Deep Recurrent Neural Network for Bandwidth Prediction in Software Defined Data Center Networks

Abstract

With the advent of 5G SDN (Software-defined networking), there has been a huge growth in demand for uninterrupted flow of data across SDN controlled devices. This is due to high bandwidth consuming applications like online games, live telecast/broadcast, video conferencing, etc., which requires continuous available bandwidth. Ensuring bandwidth availability and seamless service delivery of these applications pose great challenges. Network monitoring and load prediction would play an important role in allocating available bandwidth efficiently. However, existing art lacks prediction capability and has a significant overhead in terms of time when used in SDNs. In this paper, we propose a deep learning model using Long Short Term Memory (LSTM) to inculcate prediction capability with minimal overhead. With this built-in intelligence in routing, network and computation resources can be allocated efficiently. Our solution can predict average Tx and Rx load across a link in network based on telemetry data from data center switches like cisco nexus 9000, juniper, arista, etc. The telemetry scheme we used is push based streaming telemetry with almost negligible overhead. The proposed model predicts utilized bandwidth with about 90% accuracy. This prediction capability in networks can be exploited by service provider who in turn provides applications with higher Quality of Experience (QoE) to the end-users.