Optimizing and benchmarking MAMBA for network traffic prediction on edge devices

Abstract

In this digital environment, with the growing number of connected devices, effective network traffic prediction turns out to better network performance, lightening network load, and continuous transmission of data. Standard network monitoring solutions work mostly on cloud processing, whereas this brings about various issues-rescuing-delay, energy consumption, and security concerns. Hence, the edge computing proposes a new technique whereby data is processed at real-time right on the device near the data source, i.e., at the network edge.

However, challenges arise with ML model deployment on edge devices, especially concerning computation efficiency, memory, and energy consumption.

This work explores the efficiency of using MAMBA, the latest approach to sequence modeling, for network traffic prediction in edge devices. MAMBA performance will be benchmarked against existing models in terms of prediction accuracy, inference speed, and computational efficiency. To configure for real-time edge deployment, quantization and pruning techniques will be applied to the best among the models, reducing model size while maintaining predictive performance. Upon compressing the model, it shall be deployed and tested on Jetson Orin Nano (8GB) to evaluate its performance in handling real-time network monitoring under limited hardware resources.

The findings contribute to building lightweight, scalable, and energy-efficient automated network-monitoring systems. By optimizing these learning models for edge deployment, the work considers improving real-time traffic prediction, lessening computational overhead, and ultimate consideration for embedding AI into modern network infrastructure.