Adaptive Resource Management for Uncertain Execution Platforms

Embedded systems are becoming increasingly complex. At the same time, the components that make up the system grow more uncertain in their properties. For example, current developments in CPU design focuses on optimizing for average performance rather than better worst case performance. This, combined with presence of 3rd party software components with unknown properties, makes resource management using prior knowledge less and less feasible.

This thesis presents results on how to model software components so that resource allocation decisions can be made on-line. Both the single and multiple resource case is considered as well as extending the models to include resource constraints based on hardware dynam- ics. Techniques for estimating component parameters on-line are presented. Also presented is an algorithm for computing an optimal allocation based on a set of convex utility functions. The algorithm is designed to be computationally efficient and to use simple mathematical expres- sions that are suitable for fixed point arithmetics. An implementation of the algorithm and results from experiments is presented, showing that an adaptive strategy using both estimation and optimization can outperform a static approach in cases where uncertainty is high.