Workshop on Self-Driving Networks

点击翻译按钮获取中文摘要

This workshop brings together leading researchers from a range of disciplines across computer science to define a new research agenda in network measurement and data analytics with the goal of exploring how to design networks that manage themselves. These experts will explore taking advantage of advances in disciplines including machine learning, distributed systems, and formal methods to address growing requirements and constraints of modern networking applications. Because of the proliferation of applications and services that now run over the Internet ranging from video streaming to Internet-connected smart home devices to augmented reality---the expectations for the performance, reliability, and security of our communications networks are greater than ever, as the number and diversity of applications that run on top of the network continue to proliferate, and as the volume of traffic on the network continues to grow. To meet these expectations, network operators work tirelessly to continuously collect troves of heterogeneous data from the network, analyze this data to infer characteristics about the network, and decide whether to change the network's configuration in response to network conditions (e.g., a shift in traffic demand or a cyber attack). Today, these three steps are decoupled: operators perform them separately, on different timescales, often in a slow or manual fashion that relies on intuition, as opposed to data, analysis, and inference. The vision for this workshop is that networks might one day be able to largely manage themselves through a combination of query-driven network measurement, automated inference techniques, and programmatic control. Intellectual Merit: The research agenda lends itself to research problems that will foster advances in computer science, including the following areas: 1. Distributed systems that optimize the use of limited resources for complex tasks, including support for multiple simultaneous queries; New architectures to support programmable measurement in hardware; Algorithms that partition a network analytics query across a centralized stream processor and the distributed switches and network middleboxes. 2. New measurement techniques (beyond "ping" and "traceroute") that leverage the capabilities of P4-capable data planes (e.g., in-band telemetry); Software/hardware co-design for better network measurements; Clean-slate, problem-driven designs for new network measurement tools that might tackle problems in network measurement that have proved evasive (e.g., application quality of experience); Measurement of unified compute, storage, and networking infrastructure, including monitoring of container-based systems 3. Machine Learning and new algorithms for automated troubleshooting and "what-if" scenario evaluation; Development of parsimonious models that could be implemented (at least partially) at line rate on switch hardware; Prediction and inference over non-stationary datasets to changing traffic patterns. 4. Security and privacy through scalable algorithms and systems for detecting a broad range of attacks, from denial of service to data exfiltration; Better ways to monitor application performance without having to perform man-in-the-middle attacks on traffic. Broader Impacts: Results from this workshop will be broadly distributed so that researchers in all of the areas noted above will benefit from the discussions, conclusions and recommendations resulting from the workshop. Research inspired by the workshop could have broad societal impacts by helping network operators envision how to integrate measurement, data analysis, and configuration decisions and move toward automated network control.