Minimizing Energy Consumption For Linear Speedup Parallel Real-Time Tasks

In recent years, with a substantial increase in performance of embedded system, the demand for energy has also been on the rise. Not only does the boosting of energy consumption restrict the growing integration of embedded system, but also it requires sophisticated cooling technology, which also raises the cost of the product. Therefore, how to be more effective energy conservation has become a hot issue in embedded systems research. On the other side, Due to the boosting demand for system performance, multicore has appeared in embedded real-time systems for high performances computing recently, such as robot arm dynamics and video processing. Because of the advantages of multi-core in parallel processing and energy efficiency, energy efficient scheduling for multi-core architecture is becoming an important research direction in the field of embedded real-time systems.While much work has addressed the problem for sequential tasks where each task can run on only one processor at a time, little work has been done for parallel tasks where an individual task can be executed by multiple processors simultaneously. Meanwhile, the model of sequential tasks cannot play the advantages of multi-core technology in parallel processing. So, in this paper, this thesis studied energy minimization problem for parallel task systems with timing guarantees.Power and frequency in actual system correspond to a discrete set of values. However, in order to simplify the model of energy efficiency in real-time system, the frequency value of processor is often assumed to be continuously changed, which is difficult to be used in actual system. Hence, this thesis focused on a system with DVS (Dynamic Voltage Scaling) enabled processors with discrete operation modes and workload satisfying linear speedup ratio model when combining timing constraints and gang scheduling scheme.This thesis first proved a lemma, a sufficient condition for minimizing the system energy, which indicates that the overall system energy is minimized when each task runs on all of the processors simultaneously. Then, adopting the earliest deadline first (EDF) scheduling policy and the lemma proved, this thesis employed a 0-1 Integer Linear Program (0-1 ILP) to derive the optimized frequency assignment with minimized energy consumption. Furthermore, two polynomial-time complexity heuristics with opposite frequency assignment searching directions were also proposed. This thesis conducted extensive simulation experiments and the results show that the proposed heuristics can significantly reduce the system energy consumption and consume nearly the same energy as does 0-1 ILPs.