<title>Predictive loss pattern queue management for Internet routers</title>

Abstract

Widespread availability of real-time services is the next challenge for the Internet after the introduction of the WWW has already changed Internet tra c patterns once. As the Internet provides a best e ort" datagram service only, n o assurance for actual packet delivery for real-time ows can be given. Most real-time applications exhibit tolerance against occasional loss of packets, but are sensitive to losses that occur in bursts. This is especially signi cant for a voice service, which we consider as our primary target application in this paper due to its importance in the future Internet, its relatively well-known subjective properties in the presence of packet loss and its simple ow structure. Currently all hop-by-hop approaches to enhance the Quality-of-Service for real-time ows either use strict per-ow setup and state maintenance of reservations Integrated Services or rely on the sender ingress router which is unaware of the amount and the location of congestion in the network to mark packets for preferred treatment Di erentiated Services. This results in either high resource consumption in the network due to a conservative characterization of the application's requirements or dissatifactory perceived quality due to the toleration of too many burst losses in the network. For interactive v oice, end-to-end adaptation to the current network load is also di cult to apply, considering the per-ow o verhead and usual tra c properties low bitrate.We propose an active queue management algorithm called PLoP Predictive Loss Pattern, that instead of enforcing static reservations tries to give preferred service over short time intervals to particular ows that have previously been discriminated i.e. lost packets. The probability of a packet drop is made dependent o n o w history previous drops and ow type. This allows a basic protection of burst-loss-sensitive ows during transient and persistent congestion, keeping only partial per-ow state. The proposed algorithm does not require any speci c cooperation of the applications and gives an incentive to build loss-resilient applications using end-to-end QoS enhancement FEC and or error concealment, which can build on the enforced predictive loss pattern.Simulation results show the performance at a congested network element in terms of conditional and unconditional loss probability and processing state overhead. It it shown that burst loss protection in the data path for periodic foreground tra c voice is feasible for a wide range of load conditions. Any impact on non-real-time tra c in terms of the conditional and unconditional loss probability i s a voided, as long as the link-speed equivalent bu er is larger than the maximum expected tra c period.