Analysis of an energy proportional data center

Energy proportionality is a desirable property of an energy efficient data center that can be achieved by making servers available on demand, dynamically enabling enough computing capacity to handle the workload. However, reducing the number of running servers can impact job performance and may potentially lead to breaches in the service level agreement. We analyze the optimal (minimum) energy requirement of servers in an energy proportional data center to maintain a selected performance service level objective from the following possibilities: (i) running servers at or below a maximum utilization level; (ii) keeping the average job response time below a given limit; and (iii) limiting the probability of job response times exceeding a turnaround deadline. Performance and power measurements from a real server allow to define realistic parameters for theoretical and simulated models and to obtain realistic results.