Predicting computation time for advanced processor architectures

Estimating computation times using analysis techniques is always safe but is becoming prohibitively complex or pessimistic with modern processors. The only alternative approach is to use measurement, but this has the significant disadvantage of optimism - the largest value seen during testing may not be the largest experienced during deployment. In this paper, we subject data obtained from measurement to statistical analysis using the techniques of extreme value estimation. A simple case study is described and the approach is illustrated via this study which focuses on the superscalar technique of branch prediction. The approach is applicable to all forms of hardware-induced temporal variability.