Reducing energy consumption of single and multiple processors core systems using dynamic voltage scheduling

Dynamic voltage scheduling has been proven to be one of the most effective ways to reduce the power consumption of digital systems. Most of the work done previously in this area focused on single-processor-core systems. Therefore task allocation was usually not considered during the voltage-scheduling phase. In the work done in regard to this thesis, we tackled the problem of minimizing the energy consumption of both single and multiple processor-core systems. We investigated both the offline and online phases of the voltage scheduling problem. For the multiple-processor core systems, we developed a method which carries out task allocation on processors, task scheduling and voltage assignment for each task concurrently. The problem was formulated as a Mixed Integer Nonlinear Programming (MINLP) optimization problem and optimal solution can be obtained. A heuristic algorithm is implemented to reduce the solving time. We also studied the impact of the offline scheduling decisions on the potential energy saving using the slack time during the on-line dynamic voltage scheduling. Two design strategies were proposed for early consideration in the static phase. One method involves re-ordering the tasks by taking into account the statistical information such as the mean of the execution time of each task. The objective is to create, on average, as much slack as possible during the task execution. The other method involves considering the average energy consumption in the off-line static phase. This provides the minimum average energy after the online slack distribution. During the online voltage-scheduling phase, the energy consumption is further reduced by optimally re-distributing the slack time and re-assigning the voltage for each task by estimating the execution cycle of the current task. The correlations between the tasks are considered and this provides a more accurate prediction of the properties of the incoming tasks. Experimental results show that large energy reductions are achieved using the exact approach with reasonable solving time for small sized systems. For medium to large systems, a near-optimal solution can be obtained in a relatively short solving time. It is also shown that significant improvement can be made when the real-time characteristic of the tasks are catered for early in the static phase