Realism in Statistical Analysis of Worst Case Execution Times

Abstract

This paper considers the use of Extreme Value Theory (EVT) to model worst-case execution times. In particular it considers the sacrifice that statistical methods make in the realism of their models in order to provide generality and precision, and if the sacrifice of realism can impact the safety of the model. The Gumbel distribution is assessed in terms of its assumption of continuous behaviour and its need for independent and identically distributed data. To ensure that predictions made by EVT estimations are safe, additional restrictions on their use are proposed and justified.