Validating complex computer system availability models

The authors report on experiences in validating complex computer-system availability models. A validation process, the availability models, and the data-collection process are described. An iteration of the model validation process emphasizing discrepancies between observed system behavior from data and from the model assumptions is presented. Analysis of data from five sites revealed that interruption dependencies on a device and across devices within a system exist. Furthermore, the frequency of, dependencies between, and downtime of device interruptions constitute differentiating characteristics of the causes of device interruptions. Data on partially available machine states, nonexponentially distributed machine downtimes, and nonstationarity of the machine interruption process are included. Validating the model assumptions is essential to model interpretation. Model accuracy is assessed. Exponentially distributed times between interruptions on a device and exponentially distributed device downtimes are not significant with respect to average measures of availability. The benefits of and issues for availability-model validation are summarized