UC San Diego

Data centers need efficient energy management mechanisms to reduce their consumption, energy costs and the resulting negative grid and environmental effects. Many of the state of the art mechanisms come with performance overhead, which may lead to service level agreement violations and reduce the quality of service. This thesis proposes novel methods that meet quality of service targets while decreasing energy costs and peak power of data centers. We leverage short term prediction of green energy as a part of our novel data center job scheduler to significantly increase the green energy efficiency and job throughput. We extend this analysis to distributed data centers connected with a backbone network. As a part of this work, we devise a green energy aware routing algorithm for the network, thus reducing its carbon footprint. Consumption during peak periods is an important issue for data centers due to its high cost. Peak shaving allows data centers to increase their computational capacity without exceeding a given power budget. We leverage battery-based solutions because they incur no performance overhead. We first show that when using an idealized battery model, peak shaving benefits can be overestimated by 3.35x. We then present a distributed control mechanism for a more realistic battery system that achieves 10x lower communication overhead than the centralized solution. We also demonstrate a new battery placement architecture that outperforms existing designs with better peak shaving and battery lifetime, and doubles the savings. Data centers are also good candidates for providing ancillary services in the power markets due to their large power consumption and flexibility. This thesis develops a framework that explores the feasibility of data center participation in these markets, focusing specifically on regulation services. We use a battery- based design to not only help by providing ancillary services, but to also limit peak power costs without any workload performance degradation