Sensitivity Analysis Based Predictive Modeling for MPSoC Performance and Energy Estimation

Multi-processor system on chip (MPSoC) has been a de facto standard for embedded processor architectures. However, the architectural design space of MPSoC is so huge that it is time prohibitive to exhaustively simulate all design points to evaluate their design metrics (such as performance, energy, etc.). Thus, many architects have resorted to predictive modeling methods to fast estimate the design metrics of design points. An essential task in these techniques is input variable selection. Input variables of the predictive model consist of architecture parameters and their interactions, but not all input variables should be included in model. The inclusion of significant input variables in model can improve the prediction accuracy of model, but the inclusion of insignificant input variables will increase the risk of over fitting. So, how to identify and include the significant input variables while exclude the insignificant ones is a great challenge. In this paper, we propose an adaptive component selection and smoothing operator (ACOSSO) regression technique for predictive modeling of MPSoC performance and energy. The ACOSSO regression technique allows simultaneous global sensitivity analysis (which performs input variable selection) and model computing through solving an L1-norm penalized least squares fitting problem. We compare the proposed ACOSSO model with the state-of-the-art restricted cubic splines (RCS) model and two enhanced RCS models by applying them to an MPSoC performance and energy estimation problem. One enhanced RCS model performs input variable selection by use of ACOSSO regression based sensitivity analysis technique and the other by a stepwise regression modeling technique. Experimental results show that the ACOSSO regression model has better prediction accuracy than the other models, and the results of ACOSSO regression based sensitivity analysis are also useful for RCS modeling.