三維同質多核心處理器之熱限制下任務排程研究

Abstract

相比於傳統二維積體電路，三維積體電路提供了許多優勢，尤其是三維多核心處理器更是大幅提昇電路的效能表現。然而，三維多核心處理器的高功耗密度與散熱路徑的匱乏，導致三維多核心處理器面臨嚴重的熱議題。針對三維多核心處理器的熱議題，需共同考量三個面向，分別是熱效應，能量與產能。本論文主要包含線下(非即時)與線上(即時)兩方面研究，分別從線下與線上之結構系統層次的研究，針對三維多核心處理器的熱議題，提出熱效應，能量與產能的最佳化。本論文研究與摸擬電壓與頻率調節對三維多核心處理器之熱效應，能量與產量的影響，並分別提出線下與線上作業系統層級的任務排程演算法，為三維多核心處理器在熱限制條件下，達成耗能最小化與產能最佳化。

The implementation of chip designs in three dimensions (3D) offers several advantages over traditional two-dimensional chip implementation, such as shorter interconnect delays and better performance. 3D multi-core processors (3D-MCPs) have the potential to significantly improve system performance, but they are more likely to exhibit severe thermal problems due to their high power density and lack of heat dissipation paths. The thermal challenges in 3D-MCPs require a joint assessment of performance, energy, and temperature trade-offs. This thesis consisting of an off-line work and an on-line work focuses on optimizing the system performance, energy efficiency, and temperature of 3D-MCPs. The performance, power, and thermal behaviors of the 3D-MCPs are modeled to analyze how voltage-and-frequency assignments impact the thermal control. Moreover, novel OS-level task-scheduling algorithms for 3D-MCPs to optimize the performance and energy under thermal constraints are proposed.

First, a novel architecture-level algorithm to control over-heating and large temperature variations while optimizing the system performance and minimizing energy consumption is proposed in the off-line work. A key challenge in 3D technology is that too much heat is generated from the internal active layers because the power density per unit volume is drastically increased in 3D-MCPs. Additionally, 3D-MCPs consume much more energy, which complicates low-power design implementations. To handle the severe thermal problem and energy consumption issue in 3D-MCPs, architecture-level solutions to optimize the system performance and minimize energy consumption in 3D-MCPs under thermal constraints are implemented. More specifically, the unique thermal behaviors in 3D-MCPs are studied and a novel layer-by-layer task-to-core mapping strategy to balance the temperatures among the cores in 3D-MCPs is developed. To further optimize the performance optimization and minimize the energy consumption, a thermal-and-energy-aware voltage scaling strategy is proposed to control thermal emergencies in 3D-MCPs.

Second, our work is further implemented in on-line 3D-MCP systems by proposing an on-line task scheduler with a machine-learning model to dynamically assign operating voltages and frequencies to control temperature increase and improve system performance. Mitigation of the thermal issue by utilizing machine learning techniques to identify regions of temperature increase curves is developed in the on-line work. Two different thermal regions of 3D-MCPs are discovered and different key features of these regions are extracted. A machine-learning model with these key features to predict the thermal behavior and the best operation mode of 3D-MCPs is built. This work breaks away from the common assumption of initially static operation modes in 3D-MCP systems and to develop pre-emptively dynamic operation-mode assignment with machine-learning models to handle thermal issues and optimize system performance. Furthermore, a new vertical-grouping voltage scaling (VGVS) strategy that considers thermal correlation in 3D-MCPs is used to handle thermal emergencies. The experimental results demonstrated that the proposed on-line task scheduler reduced hotspot occurrences by over 50%, improved performance by 30% and saved energy by 40% in 3D-MCPs.

The thesis focuses on the study to mitigate the thermal and energy issues associated with 3D-MCPs. The unique thermal behaviors of 3D-MCPs are extensively analyzed and both off-line and on-line architecture-level solutions to effectively control thermal issues while optimizing throughput and energy consumption are developed. This study increases the understanding of the thermal behaviors of 3D-MCPs under different operation modes, allowing increased control of thermal issues. As a result, the proposed design techniques successfully suppress hotspot occurrences, optimize system performance, and minimize energy consumption for 3D-MCPs under thermal constraints.