Measurements and Models for Resiliency Assessment of VM clusters under Aging and Rejuvenation

We present an empirical assessment of a heterogeneous VM (Virtual Machine) network. We evaluate the VM network performance using six resiliency strategies. For each resiliency strategy, we apply three software rejuvenation approaches. Therefore, our experimental design is composed of eighteen experiments. For each measurement, we present the experiment average normalized distance from the scalability requirement. Here, the scalability requirement is the average workload response time plus three standard deviations of response time, measured at low load. In addition, we plot, for each experiment sample, the instantaneous available memory and normalized distance of response time from the scalability requirement. We have also parameterized the discrete-time embedded Markov chain with rewards for each experiment, using a discrete representation of the available memory as the Markov chain state variable, and, the normalized distance from the scalability requirement as the state reward. Coloring the states of the Markov chain state transition diagram enabled new insights into the software aging and rejuvenation process, for each strategy, and software rejuvenation granularity level. We conclude by presenting an optimal synthetic resiliency strategy that was derived using Markov decision processes (MDPs). These MDPs are based on state spaces that map to a partition of the amount of available memory into disjoint, contiguous intervals.