Model-based Analysis of Next-Generation Networks

点击翻译按钮获取中文摘要

It is foreseen, that new paradigms in computer networking like Network Function Virtualization (NFV), Network Virtualization (NV) and Software Defined Networking (SDN), will increase the flexibility and openness of modern communication infrastructure. The new approaches will enable virtualization of network functions (NFV), network slicing (NV) and a (logically) centralized control of the network (SDN), such that functionality can be implemented in a physically decoupled way, e.g. running in a data center. SDN enables novel approaches to analyze and control the traffic in the network, however, requiring a deep understanding of the underlying hardware. While the packet processing pipeline of well-designed, specialized hardware is specified in detail, virtualized software on commodity hardware is more difficult to tackle. On the other hand, the additional degrees of freedom enabled by virtualizing network functions (NFV), such as custom packet processing pipelines and dynamic placement, create new possibilities for optimization.This project aims to evaluate, combine, and enhance performance models of networks and their components concerned with packet processing. The performance assessment of novel networks requires suitable modeling tools, e.g. to represent interaction on the control plane or take software-based packet processing into account. New models are also required to account for changes in the networking hardware that describe the effects of limiting factors like CPU interconnects or the bandwidth of memory, PCIe, or Ethernet. We aim to combine models, that describe individual effects, into larger entities that allow the modeling of complex packet processing tasks.ModANet covers various methods for performance analysis, e.g. resource-based models, models based on Network Calculus, and simulations. We plan to build a framework for model evaluation, which allows for automated determination of the quality and scope of models, thus enables covering a large parameter space. We plan to automate the calibration and evaluation of the analyzed models using machine learning techniques. We also aim to evaluate the modeling approaches with regard to their complexity to obtain feasible models.The outcomes of the project should provide a deep understanding of the characteristics of the different processing pipelines in SDN-based networks. Considering separate packet processing steps in network nodes allows to make reliable statements about the performance of composed packet processing chains. Furthermore, the enhanced configuration possibilities lead to new optimizations.