1 Assessing the Impact of Multiple Active Queue Management Routers Michele C. Weigle, Deepak Vembar, and Zhidian Du {mweigle, dvembar, zdu}@cs.clemson.edu Department of Computer Science, Clemson University Abstract— Recent studies have shown that a non-negligible number of packets face multiple congested links on Internet paths. We investigate the impact of using multiple Active Queue Management (AQM) routers on paths that consist of multiple congested links. We present the results of an ns-2 evaluation study of various AQM techniques that, in contrast to previous studies of AQM, uses a complex network topology including up to 3 congested links, reverse path traffic, and realistic round-trip times. We consider the effects of multiple AQM routers on the throughput of long-lived FTP flows and HTTP response times of web traffic. We find that, especially for web traffic, performance is improved as the number of AQM routers is increased, but that significant improvements occur with even a single AQM router. I. I NTRODUCTION Active Queue Management (AQM) techniques were de- signed to keep router queues small by providing early con- gestion notifications to end systems. The most prominent of these techniques is Random Early Detection (RED) [13], first proposed in the early 1990s and used today in many commercial routers. RED has been the subject of much study and discussion [4], [5], [8], [17], yet there is still a question of the effectiveness of RED, especially in environments heavily loaded with web traffic [8], [17]. In the past few years, two other AQM techniques have become particularly popular: the Proportional Integral (PI) controller [16] and Random Early Marking (REM) [3]. Both of these techniques were developed with the goal of improving the stability of RED and involve decoupling the probability of congestion notification from the average queue length. Many AQM techniques can be coupled with Explicit Con- gestion Notification (ECN) [21] enabling routers to provide congestion notification by marking segments instead of drop- ping them. If an incoming TCP segment from an ECN-enabled connection is chosen for marking, the congestion bit is set in the segment’s header, and the segment is forwarded to its des- tination. The destination will set a bit in the acknowledgment (ACK) to signal congestion to the sender. The receipt of a marked ACK is taken as a sign of congestion in the network, and the TCP sender reduces its sending rate by half. Many previous evaluations of AQM techniques [4], [8], [17] have used a simple dumbbell topology that contained a single congested link. This type of topology is not representative of the wider Internet. Anagnostakis et al. [2] have shown that a non-negligible number of packets on the Internet experience multiple congested links on the path from sender to receiver. A series of fluid flow-based studies of AQM techniques were performed that included scenarios with multiple congested links [15], [18], [20]. These studies focused on the queue sizes obtained when the traffic was comprised of long-lived TCP flows. In contrast, we perform an empirical evaluation of the performance of both web traffic and long-lived flows over multiple congested links and consider end-user metrics, such as goodput and HTTP response time. Following the advice of Floyd and Kohler [14], we strive to create as complex a simulation scenario as is feasible, while still being able to effectively evaluate the performance of the AQM mechanisms. We have a mix of short and long- lived flows, each with a different RTT, facing reverse-path traffic, and traversing multiple congested links. We realize that this scenario does not completely model the Internet – no simulation scenario can claim to do this. We do argue that this simulation setup is more representative of some Internet path than one with a dumbbell topology, a single congested link, a narrow range of RTTs, and only long-lived flows. We are interested in investigating the impact of multiple AQM routers in an environment with multiple congested links. We consider situations where all routers in the path use AQM and where only some routers in the path use AQM. Our experiments measure the performance of end-to-end HTTP and FTP traffic that encounters congestion caused by HTTP cross-traffic entering the network at three different points. We designed the experiments so that the level of cross-traffic entering each congested link is similar. End-to-end flows receiving congestion notification from any of the congested routers are expected to reduce their sending rates, reducing the load on all of the links. However, short-lived HTTP flows may be too short to react to congestion notifications. We include several FTP flows in the traffic environment in order to evaluate the effect of multiple AQM routers on long-lived TCP flows. Previous studies [8] have shown that some parameter settings of AQM techniques, especially RED, perform well for shorter flows, but harm longer flows. The performance of the FTP flows is representative of the performance of longer HTTP flows. The evaluation of the performance of both short-lived HTTP flows and long-lived FTP flows will highlight the fundamental tradeoff between response time and throughput. Response times are typically a function of queuing delays and loss rates. The smaller the router queues, the lower the queuing delays and the less likely that a packet will be dropped due to buffer overflow. In order to keep queues small, TCP senders must be alerted to reduce their sending rates when the router