Scientific applications doing complex numerical simulations may run for very long times. Performance of such applications is a subject for significant runtime dynamics which may evolve with time into severe performance degradations. However, performance analysis tools provide very limited support in this respect, leaving the job of identifying performance changes to the user. In this paper we present a model which explicitly represents and quantifies performance variability found in performance profile time-series. The model defines a hierarchy of entities explicitly expressing different aspects of performance dynamics. These are supplemented with severity functions quantifying performance variability associated with the entities based on the Wavelet transform of the underlying profile time-series. We demonstrate our technique by prototyping an extension of the CUBE tool towards efficient exploration of performance dynamics and evaluate it on a real-world application.