Luca Giacomoni

Luca Giacomoni, Basil Benny, George Parisis

Reinforcement Learning (RL) has gained significant momentum in the development of network protocols. However, RL-based protocols are still in their infancy, and substantial research is required to build deployable solutions. Developing a protocol based on RL is a complex and challenging process that involves several model design decisions and requires significant training and evaluation in real and simulated network topologies. Network simulators offer an efficient training environment for RL-based protocols, because they are deterministic and can run in parallel. In this paper, we introduce \textit{RayNet}, a scalable and adaptable simulation platform for the development of RL-based network protocols. RayNet integrates OMNeT++, a fully programmable network simulator, with Ray/RLlib, a scalable training platform for distributed RL. RayNet facilitates the methodical development of RL-based network protocols so that researchers can focus on the problem at hand and not on implementation details of the learning aspect of their research. We developed a simple RL-based congestion control approach as a proof of concept showcasing that RayNet can be a valuable platform for RL-based research in computer networks, enabling scalable training and evaluation. We compared RayNet with \textit{ns3-gym}, a platform with similar objectives to RayNet, and showed that RayNet performs better in terms of how fast agents can collect experience in RL environments.

Mihai Mazilu, Luca Giacomoni, George Parisis

Learning-based congestion control (CC), including Reinforcement-Learning, promises efficient CC in a fast-changing networking landscape, where evolving communication technologies, applications and traffic workloads pose severe challenges to human-derived, static CC algorithms. Learning-based CC is in its early days and substantial research is required to understand existing limitations, identify research challenges and, eventually, yield deployable solutions for real-world networks. In this paper, we extend our prior work and present a reproducible and systematic study of learning-based CC with the aim to highlight strengths and uncover fundamental limitations of the state-of-the-art. We directly contrast said approaches with widely deployed, human-derived CC algorithms, namely TCP Cubic and BBR (version 3). We identify challenges in evaluating learning-based CC, establish a methodology for studying said approaches and perform large-scale experimentation with learning-based CC approaches that are publicly available. We show that embedding fairness directly into reward functions is effective; however, the fairness properties do not generalise into unseen conditions. We then show that RL learning-based approaches existing approaches can acquire all available bandwidth while largely maintaining low latency. Finally, we highlight that existing the latest learning-based CC approaches under-perform when the available bandwidth and end-to-end latency dynamically change while remaining resistant to non-congestive loss. As with our initial study, our experimentation codebase and datasets are publicly available with the aim to galvanise the research community towards transparency and reproducibility, which have been recognised as crucial for researching and evaluating machine-generated policies.