Minimizing Power Consumption At Module, Server And Rack-levels Within A Data Center Through Design And Energy-efficient Operation Of Dynamic Cooling Solutions
Fernandes, John Edward
MetadataShow full item record
Data center energy consumption continues to increase with the proliferation in online services such as social networking, banking, entertainment, cloud computing, etc. Recent estimates show that associated power consumption accounts for around 2% and 1.8% of electricity production in the United States and worldwide respectively. In addition to growth in the industry, power densities continue to rise in IT equipment increasing the criticality of thermal management in data center operation. Cooling systems account for around a third of overall power consumption requiring more energy-efficient solutions across all levels within the facility.Liquid cooling of high power modules using cold plates has been around since the early 1980s. However, till today, designs remain fairly static and fail to adapt to variations in power dissipation at the device. This can be alleviated through the introduction of a "dynamic" cold plate design. Through implementation of sensing and control, the solution can distribute available resources based on local cooling requirements. A high-power multi-chip module (MCM) platform is chosen as reference for the design of such a solution. In-depth computational fluid dynamics (CFD) analysis is conducted to select appropriate heat transfer surfaces and predict thermal performance of the cold plate, when assembled with the MCM. A cost-effective MCM thermal test vehicle is assembled to enable experimental testing of both dynamic and static cold plates. Components of a liquid cooling test bench and control system are selected and permit future evaluation of thermal performance and energy-efficiency of both solutions.The current trend in air cooling of data centers involves higher ambient temperatures to maximize use of free cooling. However, power consumption at the server-level may increase due to elevated fan activity and CPU leakage current. Minimizing power consumption of web servers (1.5U profile) is achieved by studying the effect of ambient temperature on performance and investigating means to improve chassis fan control through accurate selection of CPU target temperature. Multiple servers are instrumented and deployed in a test bed data center and are subjected to different air supply temperatures and fan speeds to achieve the same. Limits of energy-efficient operation and available savings will be discussed.Fan efficiency is known to increase with size. Departing from conventional server designs, wherein fans are installed within the chassis, and consolidating air moving devices at the rear of a rack or `stack' of servers permits increase in size and cooling efficiency. Preliminary studies have shown that replacing server-enclosed 60mm units with a rear-mounted wall of larger fans (80mm or 120mm) enables savings in fan power of the order of 50%. A methodology for row-wise control of such rack-level fans, with the purpose of simulating an actual product, is previewed and savings are reported. In addition, performance under real-life scenarios such as non-uniform loads and fan failure is investigated. Each rack-level setup has distinct advantages. However, selecting between configurations would necessitate a compromise between efficiency, redundancy and cost.