超级计算系统预算约束的可靠性与能耗动态融合任务调度策略

The amount of energy needed to operate Supercomputer systems increases regularly since some years at a high pace, and the energy consumption has attracted a great deal of attention. Moreover, high energy consumption inevitably contains failures and reduces system reliability. However, there has been considerably less work of simultaneous management of system performance, reliability, and energy consumption on Supercomputer systems. In this project, we first aim at the Supercomputer large-scale and heterogeneity, and propose an architecture based on energy consumption angle of view. Then, we take the petroleum domain seismic imaging prestack reverse time migration processing as an example of large-scale complex high performance parallel application, and use structured and unstructured grids.to partition their data set dynamically, and propose a multi-constrained dynamic task partitioning method based on Supercomputer architecture. Third, aiming at the mutual dependence and restriction relation of system energy consumption, room temperature, reliability and task execution behavior relation, we propos a application execution reliability and energy consumption dynamical combination theory. Fourth, we use economics non-cooperative game equilibrium theory to compute application execution cost. Finally, we propose a budget restraint reliability and energy consumpiton dynamical combination multi-objective Pareto optimal task scheduling theory. The project research results will enrich the basic theory about Supercomputer systems energy consumption management, reliability assurance, task scheduling, and so on, and also provide a feasible method to improve the efficiency of Supercomputer application.

近年来，超级计算系统在社会经济、科学技术和国防安全中的应用日益广泛。然而，超级计算机固有的可靠性墙、能耗墙、并行墙等问题一直制约着高性能计算的应用效率。项目首先针对超级计算机的大规模性与异构性，提出能耗视角的体系结构，采用DVFS、任务整合与资源休眠机制建立能耗计算模型；然后，以石油领域地震成像叠前逆时偏移处理为复杂高性能计算应用实例，采用结构化网格和非结构化网格复合技术对数据集进行动态剖分，提出适合超级计算机体系结构的多约束动态任务划分方法；其次，研究能耗、温度、可靠性与任务执行行为关系，提出应用程序执行可靠性和能耗动态融合理论；最后，采用经济学非合作博弈均衡理论计算应用程序执行成本，提出成本预算约束下的可靠性与能耗动态融合多目标帕累托优化调度理论。项目研究成果将丰富超级计算系统能耗管理、可靠性保障和任务调度等方面的基础理论，为提高超级计算机应用效率提供一种可行方法。

近年来，大规模计算系统在社会经济、科学技术和国防安全中的应用日益广泛。然而，计算机系统固有的可靠性墙、能耗墙、并行墙等问题一直制约着高性能计算的应用效率。项目基于DVFS功耗技术提出元器件电压/频率与可靠性关系，并采用最小二乘曲线拟合法获得其关系模型。其次，项目组系统研究元器件能耗与系统能耗、能耗与温度、温度与可靠性关系，建立系统能耗、温度与可靠性相互关联依赖模型。最后，我们利用经济学非合作博弈均衡理论计算应用程序执行成本，采用遗传算法实现预算约束下的能耗、性能与可靠性动态融合多目标帕累托最优任务调度算法。项目研究成果将丰富计算系统能耗与可靠性约束关系、任务调度等方面基础理论，为提高其应用效率提供一种可行方法。