pChanger
Published:
Heavy changer is flows change dramatically between two consecutive periods exceeds a specific threshold and can be viewed as a signal of congestion or malicious attacks. Identifying heavy flows (e.g., heavy hitters and heavy changers) in network traffic helps to improve network performance. It has many benefits to perform prediction of heavy changers within the data plane including eliminating extra communication between the control plane and the data plane and detecting the very beginning packages arrived at the switches. However, it is challenging due to extremely limited resource availability and the stringent requirements of fast packet processing. There are several proposed methods and systems in the research community for conducting network measurements on high-speed networks, yet there currently exists no advanced method to predict heavy changes in the data plane of programmable switches. In this thesis, we introduced pChanger, a programmable heavy changer prediction method based on time slots. This approach enables data planes on programmable switches to predict heavy changers by balancing performance and resource utilization, employing supervised machine learning (ML) techniques. In our evaluation, we first tested pChanger with traces and deployed it on bmv2 software P4 switches. The evaluation results indicate that pChanger achieved an accuracy of 99% and a recall of 75% on the data plane. We also demonstrate its near-linear runtime for predictions.
Key Words: Machine learning (ML), P4, heavy changer, programmable data plane, time slot