The Choke Router Algorithm Provides an Approximately Fair Bandwidth Allocation at a Low Implementation Cost. the Authors Consider Performance and Implementation Issues Associated with Choke, with Particular Emphasis on Variable-length Packets

To provide high-quality service under heavy user loads, the Internet depends on congestion avoidance mechanisms implemented in the transport-layer such as the transmission-control protocol (TCP). However, many TCP implementations don’t include—either deliberately or by accident— a congestion avoidance mechanism. Moreover, a growing number of user datagram protocol (UDP)-based applications running on the Internet don’t back off properly when they receive congestion indications. As a result, these applications aggressively use more bandwidth than other TCP-compatible flows. Therefore, it’s necessary to have router mechanisms to shield responsive flows from unresponsive or aggressive flows and to provide good quality of service.1 All of the router algorithms (scheduling and queue management) discussed in the “Router algorithms” box have either provided fairness or were simple to implement, but not both simultaneously. In this article, we explore the CHOKe 2 (choose and keep for responsive flows, choose and kill for unresponsive flows) algorithm, which combines fairness and simplicity, and we address approximating byteby-byte fairness and implementation issues of the algorithm. Background CHOKe uses the observation that the FIFO-buffer contents form sufficient statistics about incoming traffic to penalize misbehaving flows in a simple fashion. The state, taken to be the number of active flows and the flow identification of each of the packets, is assumed to be unknown to the algorithm. The only observable data for the algorithm is the total occupancy of the buffer. Specifically, CHOKe calculates the average occupancy of the FIFO buffer using an exponential moving average window exactly as in the random early detection (RED) algorithm.3 It marks two thresholds on the buffer, a minimum threshold, minth and a maximum threshold, maxth. When the average queue size is less than minth, every arriving packet is queued into the FIFO buffer. When the average queue size is larger than minth, each arriving packet is compared with a randomly selected packet, called a drop candidate packet, from the FIFO buffer. Packets with the same flow identification are both dropped; otherwise, the randomly chosen packet remains in the buffer, and the arriving packet is dropped with a probability dependent on queue size. The drop probability is computed exactly as in RED. In particular, when the average queue size is greater than Konstantinos Psounis Rong Pan Balaji Prabhakar Stanford University THE CHOKE ROUTER ALGORITHM PROVIDES AN APPROXIMATELY FAIR