JCSE, vol. 13, no. 2, pp.39-65, 2019

DOI: http://dx.doi.org/10.5626/JCSE.2019.13.2.39

Analysis of the Factors Influencing Architectural Time-Predictability of Superscalar Processors

Mutaz A. B. Al-Tarawneh
Department of Computer Engineering, Faculty of Engineering, Mutah University, Karak, Jordan

Abstract: Architectural time-predictability (ATP) is a major component in the design of real-time systems (RTS). In these systems, it is imperative to maintain temporal and logical accuracy. Hence, designers need to estimate the worst-case behavior of such systems before deploying them in real-life applications. This paper presents an experimental framework to quantify the ATP of modern superscalar processors. In addition, it aims at characterizing, modeling, and analyzing the factors that influence the ATP of such processors. Hence, a total of 33 benchmark programs derived from the Malardalen WCET benchmark suite have been simulated on a specialized superscalar processor simulator. The standard deviation of the cycles per instruction (CPI) of the simulated superscalar processor has been used to quantify ATP. In addition, an unsupervised machine learning approach has been employed to classify the benchmark programs into several clusters based on their sensitivity to different processor components. Moreover, a design of experiment (DoE) methodology has been utilized for factor screening, sensitivity analysis, model building, and design space navigation. It has been shown that the benchmark programs exhibit a varying degree of dissimilarity in their temporal behavior. In addition, the proposed framework can be used to model the ATP of superscalar processors adequately and optimize their design to increase their ATP levels.

Keyword: Real-time systems; Time-predictability; Cycles per instruction; Standard deviation; Design of

Full Paper:   360 Downloads, 1593 View