Linear Programming for Optimal Power Control of Data Center Computers

In a modern data center the cost of power, for computers and air conditioning, can be more than the cost of the computer hardware. Modern computers have a variety of power states with different power vs. response time tradeoffs: off, sleep, hibernate, etc. With thousands of computers in a typical data center it is challenging to determine what power state each computer should be in at any moment in order to minimize power while maximizing responsiveness. I developed an algorithm which breaks the problem into two pieces: predicting future demand and determining power state transitions to minimize power while meeting demand in the best way. Any prediction scheme can be used but in our first implementation we used simple linear prediction. The optimal power state transitions are computed with linear programming. In the general case this is an integer, rather than a linear, programming problem, but a novel representation of the system allows linear programming to be used, while guaranteeing integer results. This makes the algorithm very fast even for data centers with tens of thousands of computers. Our evaluation on three very different data center workloads shows that the energy savings are close to optimal, saving 96%-99.5% of the maximum possible.