Analysis of dependability and sustainability requirements to support the deployment of dense data center architectures

Abstract

The convergence of communication networks and the demand for storing and processing capacities of large amounts of information, especially in recent years, has driven requests for everything-as-a-service (XaaS) and has been generating, on an increasing scale, demands for new data centers (DC) infrastructures. Energy consumption per rack unit has increased due to the consolidation of spaces via server virtualization and as an effect of the density achieved by modern IT equipment. As these factors contribute to significant improvements, such as gains in availability and performance, they also contribute to global warming due to carbon dioxide (CO2) emission, one of the leading greenhouse gases. This occurs both in the production of energy, which feeds these large infrastructures, and in the consumption of the entire IT load. Given this context, we analyze dependability and sustainability requirements to support data centers’ implementation in this work. We initially performed a systematic literature review to identify possible problems in the area. We propose a set of models for power architectures and data center networks to quantify specific dependability and sustainability metrics. The proposed approach is based on the hybrid and hierarchical modeling for evaluating costs, exergy, electrical demand, efficiency, availability, reliability, etc., of the generated models. We propose energy flow models (EFM) to determine the minimum energy required for a data center project, considering the computer room (DC network). We created some scenarios to calculate the impact of environmental sustainability from raw material used in energy production. We also propose reliability block diagram (RBD) models to represent the power and network architectures of the data centers to identify the reliability importance of the components for the systems’ availability. Besides, we use parametric sensitivity analysis techniques to carry out experiments in different scenarios to identify possible bottlenecks for architectural failure. Based on the formalism of the stochastic Petri nets (SPN), we propose a maintenance policy to quantify the impact caused on the availability of the networks in the face of the variation in the mean time to failure (MTTF) and repair (MTTR) of the components. Additionally, we check if the availability achieved reaches service levels specified for data centers given their tier classifications. The set of modeling techniques used serves as a subsidy for the experiments to represent real scenarios. Thus, we can quantify essential design requirements, identifying possible vulnerabilities to the system’s failure before the implementation of a real architecture has been done.