Adaptive Software for Extreme-Scale Scientific Computing: Co-Managing Quality-Performance-Power Tradeoffs

本提案试图通过开发自适应软件工具来共同管理质量-性能-功率权衡来解决这两个主要挑战。首先，我们将开发组合和统计自适应技术来动态地选择方法，在产生满足应用程序质量要求的解决方案的同时交付改进的性能。接下来，使用计算和通信成本的注释模型以及稀疏数据访问模式，我们将开发在不损害性能的情况下降低功耗的技术。例如，即使处理器之间相对较小的负载平衡被利用，也可以显著节省电力。这些不平衡可以很容易地达到数万亿个cpu周期的数量级，并且可以通过动态电压缩放（DVS，其中时钟频率和电源电压都被调整）来调整负载轻/负载重的处理器的功耗。更重要的是，由此得出的见解可以引导未来的系统，在这些系统中，功率预算可以有效地针对处理器-内存互连子系统，以提高应用程序性能。我们计划通过在高端多处理器上开发自适应组件软件系统来实现我们的技术

This proposal seeks to address these two primary challenges by developing adaptive software tools toco-manage quality-performance-power tradeoffs. First, we will develop combinatorial and statistical adaptive techniques to select methods dynamically, delivering the improved performance while producing a solution that meets application quality requirements. Next, using an annotated model of computation and communication costs and sparse data access patterns, we will develop techniques for power reduction withoutperformance impairment. For example, power savings can be significant even when relatively minor loadimbalances among processors are exploited. These imbalances can easily be on the order of trillions ofCPU cycles, and power consumption can be tuned through dynamic voltage scaling (DVS, where both theclock frequency and the supply voltage are tuned) for lightly/heavily loaded processors. More importantly,resulting insights can lead to future systems where the power budget is directed effectively over processor-memory interconnect subsystems to improve application performance. We plan to implement our techniques by developing an adaptive component software system on high-end multiprocessors