NeTS: Small: ML-Driven Online Traffic Analysis at Multi-Terabit Line Rates

Purdue University, West Lafayette IN

Investigators

Abstract

Internet network operations, whether run by humans or via automated systems, need real-time operational data and data analysis for timely decision making in support of a more secure and reliable Internet. This project supports drastically reducing response times in detecting and responding to network traffic anomalies, and aid with performance diagnosis and repair. Realizing network capabilities often requires machine learning (ML) inferencing algorithms (e.g., to detect anomalous traffic). Unfortunately, as network bandwidth grows to hundreds of gigabits to even terabits per second, it is challenging to analyze network traffic at line rates today. Consequently, network operators resort to out-of-band traffic analysis resulting in slow reaction times (e.g., to detect network intrusions). This project seeks to enable an Internet that can run ML-driven inference algorithms for applications such as network security at line rate of hundreds of Gbps or even Tbps by advancing and leveraging programmable switch technology. Doing so is challenging since programmable switches are constrained in their compute and memory capabilities, have limited expressivity, and limited support for runtime programmability (i.e., ability to make changes without a switch reboot). The project is developing (i) novel methods that can efficiently map an entire class of popular ML models such as decision trees and neural networks to a programmable switch pipeline in a manner that efficiently uses limited switch computation and memory resources, while allowing any model in the class to be supported at runtime; (ii) new switch primitives for computing ordered statistics (e.g., medians, percentiles) of flow features; and (iii) techniques to distribute large ML models on programmable switches across multiple pipelines, and multiple switches in a runtime programmable fashion while handling switch failures achieving resource efficiencies through a combination of new switch primitives, and new network-wide optimization models. The project team is collaborating with campus network operators for larger scale validations. The project will train Ph.D, Masters and undergraduate students, and lead to material on programmable switches in the networking curriculum. 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.