CSR: Medium: Optimal Control of Approximate Computing Systems

计算越来越受到能量的限制，无论是低端的移动设备，如手机，还是高端的大型数据中心。因此，减少在计算中消耗的能量是当今计算机科学面临的最重要的问题之一。通过采取计算捷径，例如不执行程序的某些部分或以较低的精度执行它们，在许多应用中（例如视频处理）可以在不显著影响输出质量的情况下减少能量需求。Capri项目的目标是构建一个系统，该系统可以以这些方式优化程序的能耗，同时保证输出质量保持在程序员指定的某个范围内。有两个主要问题必须解决：(i)建立模型来表征程序的输出质量和能源行为，以及（ii）使用这些模型来解决约束优化问题，以确定如何使用近似。Capri将使用最先进的机器学习技术来构建输出质量和能源行为的程序模型，并将采用现代非线性优化算法来解决约束优化问题。该项目将生产可扩展的开环和闭环控制系统，以优化能源效率的应用。该项目包括在诸如并行编程原理与实践（PPoPP）和高性能计算机体系结构（HPCA）等会议上提供有关信息高效计算的教程，并将这些材料纳入课堂并使其公开可用。Capri项目产生的所有数据和产出将在这个网站上提供：http://iss.ices.utexas.edu/。

Computing is increasingly constrained by energy both at the low end in portable devices like cell phones and at the high end in large-scale data centers. Therefore, reducing the energy expended in computation is one of the most important problems facing Computer Science today. By taking computational shortcuts such as not executing certain portions of the program or executing them with lower accuracy, it is possible in many applications, for instance video processing, to reduce energy requirements without substantially affecting the quality of the output. The goal of the Capri project is to build a system that can optimize the energy consumption of a program in these ways while guaranteeing that the output quality stays within some bound specified by the programmer.There are two major problems that must be solved: (i) building models to characterize the output quality and energy behavior of a program, and (ii) using these models to solve a constrained optimization problem to determine how approximation can be used. Capri will use state-of-the-art machine learning techniques to build program models for output quality and energy behavior, and will employ modern non-linear optimization algorithms to solve the constrained optimization problem. The project will produce scalable open-loop and closed-loop control systems for optimizing applications for energy efficiency. The project includes presenting tutorials on information-efficient computing at conferences like Principles and Practice of Parallel Programming (PPoPP) and High-Performance Computer Architecture (HPCA), and incorporate this material into classes and make it publicly available. All data and outputs produced by the Capri project will be made available at this website: http://iss.ices.utexas.edu/.