Evaluation of the impact of different network telemetry collection approaches on machine learning models

• Grant number: 24/13591-1
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Start date: November 18, 2024
• End date: February 17, 2025
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Fabio Luciano Verdi
• Grantee: Gabriel Santos de Andrade
• Supervisor: Chrysa Papagianni
• Host Institution: Centro de Ciências em Gestão e Tecnologia (CCGT). Universidade Federal de São Carlos (UFSCAR). Campus de Sorocaba. Sorocaba , SP, Brazil
• Institution abroad: University of Amsterdam (UvA), Netherlands
• Associated to the scholarship: 24/03181-0 - Evaluation of the impact of different network telemetry collection approaches on machine learning models, BP.IC

Abstract

Traditionally, the monitoring and management of computer networks is based on the analysis of a set of performance metrics, such as: the CPU consumption of a given equipment, the number of packets sent through a network interface, or even the delay on a given connection. However, collecting these fine-grained metrics has not always been easy or even possible. This scenario has changed recently thanks to the programmable data plane with the P4 language in conjunction with INT (In-band Network Telemetry). In this sense, telemetry instructions direct the collection of new fine-grained metrics from network equipment at low granularity and high frequency, allowing network operators to obtain a clearer snapshot of the network state from a large amount of data. Such data is then used to feed ML (Machine Learning) models for different applications, such as QoS and congestion control predictions. However, while the amount of telemetry data collected is, on the one hand, extremely important for ML models, on the other hand, it can cause excessive consumption of network resources just to collect such metrics. In this sense, there are several techniques in the literature that use deterministic and probabilistic approaches to reduce the amount of telemetry data collected. The goal of this research project is to evaluate how different INT solutions affect the accuracy of ML models. We will evaluate deterministic and probabilistic INT mechanisms and the tradeoff between the quantity of data collected and accuracy. The most known ML models will be evaluated such as Random Forest, Decision Tree, KNN, and Reinforcement Learning.