Availability Modeling and Analysis of a Single-Server Virtualized System with Rejuvenation

Abstract

Availability of business-critical application servers is an issue of paramount importance that has received special attention from the industry and academia in the last decade. This paper presents two stochastic reward net based availability models for a single-server virtualized system. The similarity in both models is that software rejuvenation is applied at not only virtual machine monitor (VMM) level using a time-base policy but also at virtual machines (VMs) using a prediction-based policy. The key difference is that the passive software replication and the active software replication are respectively adopted at the VM level of both models. We compare these models in terms of steady-state system availability by numerical analysis. Results show steady-state system availability with the active replication style gets a bit better than that of the passive one. Further, we study the impact of two critical parameters, the VMM rejuvenation interval and the VM aging detection probability, on downtime and on the number of transaction lost by sensitivity analysis.