Entwicklung einer auslastungsabhängigen Energieeffzienzmetrik für Rechenzentren

Data centers' total operational expenditures (OpEx) are rising due to the need for data center extensions caused by the increasing data center service usage as well as continuously growing data volumes. In order to save energy costs, which represent a signi cant share of the OpEx, the data center operator's focus is shifting towards energy e ciency improvements. However, before optimizations can be implemented, one has to gather information about the actual state of the data center which can be done with the aid of energy e ciency metrics. Within the scope of this thesis, the Load Dependent Energy E ciency (LDEE) an energy e ciency metric for data centers which is applicable on a broad spectrum of use cases has been developed. Among providing a general overview of energy e ciency (1) further use cases are fair comparisons of data centers' energy e ciencies (2), the identi cation of sources causing ine ciencies (3), and estimations of achievable savings by potential optimization measures (4). While common metrics may only provide a general overview on energy e ciency, the LDEE may also be used for supporting the optimization processes. The LDEE makes use of load dependent performance and power models for single data center components as well as an IT load model abstracting the distribution of system utilization on all IT systems. These models are then combined to represent the total data center performance and power in relation to estimated usage. The component models base on publicly available data sources such as benchmark results in databases or operating parameters in data sheets. Thus, there is no need in actual measurements which may disrupt a data center's operation or at least constitute a high e ort. In the case of missing models or data for infrastructure components, the Infrastructure Power Adaptability (IPA) metric, which also has been developed in the context of this work, may be applied. The IPA uses traces from the Power Usage E ectiveness (PUE) to derive the actual dependency of infrastructure power to IT power. By combining PUE with IPA the load dependent infrastructure power can be estimated. The component models and the LDEE as a whole were analyzed in several evaluations. The IT load model and the power models were compared directly to available measured data and have relative errors of less than 5%, except for the free cooling model. In contrast, the performance models as well as the whole metric were applied on synthetic data center data and compared with alternate approaches. In a discussion, the results are examined regarding the requirements for energy e ciency metrics, which have been derived from the aforementioned use cases. Further evaluations demonstrate the broader application possibilities of the LDEE compared to the alternative approaches.