CSR: Small: Cross-layer HW/SW solutions towards Energy-Efficient Datacenters

CSR：小型：面向节能数据中心的跨层硬件/软件解决方案

• 批准号: 1321047
• 负责人: Lingjia Tang
• 金额: $ 30万
• 依托单位: University of California-San Diego
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2013
• 资助国家: 美国
• 起止时间: 2013-10-01 至 2014-02-28
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1321047&HistoricalAwards=false
• 关键词: CSR; Small; Cross; layer; HW

Facilitating workload consolidation and improving server utilization are critical for reducing cost and improving energy-efficiency of modern datacenters. One major challenge for improving utilization is to do so without affecting the quality of service (QoS). On modern servers, various co-located jobs may share critical resources including 1) micro-architectural resources such as last level cache, memory bandwidth, and functional units, etc., and 2) energy resources including grid power and distributed batteries that provide an additional power source especially during high load period. Multiple co-located jobs may contend for shared these resources, causing interference, threatening application QoS or even triggering the circuit breaker and resulting in costly downtime and severe QoS violations.This project addresses these challenges by designing a cross-layer system that effectively manages workload consolidation, quality of service (QoS) and various energy sources to optimize for energy efficient computing. Our system spans several layers, including profiler, static compiler, online lightweight monitoring and prediction, runtime execution management, hardware power state control and energy sources control. The compilation technique identifies and inserts markers in contentious code regions in low-priority applications, as well as critical regions in high-priority applications that require QoS protection. The lightweight runtime utilizes the compiler hints, monitors the QoS, power consumption, etc., and adaptively adjusts the pressure applications generate to the shared resources such as shared cache and memory bandwidth, manages battery (dis) charges and hardware power states to guarantee QoS and achieve efficient power shaving.

促进工作负载整合和提高服务器利用率对于降低成本和提高现代数据中心的能效至关重要。提高利用率的一个主要挑战是在不影响服务质量（QoS）的情况下这样做。在现代服务器上，各种协同定位的作业可以共享关键资源，包括1）微架构资源，诸如末级高速缓存、存储器带宽和功能单元等，以及2）包括电网电力和分布式电池的能量资源，其尤其在高负载期间提供附加的电源。多个位于同一地点的作业可能会争夺共享的这些资源，造成干扰，威胁应用程序的QoS，甚至触发断路器，导致代价高昂的停机时间和严重的QoS违规。本项目通过设计一个跨层系统，有效地管理工作负载整合，服务质量（QoS）和各种能源，以优化节能计算来解决这些挑战。该系统包括分析器、静态编译器、在线轻量级监控和预测、运行时执行管理、硬件电源状态控制和能源控制等多个层次。编译技术识别并插入标记在低优先级应用程序中的有争议的代码区域中，以及在需要QoS保护的高优先级应用程序中的关键区域中。轻量级运行时利用编译器提示，监控QoS、功耗等，并且自适应地调整应用对诸如共享高速缓存和存储器带宽之类的共享资源产生的压力，管理电池（放电）充电和硬件功率状态以保证QoS并实现有效的功率节省。