This grant provides funding to develop models and methodologies for reducing energy consumption in data centers to the maximum extent possible without degrading the quality of service experienced by the users. The models and methodologies are based on integrating stochastic optimization and stochastic optimal control algorithms to determine: a) the optimal set of meta-applications for virtualization in multiple servers; b) the optimal strategy to control server speeds by dynamic voltage scaling; c) optimal rules for real time cluster sizing. The algorithms will be developed and integrated under a unified multi-time scale platform to exploit their benefits. They will be implemented in software using existing advanced computational infrastructure that is capable of interfacing with sensors and actuators in data centers. Software simulations will be performed using workloads experienced by standard enterprise servers to test, benchmark and validate the algorithms.
If successful, the results of this research will lead to improvements in the efficiency of data center energy usage and methods for handling multi-dimensional, multi-time scale, and nonlinear stochastic optimal control problems. The primary goal of this research is to combine stochastic optimization and stochastic optimal control for capacity planning and resource management in data centers. These methodologies will be used to integrate pro-active planning with real time control to reduce energy consumption to the minimum extent possible. Planning capacity and managing resources will help reduce the operating cost and the greenhouse gas emissions from data centers. The proposed work will also contribute to the analytical tools and software for stochastic optimization and optimal control problems.
