Dynamic Data Center Power Management: Trends, Issues, and Solutions

In this paper we examine the challenges of increasing data center power consumption and higher energy costs in the face of ever-increasing computing needs. An examination of how power is allocated to computing resources in data centers shows that current methods do not result in optimal use of available data center power and space. We identify requirements that server platforms must address to solve data center power problems, and we offer a solution that includes a platform resident Policy Manager (PM). The PM monitors power and thermal sensors and enforces platform power and thermal policies. We explain how the PM can be used as the basis of a data center power management solution. We present results from a Proof of Concept (PoC) implementation, and we conclude by showing that a policy-based approach is powerful for maximizing power allocation within a given power envelope and increasing server density in data centers. INTRODUCTION One of the biggest challenges for data center operators today is the increasing cost of power and cooling as a portion of the total cost of operations. As shown in Figure 1, over the past decade, the cost of power and cooling has increased 400%, and these costs are expected to continue to rise. In some cases, power costs account for 40-50% of the total data center operation budget. To make matters worse, there is still a need to deploy more servers to support new business solutions (Figure 2). Data centers are therefore faced with the twin problem of how to deploy new services in the face of rising power and cooling costs. In a recent survey of data centers (Figure 3), 59% identify power and cooling as the key factors limiting server deployment. If these trends continue, the ability of data centers to deploy new services will be severely constrained. To overcome this constraint, data centers have three choices: expand power and cooling capacity, build new data centers, or employ a power management solution that maximizes the usage of existing capacity. The first two choices can be very expensive because they involve capital expenditure for purchasing and installing expensive new power delivery equipment. For this reason, the power management approach bears close examination, and this approach is the focus of the rest of our paper. For previous work in this area, the reader is referred to Felter et al. [4] who examine the benefits of dynamic power budget allocation, Femal [5] who examines the benefits of monitoring and coordinating power distribution to achieve higher application throughput, [6] where a framework to monitor power is discussed, and Bianchini [7] who presents a survey of energy management techniques by type of server system. Intel Technology Journal, Volume 12, Issue 1, 2008 Dynamic Data Center Power Management: Trends, Issues, and Solutions 60 Figure 1: IDC Report of data center cost structure and trend Figure 2: Expected growth in server count Figure 3: Factors limiting server growth Our paper is organized as follows. In the next section we describe the current data center power allocation approach and the resulting problems; we follow this by proposing a new method for dealing with data center power allocation and describe the resulting server requirements; we then describe the role and functions of a platform resident Policy Manager (PM) and show how it addresses these requirements. Finally, we present the results of a PM Proof of Concept (PoC) and the benefits it offers data centers. CURRENT POWER ALLOCATION METHODS A typical data center power distribution hierarchy is designed to deliver a fixed amount of power to the room and then to each rack. The challenge of the data center operator is to determine the number of servers for each rack while ensuring that the overall rack (hence room) power consumption does not exceed the limit. To do this, the operator must make certain assumptions about the maximum power consumption per server. For most data centers, there are two ways of determining this: 1) using server nameplate power value, and 2) using a derated nameplate value. The server nameplate value, which is marked on the server by the manufacturer, is the maximum possible power value that the server can consume. Actual power consumption is typically much less that the nameplate power. Most data center operators are aware that typical server power consumption never reaches the nameplate value, and one way for them to increase server density is to derate the nameplate power by a certain percentage— depending on the workload that is deployed on the server. Change in Server Requirements