Characterizing Fairness for 3G Wireless Networks Vaneet Aggarwal, Rittwik Jana, Jeffrey Pang, K. K. Ramakrishnan, N. K. Shankaranarayanan AT&T Labs Research, Florham Park, NJ, USA Abstract—The end to end system data performance over a 3G cellular network depends on many factors such as the number of users, interference, multipath propagation, radio resource management techniques as well as the interaction between these mechanisms and the transport protocol’s flow and congestion mechanisms. Using controlled experiments in a public cell site, we investigate the interaction between TCP and the 3G UMTS/HSPA network’s resource allocation, and its effect on fairness in the throughput achieved across multiple (up to 26) TCP flows in a loaded cell sector. Our field measurement results indicate that TCP fairness fluctuates significantly when the air interface (radio link) is the bottleneck. We also observe that TCP fairness is substantially better when the backhaul link (a fixed wired link) is the bottleneck, instead of the air interface. We speculate that the fairness of TCP flows is adversely impacted by the mismatch between the resource allocation mechanisms of TCP’s flow and congestion control and that of the Radio Access Network (RAN). I. I NTRODUCTION There has been tremendous growth in the use of data over wireless cellular networks with the advent of smartphones over the last few years. As in wired networks, the dominant transport protocol used in cellular networks is TCP, which comprises over 95% of flows [1]. Even real-time stream- ing traffic, such as video, is more often transported over HTTP/TCP than UDP. Hence, there is significant interest in understanding how the allocation and scheduling of cellular resources impacts the behavior of TCP flows. A major part of this interest is on the downlink in UMTS/HSPA cellular networks, i.e. from the base station to the user equipment (UE), such as a smartphone or a cellular data modem. Radio access network (RAN) resources are allocated to each UE in a highly complex and dynamic manner, taking into account the radio resources available, the signal strength observed by the end-device, outstanding data to other receivers, and other considerations. A primary notion is that there is a “channel” for each UE to which RAN resources are allocated (this is analogous to a virtual circuit-switched net- work’s connection). Hence, recent work (e.g., [2], [3], [4]) has examined the interaction between TCP’s flow and congestion control mechanisms, the RAN resource allocation methods, and the wireless channel. These studies observed negative performance impact on individual TCP flows – for example, poor efficiency due to link-layer retransmissions that result in variability in RTT. It is important to recall, however, that TCP has two major goals: efficiency and fairness. Previous work focused primarily on how RAN resource allocation impacts the efficiency of flows. Our contributions are observations on the fairness in the performance obtained for competing TCP flows based on an extensive set of measurements on an operational UMTS/HSPA cellular network. We characterize both temporal (long term time average) and spatial (distribution of throughout within a cell or sector) fairness. Understanding how RAN resource allocation interacts with TCP is not straightforward due to its complexity and because it functions independently at the radio link control (RLC) layer, a protocol layer below IP. For instance, High Speed Downlink Packet Access (HSDPA) has fast scheduling and hybrid ARQ. The RAN resource allocation seeks to provide equitable proportioning of radio resources depending on the channel condition, user location, interference, and scheduling discipline. For each transmission time interval (TTI, typically 2 ms), the scheduler located in the Node B (base station), decides on the users to be scheduled and their corresponding data rates. This is based on the reported channel quality by the UEs, as well as fairness metrics. There is also a radio link control (RLC) flow control mechanism between the RNC and the node B. For the same traffic, we also have TCP’s flow and congestion control, which operate on timescales of one or more end-to-end RTTs (typically 100-300 ms in cellular networks), while the HSDPA scheduler algorithms operate on finer timescales of the order of 2ms TTIs. TCP’s congestion control mechanism also seeks to achieve proportional fairness across connections [5], [6]. To evaluate how these two layers interact, we measure how TCP flows perform when the RAN is loaded with a controlled number of high-demand, long-lived flows. We vary the number of sources, although we expect an individual source to be capable of offering a load that is a high fraction of the air interface capacity of a single sector. We conducted measurements in two types of environments: (a) where we were confident that the air interface was the bottleneck and (b) where we were confident that the air interface was not the bottleneck. Our contributions are as follows: First, we use real-life controlled experiments with an operational 3G UMTS/HSPA cellular network with multiple datacards and smartphones (total up to 26) to characterize the fairness across TCP flows as a function of increasing load and varying traffic dynamics across 1, 2, and 3 cell site sectors. Second, we investigate how fairness is impacted if the radio air interface is the bottleneck, and compare this with the case when the backhaul is the bottleneck. Our findings indicate that there is greater variability in fairness when the air-interface is congested as opposed to the backhaul being congested. We believe this is because TCP’s congestion control takes precedence over the RAN resource allocation algorithms when the backhaul is the bottleneck, thus leading to better fairness.