Characterizing and exploiting task load variability and correlation for energy management in multi core systems

We present a hybrid energy management technique that exploits the variability of and correlations among the computational loads of tasks in a real-time application with soft timing constraints. In our technique, task load variability and correlations are captured in stochastic models that incorporate certain salient features and essential characteristics of the application. We use the stochastic models in formulating and solving the energy management problem for applications with soft timing constraints running on multiprocessor systems with dynamic voltage scaling (DVS). We present a novel optimization formulation for minimizing average energy consumption while providing a probabilistic guarantee for satisfying timing constraints. We compare our stochastic models and energy management scheme with other models and schemes that do not capture/exploit either the variability of or the correlations among the computational loads of tasks.