当前位置 >>  首页 >> 合作交流 >> 学术交流

Designing Energy-Efficient Servers and Data Centers

撰稿: 摄影: 发布时间:2012年08月07日

时间:2012年5月8日(周二)上午10:30-12:00
 

地点:四层报告厅


摘要
The performance, cost, and reliability of modern supercomputers,servers, and data centers are dictated in large part by the effectivemanagement of limited power and thermal budgets. Buildingenergy-efficient systems entails solving problems at multiple systemlevels, e.g., circuits, architecture, system software, and applications.This talk will present a variety research projects underway at IBMResearch - Austin, including: (i) mechanisms that manage the powerconsumed by individual system components (e.g., processors, memory,fans, and power supplies), (ii) storage servers that exploit energy- andreliability-aware flash caching/replication to reduce storage power, and(iii) integrated IT and facilities energy management. The talk willintroduce a set of design principles for building more energy-efficientsystems that IBM researchers have developed based on the insights from adecade-long focus on energy efficiency, and suggest challenges that mustbe addressed to substantially improve the energy-efficiency of futuresystems.

The talk will include a "deep dive" into technology that activelymanages voltage margin to save energy in POWER7 processors. Voltagemargin allows processors to operate correctly during worst-caseconditions, but during typical conditions it wastes power. We developeda mechanism that reduces excess voltage margin by (1) introducing acritical path monitor (CPM) circuit that measures available timingmargin in real-time, (2) coupling the CPM output to the clock generationcircuit to adjust clock frequency within cycles in response to excess orinadequate timing margin, and (3) adjusting the processor voltage levelperiodically in firmware to achieve a specified average clock frequencytarget. We implemented this mechanism in a prototype IBM POWER7 serverand were able reduces the average voltage setting 137-152 mV belownominal during typical conditions, resulting in average processor powerreduction of 24% with no performance loss. This work was awarded BestPaper at ACM MICRO 2011.

主讲人简介
John Carter leads the Data Center Networking research group at the IBMResearch - Austin. From 2008-2010, Dr. Carter led the Power AwareComputing team at IBM Research - Austin that developed technologies toreduce server and data center energy consumption as part of IBM's SmartPlanet initiative. Prior to joining IBM Research, Dr. Carter was theAssociate Director of the University of Utah's School of Computing,where he led research projects in the areas of multiprocessor computerarchitecture, distributed systems, and memory system design (e.g.,Impulse and Khazana).
附件下载: