Low-level estimation at high-levels of abstraction in system-level design

Embedded systems are becoming increasingly complex with shortening time-tomarket demands. System-level modeling and design have been proposed to help embedded system development keep pace with this complexity. In a system-level design environment, a designer is able to delay critical design decisions until late in the design cycle, reducing the risk of making incorrect decisions which could require a costly redesign. New methods of estimating system-level performance must be devised to accommodate these needs.;In embedded systems composed of off-the-shelf parts, performance can be roughly estimated using part documentation. However, this process can provide poor estimates. Additionally, if the design includes a custom part, there may not be detailed documentation from which to gather performance estimates. The exhaustive gathering of estimates is error prone and tedious. In this thesis we present a novel estimation technique called minimal characterization for creating system-level estimation metrics. We show that estimates can be orders of magnitude more accurate, without any loss in fidelity, using a small number of source-level metrics. We show results from applying a source-level performance estimation technique generally used on software systems to a system-level design that is implemented in both software and hardware targets. Finally, we present a categorization of secondary execution factors which can greatly affect the accuracy of system-level estimates but have only been peripherally addressed in other approaches