Congestion Control for QUIC protocol using RNN And LSTM as deep learning models (COCO-QUIC)

Abstract

QUIC (Quick UDP Internet Connections) is a new transport protocol that was designed on the motivation of improved speed, better performance and enhanced security. This protocol has been adopted in several use IoT cases and applications. Several research efforts have demonstrated performance improvements by switching to QUIC however, new challenges continue to come up. As Internet networks grow due to massive demand attributed to the Internet of Things (IoT), network congestion continues to cause huge challenges along the transport Layer. Tremendous efforts, including deep learning methods, have been used to improve network congestion over time. Although deep learning methods have been used to address network congestion challenges in many works, less research has been directed towards the QUIC protocol, therefore, this study proposes Congestion Control for QUIC using Recurrent Neural Network (RNN) and Long-Short Term Memory (LSTM) as Deep Learning models (COCO-QUIC). The datasets for this study were collected usingWireshark packet sniffer from Makerere University Network Operations Center (NOC) and contains 1,549,088 samples of records with parameters including e-mail, video streams, audio streams, social media, browsing, and anomaly-related traffic. Our main objective is to develop a congestion control model and two supervised deep learning models have been stacked together and used in this study which are the Simple Recurrent Neural Networks (Simple RNN) and Long-Short Term Memory (LSTM), formulating COCO QUIC model. The Simple RNN model has been used as the initial model with the dataset for training, testing, and validation and achieved an overall accuracy of 92.57% on the collected data samples, the LSTM model attained an overall accuracy of 92.90% on the same collected data samples and COCO QUIC attained an overall accuracy of 93.33%. The two stacked models of LSTM and RNN were compared with RNN,LSTM and RF as existing models for performance evaluation and were found to perform much better with less Latency, Delay, Packet loss and improved bandwidth. The new COCO QUIC model performed better than Simple RNN, LSTM, and RF, indicating that the designed model show improvement in performance on congestion compared to existing models. This research study’s contribution is the performance of COCO QUIC model on congestion control measures of different learning rates applied to the Simple RNN, LSTM and R.F models on parameters such as throughput, delay, packet loss, latency and it was also observed that COCO QUIC performed much better than Simple RNN, LSTM and RF.