Moving toward gray box predictive models at micro-architecture level by investigating program inherent parallelism

Predictive modeling has gained much attention in last decade aiming the evaluation of different design points in Design Space Exploration (DSE) process. However, predictive model construction still requires costly simulations for every new and unseen application. To reduce the number of simulations, several cross program performance prediction schemes are developed. These schemes can be more efficient if they use application characteristics or signature. The main challenge is to find the best representative characteristics and how they contribute to predictive models. In this paper, we introduce a characteristic to represent the inherent parallelism of applications and to measure performance sensitivity to issue width changes. We show that this characteristic can better predict inherent parallelism than data dependency distance and instruction mix, which are proposed by previous works.