Summary: | 博士 === 國立臺灣科技大學 === 電機工程系 === 105 === Three-dimensional (3D) multicore processors have been recently developed to resolve the power consumption and interconnection delay problems of embedded systems; however, thermal management has proven to be challenging due to the thermal behavior from vertically stacked cores, and the subsequent trade-off that occurs between performance requirements and overheating. This dissertation investigates thermal management for 3D multicore processors with respect to the above tradeoffs. A novel thermal-aware real-time scheduling framework to schedule dynamic workloads for 3D multicore processors is proposed. We first present the concept of thermal size to manage the heat generated by task executions and to deal with the thermal conduction from vertically aligned cores. Considering homogeneous multicore processors with independent tasks having varied power consumption, a thermal-throttling dispatcher and a thermal-throttling server are proposed to determine the execution frequency for scheduling the arrival task based on its power consumption enabling thermal awareness in all real-time multicore schedulers. When the heterogeneous-ISA multicore processors are considered, to meet the end-to-end deadline of a task, a thermal size ratio detection and run-time thermal budget reclaiming are proposed to manage the heat generated from synergistic execution on heterogeneous-ISA cores with varied power consumption. To ensure that all tasks meet the timing and thermal constraints, the admission controls are then derived for both 3D homogeneous and heterogeneous-ISA multicore processors with corresponding task models. In this dissertation, we made the contributions as follows: (1) The thermal size (budget) concept is first presented to construct the relation between heat generation, thermal conduction, and execution frequency for thermal management. (2) The admission control is first proposed for 3D multicore processors to prevent the thermal constraint violations and guarantee the quality of service of tasks. (3) The proposed thermal management can universally adapt to all 2D/3D homogeneous/heterogeneous multicore processors with existing schedulers. The evaluation results indicate that the proposed framework registered 20-60% improvement in system utilization compared with existing thermal managements while meeting the thermal constraints.
|