PITA Fiscal Year 2008 Projects - Environmental Technologies

Energy Management for Data Centers

Principal Investigators: Bruno Sinopoli, Bruce Krogh, Greg Ganger

The project aims at reducing power consumption in Data Centers, improving cooling techniques by means of coordinating traditional HVAC regulation with computational load balancing using advanced distributed control and sensor network technology. According to an EPA report published in August 2007, data servers account for nearly 1.5% of total electricity consumption in the U.S. at a cost of approximately $4.5 billion per year. At current trends without intervention, electricity usage could nearly double by 2011. The amount of peak load these systems put on the grid was around 7gigawatts in 2006, destined to reach 12 gigawatts by 2011, about the power generated by 25 baseload power plants. In addition to reducing operating costs, decreasing data center power consumption will reduce risk of power outages by lowering peak power demand and dramatically decrease carbon dioxide emissions.

The main motivation for this project comes from the observation that the temperature distribution in a data center is not uniform. Temperature gradients depend upon the geometry of the center, the position of racks relative to the cooling units, and the amount of computation not being evenly distributed. These effects will create hot spots, i.e. areas where temperature is significantly above the average. Currently, cooling units will service critical regions by increasing the cooling effort. In turn most of the remaining areas will have temperatures well below the critical operational point. In other words, they will be cooler than they need to be. Significant savings could be achieved if temperature could be made uniform across the center. In order to accomplish such a task we will adopt the following dual strategy. The first is to perform load balancing of the computation within the racks with the objective of reaching constant temperature in addition to satisfying computational demand. The second will encompass the use of sensor network technology to monitor humidity and temperature conditions at a fine grain of spatial granularity so that advanced control techniques can be used to manage the cooling system in a distributed fashion.