Learning how to QUIC-kly adapt to Wireless Network Conditions

QUIC (Quick UDP Internet Connections) is a transport layer protocol, proposed by Google and used both in Chrome and YouTube. Its aim is to improve overall performance by solving some problems of the standard transport protocol currently used by most websites, i.e., TCP. Amongst the other, those most notable are Head of Line blocking and long handshake times. Researchers have put their efforts in improving many aspects of transport layer protocols, i.e., power and bandwidth constraints and congestion control in wireless networks. Specifically, research efforts have tried to improve the performances of congestion control algorithm, either with fixed heuristic or with new machine learning algorithms. However, none of these solutions is able to exploit and combine lower-level metrics with transport layer information, thus providing a more comprehensive view of the state of the network. In this work, we enhance the congestion control mechanism of the QUIC protocol via usage of a Deep Reinforcement Learning model. We design a neural network-based model that aims at obtaining better inference performance by means of cross-layer information. Finally, the model was trained by exploiting a full-stack dataset, which was collected during this thesis. The proposed approach has been validated in a NextG wireless network architecture. Experimental results shows interesting performance trade-offs between RTT and bandwidth in a variety of conditions, from video streaming to random web browsing, proving the effectiveness of the discussed approach. The code and the dataset are available with an open-source license.