Abstract:

This thesis intends to contribute towards the realization of improved embedded design tools and techniques with the introduction and development of empirically based design methodologies. These methodologies introduced here take advantage of abundantly available data in performing statistical analysis to realize varied objectives such as the characterization of process execution time, or the prediction of actual execution time, or even the proper sizing of memory hierarchies. The solutions sought are achieved by relying on analysis such as Design of Experiments to systematically and efficiently design future experimental runs, and to distinguish between influential and insignificant factors. Other techniques included the use of factor analysis to investigate and test relationships amongst factors, multivariate regression to explore factor settings, and response surface modeling to guide solution set traversals. Such an approach is empirical in nature as it attempts to analyze and understand the dynamics of any single component of the system from the vast amount of quantitative data it turns out. The sheer magnitude of this information demands that the implementation of this methodology incorporates the usage of multivariate statistics to capably extract and process any relevant information, sift through redundancies, and to efficiently design any further experiments necessary to achieving the desired progress. In addition it provides the means to consider the effects of component interactions, and the capacity to rate their significance and overall impact. Hence, one can simultaneously investigate the special requirements of an embedded system without having to isolate them from each other. Here empirical methods to statistically characterize task execution time are presented along with other quantitative means of dynamically predicting the execution time. The other topic considered is the development of a methodology to optimize system components for multiple objectives such as power and performance. In general the purpose of the work presented here is to encourage the adoption of such a scientific method which relies on the observation of empirical evidence to drive the design of embedded system platforms. The need for specialized embedded tools stems in part from the fact that traditionally embedded systems did not receive the same amount of attention usually bestowed upon other more predominant areas of computing. Such a lack of interest is made all the more startling when considering the disproportionate amount of processing capacity being devoted to embedded systems as opposed to ones which have more of a general purpose. This has inevitably led to a situation where the majority of available design tools cater to these systems, thus neglecting to properly deal with the special conditions unique to embedded systems. Given these requirements and the limited and restrictive nature of embedded systems in comparison with general purpose ones, it's unreasonable to think that development for these environments can be sufficiently handled using only these tools. Therefore, there needs to be more effort exerted to reaching the goal of having tools dedicated to the assistance of engineers working on embedded platforms.