been made to reduce the complexity of analysis. In this paper, we remove these restrictions with the help of stochastic Petri nets. In addition to including both site and link failures/repairs events in our analysis, we investigate the effect of repair dependency which occurs when many sites and links may have to share the same repairman due to repair constraints. Four repairman models are examined in the paper: (a) independent repairman with one repairman assigned to each link and each node; (b) dependent repairman with FIFO servicing discipline; (c) dependent repairman with linear-order servicing discipline; and (d) dependent repairman with best-first servicing discipline. Using dynamic voting as a case study, we compare and contrast the resulting availabilities due to the use of these four different repairman models and give a physical interpretation of the differences. We show that ignoring concurrent site and link failures/repairs events or repair dependency can very unrealistically overestimate the availability of replicated data.