Traffic classification and sifting to improve TDM-EPON fronthaul upstream efficiency

Mobile fronthaul (MF) provides connectivity between a remote radio head (RRH) and base-band processing unit (BBU) in a cloud radio access network. Dedicated MF connections between RRH and BBU are costly due to high bandwidth requirements. The cost issue becomes more pressing as cell densification increases the number of connections. A time-division multiplexing Ethernet passive optical network (TDM-EPON) is a promising solution to reduce MF cost, as it can multiplex several fronthaul flows in the MF network. However, even though mobile users transmit intermittently, RRH samples radio signals continuously, resulting in large constant line-rate connections between RRH and BBU that require huge aggregated bandwidth and render statistical multiplexing gain marginal. To improve upstream transmission efficiency by utilizing statistical multiplexing in an MF network, we propose to enhance the conventional TDM-EPON architecture. The enhanced architecture employs two techniques: (1) traffic classification (classifying upstream traffic into useful and useless by using machine-learning classifiers); and (2) useless-data sifting (sifting useless traffic to avoid the transmission of unnecessary EPON frames). We explore two feature-selection methods together with classifier biasedness in the first technique to achieve the best classification results, which then serve as input to our proposed sifting-based hybrid bandwidth allocation scheme in the second technique. Simulation results show significant improvements in terms of per-RRH traffic load, number of RRHs supported by same EPON, and signal-to-noise ratio, while keeping the end-to-end delay under 100 μs.