Measurement Analysis of TCP Congestion Control Algorithms in LTE Uplink Ali Parichehreh, Stefan Alfredsson, Anna Brunstrom Department of Mathematics and Computer Science Karlstad University, Sweden Email:{firstname.lastname}@kau.se Abstract—The unprecedented growth of user generated con- tents yielded by the proliferation of social networks applications, cellular based video surveillance and device-to-device (D2D) communication, makes the cellular uplink communication an attractive topic. In this paper we conduct a systematic evaluation and measurement analysis to characterize cellular uplink traffic and compare its interplay with different TCP congestion control algorithms (CCA), namely NewReno, Cubic, and BBR, in both stationary and mobility scenarios. The evaluation encompasses average throughput, average round trip time (RTT), fairness among simultaneous flows, and packet retransmission. The in- tended behavior of BBR has been observed in LTE uplink, but some severe issues such as lack of fairness among simultaneous flows and massive on device packet losses have been observed. It is observed that the lack of fairness among simultaneous flows can unpredictably change the throughput of multi-flow applications. Index Terms—Measurement Analysis, Congestion Control Al- gorithms, Long Term Evolution, Uplink Communications. I. INTRODUCTION Traditionally, mobile devices support only a single applica- tion due to the low computational power. The proliferation of more powerful mobile devices enables running multiple applications by leveraging concurrent TCP streams. Further- more, user generated contents by social network applications as well as the new use cases defined by 5G networks such as machine to machine (M2M) and device to device (D2D) communication, fog based traffic offloading, and HD video chat/surveillance, cause an inevitable uplink traffic surge in future mobile networks. The Third Generation Partnership Project (3GPP) has recently standardized the advanced version of Long Term Evolution (LTE-A) and LTE-A Pro, as a road toward 5G networks, with a peak data rate of up to 1.5 Gbps in uplink [1]. Therefore studying the performance of multiple simultaneous flows with different CCAs in a real LTE-A network that leverages 20MHz bandwidth in uplink (extendable to 100MHz bandwidth when employing carrier aggregation) is crucial, and deserves further investigations. In the past decade, a large body of work has been produced by researchers to improve the performance of TCP congestion control algorithms (CCA), ranging from loss based CCAs such as [2], [3], [4], [5] to delay based [6], [7], and mixed loss- delay based protocols [8]. In addition, some CCAs such as [9], [10], [11] leverage machine learning algorithms to find the optimal transmission points. BBR (Bottleneck Bandwidth and Round-trip propagation time) [12] is one of the recently introduced model-based congestion control algorithms. It has been shown to achieve a superior performance in high through- put and wired communications such as high-speed wide- area networks [12], compared to the widely used congestion control algorithms. The long term plan of the BBR team is to enable it as a default congestion control algorithm for the Internet [13]. However, it is still questionable whether BBR is able to find an equilibrium point between latency and throughput in unpredictable mobile cellular networks [14] that is an indispensable part of the future Internet. Therefore systematic measurement analysis and experiments are required to better understand the performance of BBR CCA in different scenarios and traffic patterns. Recent studies on the performance of different TCP pro- tocols (including BBR) are conducted in both stationary and mobility scenarios over live LTE networks as well as emulators with LTE link traces, considering different metrics such as throughput, delay, and fairness. The performance of BBR in highway scenarios is investigated in [15] and results indicate that BBR achieves higher throughput compared to Cubic even in low Reference Signal Received Power (RSRP) regime or handover regions. However, in [14] it is observed that BBR is not able to estimate the available bandwidth and utilize LTE link in an emulation environment. Hence a link coupled TCP CCA is proposed that leverages the architectural trends of 5G networks to enable accurate satisfaction of the requirements of each individual application. In other works, BBR shows lack of fairness among simultaneous flows [16] and it can be especially violated in the startup phase where competing short flows with loss-based CCA struggle to get a fair share of the bandwidth [17], [18]. Although the performance of TCP congestion controls in mobile cellular networks have been investigated in downlink direction, there are only few works considering the perfor- mance of TCP CCAs in LTE uplink communications. On- device bufferbloat is an important delay factor investigated in [19], and there is no practical widely deployed solution to mitigate its effect. Further investigations revealed that qdisc based solutions are not effective enough in confrontation with on-device bufferbloating, as qdisc strategies have negligible impacts on the firmware queue [19]. In addition, using the scheduling request based access, users must transmit a grant