CNS Core: Small: Realistic Traffic Generation through Application-Agnostic Learning

点击翻译按钮获取中文摘要

Industry evaluation of new computer network applications and infrastructure - as well as research evaluation of proposed platforms and protocols - rely on high quality representations of realistic network usage. Historically these needs have largely been met through the use of workload generators that rely on generalized traffic models. However, these models have increasingly proven inadequate in modern environments with fast-evolving applications, expanding mobility, and the rising prevalence of Internet-of-Things (IoT) devices in end-user networks. As a result, novel methods, protocols, and hardware prove difficult to verify in real-world scenarios prior to production deployment. This project distills complex applications into realistic models that can be used to evaluate future systems and deployments.The tools developed through this project employ a combination of a novel expert system that separates application-specific behavior from infrastructure-specific behavior and a machine learning (ML) pipeline that captures complex application exchanges in order to provide realistic models of application traffic patterns for use in existing generators. Label selection, classifier design, and choices of existing ML algorithms will drive documentation on both future research direction as well as the environments in which current tools and models are best employed.The generation of abstract but high-fidelity models of application traffic patterns as a result of this project provides valuable data to industry users and researchers alike. The separation of application behavior from potentially sensitive and proprietary input data allows for significant expansion of the quality of traffic models available for planning and research tasks while also providing portability to environments exploring new protocols and infrastructure design. The availability of these models allows for significantly more effective validation and reproducibility of prospective studies. Furthermore, the production of these models will integrate with educational outreach efforts to high school, undergraduate, and graduate level courses on computer networking and cybersecurity where representative topologies and application traffic drive hand-on labs.Documentation for project tools and code, as well as backing project data, will be located at http://docs.uh-netlab.org/appmodel/index.html, and it will be publicly available for at least 5 years after the end of substantive project work. Source code (along with documentation source) will be made available at bitbucket (http://www.bitbucket.org/uh-netlab/) on an ongoing basis.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.