Tracking the Power in an Enterprise Decision Support System Justin Meza HP Labs Mehul A. Shah HP Labs Parthasarathy Ranganathan HP Labs Mike Fitzner Hewlett-Packard BCS Judson Veazey Hewlett-Packard BCS ABSTRACT Enterprises rely on decision support systems to influence critical business choices. At the same time, IT-related power costs are growing and are a key concern for enterprise ex- ecutives. Yet, there is little work to date characterizing the power use of decision support systems. Towards this end, we present the first holistic measurements and analysis of an audit-class system running the TPC-H decision support benchmark at the 300GB scale. We first provide a break- down of the system’s power use into its core hardware com- ponents. We then explore its power-performance tradeoffs. This investigation shows that there is ample room to im- prove its energy use without sacrificing much performance. Moreover, the most energy-efficient configuration depends on the workload. These results suggest that, going forward, database software has an important role to play in optimiz- ing for energy use. Categories and Subject Descriptors H.2 [Database Management]: Miscellaneous General Terms Measurement, Performance Keywords decision support, power, energy, energy efficiency, TPC-H 1. INTRODUCTION The decision support market is a multi-billion dollar mar- ket and is still experiencing double-digit growth rates [17]. Enterprises use decision support systems to quickly perform complex analyses over large amounts of data whose results are used to inform critical business decisions. As technology improves, these companies are demanding larger, faster, and cheaper systems so they can derive value from data that was too costly to mine in the past. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. ISLPED’09, August 19–21, 2009, San Francisco, California, USA. Copyright 2009 ACM 978-1-60558-684-7/09/08 ...$10.00. At the same time, an important and growing component of the total cost of ownership for these systems is power and cooling [1]. A recent report by the EPA shows data- center power consumption in the US doubled between 2000 and 2006, and will double again in the next five years [16]. Uncontrolled energy use in datacenters also has negative im- plications on density, scalability, reliability, and the environ- ment. These trends suggest that we should optimize decision support systems for energy-efficiency, yet there is little work on understanding the power characteristics of these systems from a whole system perspective. In this paper, we present the first power measurements and analyses of a system con- figured similarly to a performance-optimized, audited TPC- H system at the 300GB scale [3]. Our main contributions are as follows. We first provide a breakdown of the system’s power-use into its core subsystems: CPU, memory, disk, and other mis- cellaneous components (Section 3). Since decision support workloads often need many disks, the storage subsystem, as expected, used more than half of the total power. Inter- estingly, the miscellaneous components in aggregate, such as fans, disk array controllers, supporting chip sets, and so on, comprised 27% of total power. But, the memory DIMMs comprised an unexpectedly small fraction, 7% of total power. We also explore the power-performance tradeoffs that this system exhibits and reveal interesting insights from this in- vestigation. In its peak performing configuration, the hard- ware (and underlying OS) offered little automatic power re- duction between full load and at idle. Instead, we found the most effective way to trade performance for power was by repartitioning the database across fewer disks and turning- off the unused ones. Using this knob for reducing power, we found that the initial performance-optimized configura- tion is grossly over-provisioned; we reduced power use by 45% while sacrificing only 5% of peak performance. Even at smaller memory sizes, we found that the system’s energy effi- ciency is non-monotonic with increasing performance (Sec- tion 4) . That is, there is a point of diminishing returns after which performance continues to improve as disks are added but energy efficiency drops. Moreover, we found that this peak efficiency point varies and depends on the query workload. These results suggest the following. First, industry bench- marks should incorporate energy measurements since sys- tems optimized for performance are poorly balanced for power and, thus, may not reflect current customer needs. More-