A comprehensive evaluation of availability and operational cost for a virtualized server system using stochastic reward nets

Virtualized server systems, as a major underlying element in high-performance computing systems, require further studies on many aspects of dependability. Among the significant factors, the availability measures are crucial to deliver high-quality services. Previous studies presented various modeling and analysis results on system availability of a virtualized system with two servers using a continuous-time Markov chain. In this study, we propose a cluster model of m virtualized servers using stochastic reward nets (SRNs). We focused on the overall configuration of the entire system, and in the modeling, we considered the detailed interactions between the servers. The model incorporates specific techniques for high availability of the system: standby techniques, virtual machine (VM) live migration and VM failover techniques. Simplified failures and recovery behaviors of physical servers and VMs are taken into consideration. Various SRN models are developed based on different case studies in which the techniques to improve the system’s overall availability are incorporated one after another. We conducted comprehensive analyses on the models with significant metrics of interest including: steady-state availability (SSA), sensitivity analysis of the SSA, downtime cost and operational cost analyses. We propose to use reward functions featured in SRN as a solution to help ease the computation of operational costs. The study provides an analytical basis for system adjustment and configuration of virtualized systems in data centers, cloud computing in practice.