Run-time energy-driven optimisation of embedded systems: a complete solution

Abstract

Consumption of power and conservation of energy have become two of the biggest design challenges in construction of embedded systems. Energy is a resource in limited supply, but demands are increasing. Hence, much research is being performed to reduce power and energy usage or optimise performance under energy constraints.

There are very few solutions that try to cater for the applications where the data input is not easily testable before run-time. These applications require an optimisation procedure that knows the power consumption of the system and is able to dynamically optimise operation to maximise performance while meeting energy constraints.

This thesis provides a complete solution to the problem of run-time energy-driven optimisation of application performance. The complete system, from a processor that is able to provide feedback of the power consumption in parallel to execution, to applications that exploit the power feedback to provide dynamic optimisation.

A processor that estimates its own power consumption is designed by the addition of small dedicated counters that tally occurrences of power consuming events which are macro modelled. The methodology is demonstrated on a standard processor achieving an average power estimation error of less than 2% while increasing area of the processor by only 5%.

This enables energy-driven optimisation via application adaptation. Modification techniques and low-overhead algorithms are provided to demonstrate how energy feedback can be effectively used to maximise performance of algorithms within given constraints. Applications’ quality is maximised under given energy constraints using less than 0.02% of the execution time.

Finally, the dissertation discusses the systems used to demonstrate the methodologies and techniques created throughout the research project. These implementations of the energy-driven optimisation system verify the soundness of the methods and applicability of the approaches used.

This is the first time a complete solution for energy-driven optimisation has been shown, from creation of the processor to analysis of software utilising the approach. The methodologies and techniques can be applied to a variety of applications in a range of fields such as multimedia and networking that have never been possible before.