Abstract: Active queue management (AQM) refers to the process of running a decision procedure in a router when a packet arrives to determine if the end-system generating the packet should receive a signal that the network is congested. Active queue management in Internet routers has been proposed as a solution to some of the scalability issues associated with TCP's pure end-to-end approach to congestion control. However, with traditional AQM schemes, an IP signaling protocol, explicit congestion notification (ECN), is required to achieve good performance. We have shown empirically that with ECN, on congested links, loss rates, link utilization, and user response times for web-like transfers are significantly improved. Without ECN, it is difficult for AQM schemes to outperform a simple drop-tail FIFO queue.
This talk presents our analysis of state of AQM research. From this analysis we present empirical evidence that ECN is required to enhance AQM results because ECN ameliorates a set of flaws in most AQM designs. Specifically, most AQM designs treat all flows identically when deciding whether or not to signal congestion. We are experimenting with an alternate approach, called differential congestion notification (DCN), that attempts to distinguishes between "small" flows and "large" high-bandwidth flows, and only provides congestion notification to large flows. The challenge is to efficiently classify flows with minimal state and appropriately determine when to signal congestion. Experiments in a laboratory testbed demonstrate that, for interactive web traffic, DCN outperforms the other AQM designs that require ECN. We believe this result is significant because it provides evidence that the promise of AQM can be realized without the cost and new protocol deployment issues inherent in the use of ECN.