XPS: EXPL: FP: Symbiotic Power Management for Integrated CPU - GPU Platforms

功耗是移动计算设备运行的主要制约因素。通过降低功耗，将有可能延长电池供电设备的运行时间。现代计算设备将中央处理器（cpu）和图形处理器（gpu）集成在同一个芯片上，以满足各种软件应用程序的性能需求。这些软件应用程序通常编写为执行利用cpu和gpu的并行线程。与分立的CPU-GPU系统相比，集成的CPU-GPU设备提供了许多优势，包括更高的性能、更低的功耗以及CPU和GPU之间更细粒度的交互。本提案的首要目标是探索共生电源管理的新范例，该范例利用CPU-GPU设备的并行编程语言的结构和语义来提供改进的电源管理功能。为了实现这一目标，PI将首先直接从并行编程结构中确定软件应用程序的操作阶段，然后使用确定的阶段来设计自适应电源管理方法，以确定最佳资源（例如，频率电压设置和内核数量），以满足性能目标，同时大幅降低功耗。为配合总体目标，PI将推行一项教育外展计划，其中包括：扩大针对当地高中生的正在进行的项目，以增加他们对计算机系统中能源挑战的认识；开发高性能计算系统课程；让本科生参与研究经验；向工业拓展。

Power consumption is a major constraint on the operation of mobile computing devices. By reducing power consumption, it will be possible to prolong the operational time of battery operated devices. Modern computing devices integrate Central Processing Units (CPUs) and Graphic Processing Units (GPUs) on the same die to cater to the performance needs of various software applications. These software applications are usually written to execute parallel threads that make use of the CPUs and the GPUs. Integrated CPU-GPU devices offer many advantages over discrete CPU-GPU systems, including higher performance, reduced power consumption, and finer-grained interactions between the CPU and GPU.The overarching goal of this proposal is to explore a new paradigm for symbiotic power management that exploits the structure and semantics of parallel programming languages for CPU-GPU devices to deliver improved power management capabilities. To achieve this goal, the PI will first identify the operational phases of software applications directly from their parallel programming constructs, and then use the identified phases to devise an adaptive power management method that determines the optimal resources (e.g., frequency-voltage settings and number of cores) to meet performance targets while offering large reductions in power consumption. In synergy with the overarching goal, the PI will pursue an educational outreach program that includes: expanding on-going projects targeted for local high-school students to increase their knowledge of energy challenges in computing systems; developing courses on high-performance computing systems; involving undergraduate students in research experiences; and outreach to industry.