Abstract
This paper is concerned with predicting the response times an enterprise information system would provide on new server architectures. These predictions can allow a workload to be transferred onto new servers whilst maintaining quality of service levels. Two common techniques are solving queu- ing models and extrapolating from previously gathered performance data. The dynamic recalibration of a layered queuing model and a historical model are investigated experimentally using an established distributed enterprise benchmark. The conclusions provide guidelines as to how to select an appropriate technique, including how to dynamically calibrate each model at a low overhead. Using these guidelines it is shown that both techniques can make low overhead predictions for new server architectures at a good level of predictive accuracy (typically over 80%).