Supporting TCP decisions in challenged networks via in-network telemetry

Supporting TCP decisions in challenged networks via in-network telemetry Nowadays, networks are facing several challenges in developing solutions to solve intrinsic problems like achieving acceptable levels of latency, bandwidth, faults, and loss of packets that guarantee responsive applications and services together with a better user experience. TCP is still the most adopted transport protocol in network communication for its reliability, but it performs poorly in heterogeneous or wireless networks because it has no mechanism to distinguish between congestion and stochastic packet losses. Thus, the incoherence of the congestion window (cwnd) value could lead to performance and throughput degradation. For this reason, there is a recent attempt to let the TCP learn the best cwnd updates via Machine Learning (ML)-based approaches. A specific class of ML, namely Reinforcement Learning (RL), has been proved to bring advantages to TCP and on top applications, having the ability to learn and adapt to the network environment. However, despite the improvements, none of these solutions can exploit network intelligence fully. This thesis aims to understand how it is possible to give support to TCP decisions by means of data collected on networks and processed by RL algorithms. Our solution is constituted by two components: a modified instance of TCP that runs on the end-hosts, and intelligent network devices to empower network telemetry. Regarding the latter component, our choice is to use P4 (Programming Protocol-independent Packet Processors) language for network programming given the vast gamma of metrics available and the fast processing time. Switches, programmed by means of P4 language, have complete control of the traffic through the network. Then, we consider two different in-band ways to carry these metrics to the end host, and we study the advantages and disadvantages in terms of complexity and performance. The first proposal is the integration of data on IP Options field, which consists of IP packet utilization without any format change or bytes addition, just Option field exploiting. The second one is the BPP (Big Packet Protocol) protocol integration which allows controlling packets flow over the network through information encoded in packets themselves. More specifically, it consists of BPP headers addition on the standard IP packet between the IP header and the IP payload field. Here we select the metadata field to store our metrics modified by switches of the emulated network for header integration and packet management reasons. The last step is to integrate the metrics in an existing RL-based TCP protocol, Owl, that helps select the correct value of the congestion window and can react to different network conditions. We finally experienced how this solution can improve the performance of TCP congestion control, impacting only limitedly on the network traffic overhead.