Data center power and performance optimization through global selection of P-states and utilization rates

The vast majority of Internet services available today rely on the computing capabilities of data centers. The load in a data center varies according to the time of day, week and year, thus it is important to be able to optimize the use of computing resources dynamically and continuously. This paper presents a framework for modeling and simulating the power and performance behavior of data centers, as well as optimization algorithms capable of power and performance trade-offs, considering all cores in a data center simultaneously. The contributions of this paper are threefold. First, it presents a novel concept of Virtual Power States (VPSs) which are combinations of processors power states and utilization rates, which are used for continuous optimization of power and performance. Second, it presents two algorithms (one optimal and one heuristic) for selecting the power states in each core in a data center in order to optimize the overall power and performance. Third, it presents a power and performance simulation model for data centers, and it shows that for a given varying workload, the algorithms can be applied at regular time intervals, change the power states of the CPU cores to adapt to varying workload and power policies, keeping the data center running at near optimal points continuously. Examples using Intel i7-based and AMD Opteron-based web servers are presented.