Empirical evaluation of power saving policies for data centers
Michele Mazzucco, Isi Mitrani · ACM SIGMETRICS Performance Evaluation Review · 2012
It has been suggested that the conflicting objectives of high performance and low power consumption in a service center can be met by powering a block of servers on and off, in response to changing demand conditions. To test that proposition, a dynamic operating policy is evaluated in a real-life setting, using the Amazon EC2 cloud platform. The application running on the cluster is a replica of the English edition of Wikipedia, with different streams of requests generated by reading traces from a file and by means of random numbers with a given mean and squared coefficient of variation. The system costs achieved by an 'optimized' version of the policy are compared to those of a simple heuristic and also to a baseline policy consisting of keeping all servers powered on all the time and one where servers are re-allocated periodically but reserves are not employed.