一种面向异构并行系统的最大功耗管理方法

Peak Power Management Method for Heterogeneous Parallel System

高功耗已成为制约高性能计算机发展的重要问题之一.近年来,大量研究关注于如何在满足系统功耗约束的条件下优化系统执行性能.然而,已有方法大都针对同构系统,未考虑异构处理器之间的功耗或速度差异,难以高效应用于基于加速器的异构系统.对当前异构并行系统执行模型进行了抽象,并提出了融合两级功耗控制机制的系统功耗管理框架,自顶向下依次为系统级功耗控制器和异构处理引擎功耗控制器.在异构处理引擎功耗控制中,针对类OpenMP并行循环,首先分析了异构多处理器在满足功耗约束条件下达到性能最优的条件.基于该结果,给出了功耗受限的并行循环划分算法,该方法通过协调并行循环调度和动态电压频率调节技术以优化异构并行处理.在系统级功耗控制中,建立了异构处理引擎效能评估方法,以此作为功耗划分的依据,在兼顾并发应用公平性的同时,提高系统整体执行效能.最后,基于典型CPU-GPU异构系统验证了方法的有效性.

High power consumption has become one of the top considerations in high performance computing field. Recently, there are many studies focused on optimizing the system performance within a given power budget. However, most existing solutions are explored for homogeneous system without considering the differences in power consumption and processing speed between heterogeneous processors, and therefore could not be adapted for accelerator-based heterogeneous parallel system effectively. This paper first summarizes the execution model of modem accelerator-based parallel system, and then introduces the power control framework consisting of two power management hierarchies, i.e. system-level power controller and heterogeneous processing engine power controller from top to bottom. In the lower level controller, targeted for OpenMP-like parallel loop, the paper first theoretically analyzes the conditions for the maximum performance given a power budget for heterogeneous processors. Based on this result, the paper provides a power-constrained parallel loop scheduling algorithm which coordinates parallel loop partition and voltage/frequency scaling for heterogeneous processers to achieve the optimal performance given a system power budget. In the upper level controller, the paper establishes the evaluation metrics for heterogeneous processing engine to allocate power budget, in order to keep fairness between concurrent applications and improve the whole system efficiency. Finally, the paper evaluates the proposed method in a typical CPU-GPU system.