A generic availability model for clustered computing systems

We study the availability of a clustered computing system with one cluster manager and "N+M" processing nodes, where M processing nodes serve as spares for the N active processing nodes. The functionality of an individual processing node is dissected into application software, management software, OS and hardware. The dependency among these entities is considered. Stochastic Petri net models are constructed to investigate the cluster availability. In order to deal with a cluster of a very large size, a solution based on state aggregation and fixed-point iteration is proposed. The existence and uniqueness of the fixed point is proved. The impact of a cluster manager, switchover time and coverage ratio are quantitatively studied. From the numerical results of a simple cluster with "2+1" processing nodes, we find that: (1) the availability of the cluster manager does not have a significant impact on the system availability, (2) system availability increases with the coverage ratio and decreases with the switchover time. Mechanisms to improve the system availability are discussed