MODELS: performance MODELing of Software-defined data center networks

点击翻译按钮获取中文摘要

Performance modeling and prediction techniques provide a powerful tool to analyze the performance of IT systems and applications running in a modern data center. In this project, we focus on the network infrastructures of modern virtualized data centers. Network virtualization is the main trend among the various approaches to address current challenges in modern network infrastructures. However, no common approach exists for modeling and evaluating the performance of virtualized networks. Existing performance models are too coarse-detailed, too focused, or too limited in scope, to be applied directly. We propose a comprehensive modeling approach based on a generic modeling language, i.e., a meta-model, designed to capture the performance-relevant aspects of data center networks. The meta-model will serve as a basis to automatically transform its instances into predictive models by using model-to-model transformations. The presented proposal aims to bridge the following gaps: First, we envision a performance modeling approach that offers flexibility in selecting the modeling granularity (coarse-grained analytical models versus highly detailed simulation models). Second, we intend to bridge the gap between the performance of the software architecture, the computing infrastructure, and the network infrastructure. Finally, the approach will support modeling of networks based on Software-Defined Networking (SDN) as a representative network virtualization technology. The proposed project will enable flexible performance prediction of SDN-based networks providing results at a user-selected targeted level of detail and accuracy. Through integration with existing modeling techniques focused on the computing infrastructure, the proposed modeling approach will support modeling of an end-to-end data center infrastructure (including computing, networking, and software). Moreover, we will provide the first language to describe SDN-based network setups which, due to the dynamic changes in the structure of virtual networks, are more difficult to analyze in an ad-hoc manner. Finally, we will support automatic extraction of network models based on monitoring data, so that the approach can be easily applied by non-experts.