Learning Congestion State For mmWave Channels

Abstract

Millimeter wave (commonly known as mmWave) is enabling the next generation of last-hop communications for mobile devices. But these technologies cannot reach their full potential because existing congestion control schemes at the transport layer perform sub-optimally over mmWave links. In this paper, we show how existing congestion control schemes perform sub-optimally in such channels. Then, we propose that we can learn early congestion signals by using end-to-end measurements at the sender and receiver. We believe that these learned measurements can help build a better congestion control scheme. We show that we can learn Explicit Congestion Notification (ECN) per packet with an F1-score as high as 97%. We achieve this by leveraging unsupervised learning on data obtained from sending periodic bursts of probe packets over emulated 60 GHz links (based on real-world WiGig measurements), with random background traffic.