项目名称: 面向众核体系结构空间特征感知的热管理研究
项目编号: No.61303029
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 自动化技术、计算机技术
项目作者: 袁景凌
作者单位: 武汉理工大学
项目金额: 25万元
中文摘要: 随着计算机系统规模增长及"大数据"各类应用,大量异构应用程序和线程并发执行,使性能需求不断提升,能耗问题日益突出。传统体系结构分析方法,以单纯细节模拟和统计分析为主,不适应大规模异构众核体系结构的分析需求。为了更全面、更智能地分析和利用体系结构特征,我们提出将大规模众核系统复杂负载的热行为映射成2D空间特征,并采用机器学习等方法进行特征分类和预测,以便更直观地指导资源优化和热管理。主要包括三项内容:(1)研究能获取丰富信息、可扩展的空间特征挖掘和分析方法,以较低开销有效地获取并分类典型的热空间模式;(2)研究基于机器学习的空间特征智能预测模型,快速重构相应配置众核系统复杂负载的热足迹;(3)研究空间特征感知的资源优化和热管理方法,建立热冲突最小化模型,并利用反馈和学习机制,全局协作地调度任务和进行动态热管理。研究成果将对提升复杂众核系统分析和评估的有效性,节省大规模系统能耗有重要意义。
中文关键词: 绿色计算;全局协作资源优化;热管理;机器学习;GPU
英文摘要: Faced with the ever-growing scale of computing systems and the rapid adoption of "big data" applications, many-core processors today need to handle concurrent execution of a large number of heterogeneous workloads/threads. As a result, there is a heightening demand of system characterization methodology for understanding the key tradeoff between high performance and energy efficiency. Conventional characterizing methods mainly focus on small-scale cycle-accurate architecture simulation and statistic performance analysis. These approaches typically ignore the spatial diversity across many cores, and therefore fail to keep pace with the scale of future multi-core integration. In order to achieve a comprenhensive analysis and intelligent application of many-core architecture characteristics, we propose a novel characterization methodology which captures the geospatial characteristics of many-core processore energy/thermal behavirs. Our approach innovatively leverages machine learning techniques to guide the dynamic resouce management across large-scale many-cores. Some highlights of this proposal include: 1) we investigate informative and scalable spatial characteristics mining and classifcation methods for understanding spatial patterns of processor thermal bethavior, 2) we study machine learning based predictive
英文关键词: Green computing;Global collaborative resource optimization;Thermal management;Machine learning;GPU