Unifying Buffer Sizing Results Through Fairness

Buffer sizing in Internet routers is a fundamental problem that has major consequences in design, implementation, and the economy of the routers, as well as on the performance observed by the end users. Recently, there have been some seemingly contradictory results on buffer sizing. On one hand, Appenzeller et al. show that as a direct consequence of desynchronization of flows in the core of the Internet, buffer sizes in core routers can be significantly reduced without any major degradation in network performance. On the other hand, Raina and Wischik show that such reduction in buffer sizing comes at the cost of synchronization and thus instability in the network. This work unifies these results. We show that the main difference arises from the implicit assumption of fairness in packet dropping in the latter result. We demonstrate that desynchronization among flows observed by Appenzeller et al. is caused by unfair packet dropping when a combination of TCP-Reno and the drop-tail queue management is used. We also show that bringing fairness in packet dropping will introduce synchronization among flows, and will make the system unstable as predicted by Raina and Wischik. Our analysis suggests that there is an intrinsic trade-off between fairness in packet drops and desynchronization among TCP-Reno flows when routers use the drop-tail queue management. Achieving fairness, desynchronization, small buffer size, and 100% link utilization at the same time is desirable yet challenging. As an effort towards this goal, we propose a queue management scheme at routers called “rate-balancing”, and show evidence that it is possible to achieve desynchronization among flows and reduce buffer sizes without sacrificing fairness while maintaining full link utilization. I. M OTIVATION AND INTRODUCTION There have been increasing amount of interests on buffer sizing due to the important role that buffer plays in routers and the performance of the Internet. The goal of buffer sizing is to find out how small we can make Internet router buffers without any degradation in network performance. A plethora of recent work emerged to reduce buffer sizes [1]–[5] and to understand the relationships between buffer sizing and other parameters of the network [8]–[14], such as, throughput, delay, loss, stability [6], [7], and the impacts of various traffic conditions [8], [15]. Recently, there have been some seemingly contradictory results on buffer sizing in Internet core routers. Appenzeller et al. show that buffer sizes in core routers can be reduced significantly, without any major degradation in network performance [1], whereas, Raina and Wischik show that such reduction in buffer sizing can cause instabilities in the network [7]. Instability, here, is defined as the periodic variations in the aggregate congestion window size of the flows. Which result is correct? To answer this question, we studied the dynamics of the system in terms of buffer sizing through mean-field theory [16] analysis and ns2 [17] simulations. We demonstrated that there is an intrinsic trade-off between fairness among TCP-Reno flows and desynchronization among them: fairness in packet drops can create synchronization among flows and have an adverse impact on the performance of the network; conversely, unfair packet drops can lead to reduced synchronization, and thus higher throughput in the network. The implicit underlying assumption here is that intermediate routers in the network use the drop-tail queue management scheme. Fairness has always been considered to be a desirable property in the network. The network is expected to treat individual flows in a fair manner when resources are limited. Synchronization among TCP flows, on the other hand, has always been considered to be an undesirable effect. Synchronized flows need much larger buffer sizes in core routers and cause local/global instabilities in the system, as well as degradation in the performance observed by individual flows. To see the origins of the trade-off between fairness and desynchronization, let us consider a congested link in a network carrying a large number of flows. An Active Queue Management (AQM) scheme that fairly drops packets at times of congestion will impact a large percentage of flows since it will distribute packets dropped among the flows. On the other hand, an unfair AQM scheme can drop a lot of packets from a few flows, thus reducing the number of flows which see one or more packet drops. TCP-Reno flows react dramatically to packet drops by halving their congestion window sizes for each dropped packet [18]. Therefore, when a large number of flows see packet drops around the same time they will synchronously r act by reducing their congestion window sizes. This may lead to a significant reduction in instantaneous throughput of the system. In an unfair AQM scheme, however, only a few flows will take the hit, and thus only a small fraction of flows will react to packet drops. Therefore, the aggregate congestion window will change less significantly. In Section II-A we show that TCP-Reno combined with the drop-tail queue management scheme is not a fair system with regard to flow packet drops. TCP-Reno sources inject packets to the network in a bursty manner [2]. Whenever such a burst arrives at a queue which is full, or nearly full, it will encounter a large number of packet drops unfairly on a few flows. Our argument that there is a trade-off between unfairness and synchronization, along with the fact that TCPReno and drop-tail are unfair can explain why there is very little evidence of any synchronization in today’s core routers. Our results unify the seemingly contradictory conclusions mentioned above. Appenzellert al. have shown that due to t e desynchronization of flows in the core of the Internet, one can reduce buffer sizes in core routers by a factor of √ N from