POW Solutions: Increasing the Power Efficiency of Large Server Farms

According to a recent EPA report, it is projected that by 2011 IT data centers will consume over 100 billion kilowatt hours of electricity, resulting in an estimated power bill totaling a staggering $7.4 billion.  With an upward trend in power prices and increased demand for IT services, power costs are currently about 12% of all data center spending, and expected to rise further, according to IT research firm Gartner.  

The power problem is particularly pronounced in large CPU-intensive server farms, composed of tens of thousands of servers. Constraints on the amount of available power often limit the number of servers that companies can deploy to a number that is below demand requirements. Ultimately these serious inefficiencies, along with the rising cost of power, will increasingly impact companies’ ability to scale their businesses, satisfy their customers, and optimize their operational efficiencies. 

POW has developed optimal power and workload allocation policies to get more performance from the available power. Furthermore, POW has recently developed novel algorithms that can dynamically adjust server capacity depending on incoming demand, thereby increasing server utilization, and decreasing the amount of power that is being wasted by idle machines, all without compromising on performance.  POW’s novel approach merges advanced queueing, optimization, and stochastic processes’ theories to produce new, non-conventional analytical models.  Results from POW’s work have demonstrated up to a 30% decrease in response time (for the same amount of power) or up to a 30% power savings (with negligible impact on job response time). 

POW is currently developing a software prototype of the above mentioned research for large scale testing.   A company is expected to be spun out of CMU in the spring of 2011.

See Pow Poster (790 KB)

Pow! Team:

• Mor Harchol-Balter, Associate Professor, Computer Science Department – Education, Carnegie Mellon
• Anshul Gandhi, Ph.D. student, Computer Science Department, Carnegie Mellon
• Varun Gupta, Ph.D. student, Computer Science Department, Carnegie Mellon
• Timo Mechler, 2010 MBA Graduate, Tepper School of Business, Carnegie Mellon

News

POW Solutions reaches MIT Clean Energy Prize Semi-Finals
POW Solutions  is building an enterprise software that will manage power consumption of servers in IT data centers.  The software is based on cutting edge power management algorithms developed in Dr. Mor Harchol-Balter’s research group in the School of Computer Science at Carnegie Mellon.
Check out minute 16:50 to hear the announcement.

February 25, 2011 - Pittsburgh BusinessTimes
Pittsburgh a nurturing place for energy start-ups, spin-outs