大规模异构并行系统软件功耗优化与管理方法研究

Green-computing is one of the hottest topics in high performance computing fields. Controlling system power consumption and improving system power efficiency are the essential way for larger-scale supercomputer. Heterogeneous parallel system has become one of the most important trends for HPC system. Compared with traditional homogeneous parallel system, heterogeneous parallel system with special accelerator could achieve much higher peak performance and peak power efficiency. However, owing to the differences of processing speed and power consumption between heterogeneous processors, existing low-power optimization methods targeted for homogeneous system could not be efficiently applied onto heterogeneous system. Along the system hierarchy, this proposal focuses on the power optimization and management methods for accelerator, node and system levels. In order to achieve the hierarchical system power management framework, our research group especially focuses on the following topics, coordinating processing and memory units to lower power consumption for GPU accelerator, power-aware parallel loop schedule, power-constrained parallel loop schedule, hardware offload based MPI collective low-power optimization, control theory based system-level power management. The result of this study will help to advance the state-of-the-art in the research of low-power in heterogeneous parallel system, improve the power efficiency of the homemade HPC system, and have significant theoretical and practical impact on the development of low-power study.

绿色计算是当前高性能计算领域最为关注的话题之一，控制系统功耗、提高系统效能是维持超级计算机向更大规模发展的必要途径。异构并行系统已成为当前高性能计算机系统发展的重要趋势之一，与传统同构并行系统相比，集成有专用加速部件的异构并行系统具有更高的峰值计算速度和峰值效能。然而，由于异构处理器间计算速度和功耗开销的不同，已有的面向同构系统的功耗优化方法难以高效应用于异构系统。本项目拟从异构并行系统层次构成出发，依次研究部件级、结点级以及系统级三个层次的功耗优化和管理技术，重点突破GPU计算与存储协同低功耗优化、功耗感知的并行循环调度、功耗受限的并行循环调度、MPI聚合通信功耗优化、基于控制理论的系统功耗管理等关键技术，形成层次式系统功耗管理框架。本项目的研究成果有助于推进异构并行系统低功耗研究进展，提高国产高性能计算机系统运行效能，具有重要的研究意义与应用价值。

绿色计算是当前高性能计算领域最为关注的话题之一,控制系统功耗、提高系统效能是维 持超级计算机向更大规模发展的必要途径。异构并行系统已成为当前高性能计算机系统发 展的重要趋势之一,与传统同构并行系统相比,集成有专用加速部件的异构并行系统具有 更高的峰值计算速度和峰值效能。然而,由于异构处理器间计算速度和功耗开销的不同, 已有的面向同构系统的功耗优化方法难以高效应用于异构系统。本项目拟从异构并行系统 层次构成出发,依次研究部件级、结点级以及系统级三个层次的功耗优化和管理技术,重 点突破GPU计算与存储协同低功耗优化、功耗感知的并行循环调度、功耗受限的并行循环 调度、MPI聚合通信功耗优化、基于控制理论的系统功耗管理等关键技术,形成层次式系 统功耗管理框架。本项目的研究成果有助于推进异构并行系统低功耗研究进展,提高国产 高性能计算机系统运行效能,具有重要的研究意义与应用价值。