| Multi-core processors have emerged as a solution to the problem of ever-increasing demands for computing power. However, higher power dissipation levels resulting into thermal problems and increasing cooling costs are major factors limiting their scalability into larger systems. Therefore, dynamic thermal management (DTM) and dynamic power management (DPM) of multi/many-core systems have emerged as important areas of research.;The existing resource management approaches are either energy-aware or thermal-aware. In this dissertation, we focus on a new problem of simultaneous performance (P), energy (E), and temperature (T) optimized scheduling (PETOS) for allocating tasks to multi-core systems. To allocate a set of parallel tasks to a set of cores, we propose a static multi-objective evolutionary algorithm (MOEA)-based task scheduling approach for determining the Pareto optimal solutions with PET-optimized schedules defining the task-to-core mappings and the corresponding voltage/frequency settings for the cores. Our algorithm includes problem-specific techniques for solution encoding, determining the initial population of the solution space, and the genetic operators that collectively work on generating efficient solutions in fast turnaround time. We also propose a methodology to select one solution from the Pareto front given the user preference describing the related P, E, and T goals. We show that the proposed algorithm is advantageous in reducing both energy and temperature together rather than in isolation. We also propose a dynamic multi-objective optimization approach that can solve PETOS problem while taking into consideration the task and system model uncertainties.;Another contribution of this dissertation is the design of efficient heuristic algorithms that can generate a set of solutions to the PETOS problem. Central to each heuristic are strategies for task assignment and frequency selection decisions. The proposed set of heuristics includes several iterative, greedy, random, and utility function and model based methods to explore the scheduling decision space for the PETOS problem. We describe and classify these algorithms using a methodical classification scheme. The methods developed in this dissertation obtain multiple schedules with diverse range of values for makespan, total energy consumed, and peak temperature, and thus present efficient ways of identifying trade-offs among the desired objectives for a given application and architecture pair. |
