深度融合软硬件节能原理的绿色驱动异构调度研究

Green real-time computing of processors in cloud is new demand for environmental protection and sustainable development. However, there are bottlenecks in the field. On the one hand, new optimization ways of energy efficiencies of heterogeneous systems are required in scheduling software; on the other hand, hardware energy-saving space also needs appropriate extension. Some main contents in the project are as follows. Firstly, the nonlinear characteristics of CMOS circuits in cluster processors are vectorized with variable metrics via three stages (developing independently test-configurations, analog acquisition of circuit signals, regression and integration of real-time data); then the dynamic energy of cloud heterogeneous-computing systems in continuous time will be mathematically defined to overcome the research defect such as just staying in the qualitative analysis or description. Secondly, heterogeneous real-time scheduling driven by green demand is multi objectively modeled by infiltrating quantification methods of processor dynamic-energy into energy budgets for cloud deployment, while practical standards for energy efficiency evaluation are effectively implanted in the resource management middleware. Thirdly, based on data structures of parallel dependent tasks, and starting with feature encoding/decoding of bionic solution to support graph sorting by depth or coupling strength values, an intelligent cloud-scheduling algorithm is presented oriented to collaborative CPU-GPU hyper architectures. The three supplement one another and none of them is dispensable. In conclusion, the project is of great importance to laying the theoretical or technological foundation of a linkage or even higher integration-efficiencies between energy-saving software and hardware in cloud models.

云处理器绿色实时计算是环境保护、人类可持续发展新需求。然而目前该领域却存在研究瓶颈：一方面，软件调控亟待异构能效优化新途径；另一方面，硬件级节能空间需要合理延展。本课题攻关内容如下：(1)、拟历经实验配置自主研发等三个阶段，突破集群处理器CMOS非线性电路特征的变尺度向量化，进而提出云异构计算体系连续时间内动态能耗数学定义，以克服此类探索目前大多还仅停留在定性分析与描述上研究缺陷。(2)、将处理器动态能耗量化法渗透到云调度部署的能耗“预算”中，建立绿色驱动异构实时调度多目标优化模型，并同步实现能效测评科学依据及可行标准在云资源管理中间件中的有效植入。(3)、基于并行依赖任务集数据结构，从支持图深度、耦合强度值排序的仿生解个体特征编、解码入手，设计适于CPU-GPU协同超体系的云智能调度算法。三者相辅相成，缺一不可。本项目为实现云模式中软、硬件节能联动，或提高其融合实效，奠定理论和技术基础。

本课题探索融合软、硬件节能原理的绿色驱动异构调度理论与方法，为面向海量数据密集应用中的高效能计算奠定相关理论及可行性技术基础;具体成果总结如下。. (一) 实现云异构处理器CMOS电路的动态能耗量化.本研究遵循电路与电子学等相关硬件原理，历经实验配置自主研发、电路信号模拟采集以及实时数据回归集成三个阶段，突破云处理器CMOS电路特性的变尺度向量化，进而提出集群异构体系连续时间内动态能耗的数学定义，以克服此类探索目前大多还仅停留在定性分析与描述上的研究瓶颈，并为软、硬件节能原理相互渗透、联动衔接奠定理论和技术基础。.(二) 提出绿色驱动的异构实时调度多目标优化模型.异构调度，鉴于资源复杂性、环境多样性、应用新需求和目标折衷性等，目前虽在性能不降低情况下对静态功耗减少有所突破，但远不满足面向密集数据的动态实时应用需求。本课题将CMOS电路的动态能耗量化法纳入到调度部署的能耗“预算”中，提出绿色驱动多目标优化实时调度模型，为云模式实现高性能、低功耗及强扩展性建立科学而实用的计算依据。.(三) 设计面向CPU-GPU协同超体系的云智能调度算法.目前多数元启发式算法对于高维目标优化问题存在种群多样性保持不够或收敛速度过慢等缺陷。本研究基于动态依赖任务集存储结构，采用支持DAG深度、耦合强度排序的仿生解个体编、解码规则，提出面向CPU-GPU协同粗粒度与主从式模型共融的多层次并行智能调度算法，为适应环境多样性以及计算服务指标折衷性等新常态提供高效求解理论和方法。.大量仿真实验结果表明:本课题异构调度方法在时间、能效及可扩展等方面优势明显;对于高维实例,整体性能改善分别达到8%,12%和14%。.概言之，随着硬件云在科学计算、信息服务和数据(或事务)处理等方面优势日益显现，集群异构处理器的绿色实时计算与环境保护关联密切，对人类可持续发展影响深远。同时，本研究属于异构集群体系、云调度中间件和分布式人工智能等多学科交叉方向；课题为推进云模式中软、硬件节能联动，或提高其衔接实效，探索新的道路，奠定重要理论和技术基础。