Elephant Flow Classification on the First Packet With Neural Networks

Abstract

Quick and accurate identification of the largest flows in the network would allow for the management of most traffic using dedicated, flow-specific routes and policies, thereby significantly reducing the overall number of entries in switch flow tables. Our analysis focuses on utilizing neural networks to classify elephant flows based on the first packet using 5-tuple packet header fields. The findings indicate that with simple neural networks comprising solely linear layers, it is possible to accurately detect elephant flows at their inception, thereby reducing the number of flow table entries -by up to a factor of 15 -while still effectively covering 80% of the network traffic with individual flow entries.

Utilizing TabNet Deep Learning for Elephant Flow Detection by Analyzing Information in First Packet Headers

“…The best reduction rate achieved for the 80% traffic coverage was 20.14. As shown in Figure 10 this is a 25% higher reduction rate than achieved previously with neural networks comprising solely linear layers, which provided only 15-fold reduction for the best parameter combination [36].…”