Premium
Concepts for computer center power management
Author(s) -
DiRienzo A.,
Medeiros John A.,
Whitlock M.,
Wages E.,
Highfield J.
Publication year - 2010
Publication title -
concurrency and computation: practice and experience
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.309
H-Index - 67
eISSN - 1532-0634
pISSN - 1532-0626
DOI - 10.1002/cpe.1542
Subject(s) - computer science , electric power system , predictability , data center , scheduling (production processes) , nonstop , reliability engineering , real time computing , distributed computing , operating system , power (physics) , engineering , operations management , physics , quantum mechanics
Electrical power usage contributes significantly to the operational costs of large computer systems. At the Hypersonic Missile Technology Research and Operations Center (HMT‐ROC) our system usage patterns provide a significant opportunity to reduce operating costs since there are a small number of dedicated users. The relatively predictable nature of our usage patterns allows for the scheduling of computational resource availability. We take advantage of this predictability to shut down systems during periods of low usage to reduce power consumption. With interconnected computer cluster systems, reducing the number of online nodes is more than a simple matter of throwing the power switch on a portion of the cluster. The paper discusses these issues and an approach for power reduction strategies for a computational system with a heterogeneous system mix that includes a large (1560‐node) Apple Xserve PowerPC supercluster. In practice, the average load on computer systems may be much less than the peak load although the infrastructure supporting the operation of large computer systems in a computer or data center must still be designed with the peak loads in mind. Given that a significant portion of the time, systems loads can be less than full peak, an opportunity exists for cost savings if idle systems can be dynamically throttled back, slept, or shut off entirely. The paper describes two separate strategies that meet the requirements for both power conservation and system availability at HMT‐ROC. The first approach, for legacy systems, is not much more than a brute force approach to power management which we call Time‐Driven System Management (TDSM). The second approach, which we call Dynamic‐Loading System Management (DLSM), is applicable to more current systems with ‘Wake‐on‐LAN’ capability and takes a more granular approach to the management of system resources. The paper details the rule sets that we have developed and implemented in the two approaches to system power management and discusses some results with these approaches. Copyright © 2009 John Wiley & Sons, Ltd.