M&M: Machine-Learning Enhanced MPQUIC for Resilient Communication with mmWave Networks

Applications are continually seeking higher bandwidth and lower latency from wireless networks. Millimeter wave (mmWave) links, operating in the 20-60 GHz range, address these requirements. However, mmWave links have significant challenges due to environmental factors such as penetration losses, rain attenuation, and blockages. These factors lead to highly variable and intermittent connectivity for which traditional single-path transport protocols are ill-suited.Multipath transport protocols, including Multipath TCP (MPTCP) and Multipath QUIC (MPQUIC) leverage multiple wireless paths to enhance resilience and achieve higher aggregate throughput. These protocols can continue communicating over the remaining available path(s) even when one or more paths are obstructed. Nonetheless, efficiently detecting mmWave link obstructions and appropriately balancing data transmission across multiple paths remain critical challenges.This paper introduces M&M, a machine learning (ML)-based MPQUIC scheduler designed to dynamically adapt to mmWave link conditions. Leveraging a small mmWave testbed, we collect data across various environments to train an ML model that classifies environmental conditions and predicts available link-layer bandwidth. The scheduler utilizes these predictions to allocate data transmission proportions across different paths, complemented by traditional congestion control mechanisms. Evaluation results on the testbed demonstrate that our proposed scheduler effectively maintains high-bandwidth, low-latency wireless communications in dynamic and challenging environments.