On the performance and use of dense servers

IBM Journal of Research and Development, Sep-Nov 2003 by Felter, Wesley M, Keller, Tom W, Kistler, Michael D, Lefurgy, Charles, Et al

Dense servers trade performance at the node level for higher deployment density and lower power consumption as well as the possibility of reduced cost of ownership. System performance and the details of energy consumption for this class of servers, however, are not well understood. In this paper, we describe a research prototype designated as the Super Dense Server (SDS), which was optimized for high-density deployment. We describe its hardware features, show how they challenge the operating system and middleware, and describe how we have enhanced its software to handle these challenges. Our performance evaluation has shown that dense servers are a viable deployment alternative for the edge and application servers commonly found at conventional Web sites and large data centers. Using industry benchmarks, we have shown that SDS outperforms a comparable traditional server by almost a factor of 2 for CPU-bound electronic commerce workloads for the same space and roughly equivalent power budget. We have observed the same advantage in performance when SDS is compared to the alternative solution of virtualizing a high-end server to handle "scaled-down" workloads. We have also shown that SDS offers finer power management control than traditional servers, allowing higher energy efficiency per unit of computation. However, for high-intensity Web-serving workloads, SDS does not perform as well as a traditional server when many nodes must be configured into a cluster to provide a single system image. In that case, the limited memory of each SDS node reduces its performance scalability, and a traditional server is a better alternative. We have concluded that until technology advances allow denser packaging of memory or more efficient use of memory across nodes, the best performance and energy efficiency can be obtained by heterogeneous deployment of both traditional high-end and dense servers.

1. Introduction

Recently, several companies have begun to offer dense servers containing components designed for mobile computing systems [1-4].1 Typically, such servers contain a low-power x86 processor, designed for mobile computers, generally clocked at frequencies between 300 MMz and 1 GHz, with memory ranging from 128 MB to 1 GB, up to 40 GB of disk storage, and one to three 100-Mb/s Ethernet connections. Some of these servers are configured as single-purpose appliances, while others are intended for general use. However, little is understood about their performance, their energy efficiency, or the application classes for which they are best suited. For instance, there is a perception that the limited processing resources available on these servers invariably lead to poor performance compared with high-end servers. Additionally, their energy efficiency is difficult to quantify given the lack of standard and agreed-upon metrics. This paper addresses these issues, with a focus on the performance, energy efficiency, and class of applications for which these servers are well suited. We also describe the additional operating system and system management support that is needed to overcome some of the perceived performance and management problems with dense servers.

In Section 2, we provide background material and describe our prototype implementation of a dense server. Section 3 describes the necessary operating system and system management support, and Section 4 provides the bulk of the quantitative performance analysis. In Section 5, we provide a further, qualitative analysis of our research prototype. We discuss future work in Section 6 and related work in Section 7. Section 8 concludes the paper.

2. Dense servers

The case for energy efficiency in servers

There is a growing industry trend to outsource computing services to large data centers accessible through the Internet. These data centers use economies of scale to amortize the cost of ownership and system management over a large number of servers-typically hundreds or thousands, densely packed to maximize floor space utilization, providing the customer with a more cost-effective approach than the alternative of operating the same services in-house.

Large-scale deployment, however, pushes the limits of power-supply and cooling systems in data centers. Power and cooling costs are already a significant portion of total operating costs-as high as 25% for some data centers. Intermittent system failures due to insufficient cooling in densely packed data centers also add to operating costs. Furthermore, in many data centers, the power supply to the server racks is a key inhibitor to increasing server density, with a practical limit ranging from 5 to 7 kW per rack. This amount of power is often insufficient to allow the rack to be fully populated with servers, thus exacting a loss of revenue due to under-utilization of space. Dense servers solve these problems.

Built using components designed for low-power operation, dense servers are suitable for environments that require a high level of deployment density without violating power-supply limits or generating excessive heat. They also enable sophisticated cluster-based power management techniques at a fine-grained level of control [1, 6-8], further reducing energy consumption. While the use of low-power components such as mobile processors implies a reduction in the computing resources at each server node, the resulting reduction in power and heat enables more aggressive packaging that increases the density of server deployment for a given volume and power budget. Thus, dense servers trade the level of performance at each server node for a larger number of servers that can be deployed within the same space. In Section 4, we examine the impact of this tradeoff on system performance and power consumption.