Techniques for Scenario Prediction and Switching in System Scenario Based Designs

Abstract

Many modern applications exhibit dynamic behavior, which can be exploited for reduced energy consumption. In this dissertation we contribute to the development of a two-phase combined design-time/run-time methodology known as the system scenario design methodology. This methodology identifies different run-time situations and clusters similar behaviors into system scenarios. It is implemented in our framework for system scenario based designs, which dynamically tunes the hardware to match the application behavior. We develop techniques for scenario prediction and switching for data variable dependent applications, and achieve significant energy reductions when applying these techniques on an extracted control structure of a video codec widely used in hand-held devices today. Two different microcontroller boards are used to validate our techniques with the use of dynamic voltage and frequency scaling (DVFS). In addition, we developed a novel DVFS controller extension for the GEM5 computer architecture (CA) simulator in order to better evaluate the tuning of platform resources to the dynamic behavior of modern applications. The GEM5 extension is used together with our novel energy model, which is calibrated with measurements from real hardware. Finally, we develop a novel data dependent neuro-bio application that significantly reduces the amount of data to be transmitted from high-density electrode arrays. These arrays are used to study the behavior of neural cells and networks through extracellular monitoring of the action potentials in animals’ brains in vivo. This application also serves as a complex, realistic example of a data dependent application, which is applicable for our system scenario techniques.