A statistical machine learning based modeling and exploration framework for run-time cross-stack energy optimization

As the complexity of many-core processors grow, meeting performance, energy, temperature, reliability, and noise requirements under dynamically changing operating conditions requires run-time optimization of all parts of the computing stack - architecture, system software, and applications. Unfortunately, the combination of design parameters for the entire computing stack results in an operating space of millions of points that must be explored and evaluated at run-time. In this paper, we present a statistical machine learning (SML) based modeling framework that can be used to rapidly explore such vast operating spaces. We construct a multivariate adaptive regression spline (MARS) based model that uses a number of architecture and application parameters as predictor variables to predict performance and power. We then use a Pareto-front exploring evolutionary algorithm to determine operating points for optimal power and performance. The operating points constituting the Pareto front are stored in look-up tables for runtime use. The proposed framework is applied to an ×264 video encoding application executing on a quad core processor. The microarchitectural predictor variables include core and cache parameters. The application predictor variables include the video resolution, and visual quality determined by the choice of the motion estimation algorithm. The model outputs the average frames per second (FPS) and the average power consumption. The MARS model has an R² of 0.9657 and 0.9467 respectively for FPS and power. For a video frame resolution of 480x320, and FPS of 20, a power saving of 55% can be obtained by jointly tuning the microarchitectural parameters and the visual quality.