Operating System Strategies for Energy- and Resource-Aware Adaptation

美国国家科学基金会分布式系统研究CISE/ cnsabstract提案编号：0411127首席研究员：Dwarkadas， sandhyinstitution：罗切斯特大学提案标题：能源和资源感知适应的操作系统策略随着设备功能的增加，能源消耗已经成为移动和服务器级计算机的一个越来越大的问题。硬件设计人员为此开发了一些设备（如硬盘和网络），这些设备在不活动时可以“断电”（进入非操作模式），而其他设备（如处理器组件）可以“缩减”（通过降低电压或将部分重新分配给其他线程），以在其峰值性能和功率下运行。为了有效地利用这些设备，操作系统必须动态地“塑造”所提供的负载。对于“断电”设备，它应该以突发为目标，以最大化低功率间隔的长度和最小化代价高昂的状态转换的数量。对于“缩减”设备，它应该以平滑为目标，利用电源对电压的非线性依赖，避免为未使用的资源付费。对于具有这两种功能的设备，它必须平衡两种方法的好处。现有的和新兴的体系结构，包括单线程、多线程和集群处理器的工作正在进行中。预期的结果包括(1)服务器和移动工作负载在行为可变性方面的特征和(2)调度程序和设备管理策略，这些策略将运行时检测与过去行为的概要文件相结合，以模拟所有运行中的应用程序的需求，并将这些需求与可用资源相匹配。传播将通过出版物、学生、开源软件工具和/或行业合作来实现。更广泛的影响将包括对未来日益复杂的系统的资源意识适应，以提高性能和降低能耗。Brett D. fleisch CISE/ cns2004年6月16日

National Science FoundationDistributed Systems Research CISE/CNSABSTRACTPROPOSAL NUMBER: 0411127PRINCIPAL INVESTIGATOR: Dwarkadas, SandhyaINSTITUTION: University of RochesterPROPOSAL TITLE: Operating System Strategies for Energy- and Re-source-Aware AdaptationWith increases in device capabilities, energy consumption has emerged as a large and growing concern for both mobile and server-class computers. Hard-ware designers have responded by developing devices (e.g. hard disks and net-works) that can be "powered down" (dropped into a non-operational mode) when inactive, and other devices (e.g. processor components) that can be "scaled back" (by reducing voltage or reallocating in part to other threads), to operate below their peak performance and power. To utilize such devices effectively, the operating system must dynamically "shape" the offered load. For "power down" devices it should aim for burstiness, to maximize the length of low-power inter-vals and minimize the number of costly state transitions. For "scale back" de-vices it should aim for smoothness, to leverage the nonlinear dependence of power on voltage and to avoid paying for unutilized resources. For devices with both capabilities it must balance the benefits of the two approaches.Work is addressing both existing and emerging architectures, including single-threaded, multi-threaded, and clustered processors. Expected results include (1) the characterization of server and mobile workloads in terms of behavior variability and (2) scheduler and device management strategies that combine run-time instrumentation with profiles of past behavior to model the needs of all running applications, and to match those needs to available resources. Dis-semination will be achieved via publications, students, open source software tools, and/or industry collaboration. Broader impacts will include the re-source-aware adaptation of future increasingly complex systems for improved performance and reduced energy.Dr. Brett D. FleischProgram Director, CISE/CNSJune 16, 2004..