Accurate run-time prediction of performance degradation under frequency scaling

Dynamic voltage and frequency scaling is employed to minimise energy consumption in mobile devices. The energy required to execute a piece of software is highly depedent on its execution time, and devices are typically subject to timeliness or quality-of-service constraints. For both these reasons, the performance at a proposed frequency setpoint must be accurately estimated. The frequently-made assumption that performance scales linearly with core frequency has shown to be incorrect, and better performance models are required which take into account the effects, and frequency setting, of the memory architecture. This paper presents a methodology, based on off-line hardware characterisation and runtime workload characterisation, for the generation of an execution time model. Its evaluation shows that it provides a highly accurate (to within 2% on average) prediction of performance at arbitrary frequency settings and that the models can be used to implement operating-system level dynamic voltage and frequency scaling schemes for embedded systems