A comparison of queue management algorithms for intra-flow loss control

Multimedia flows are inherently inhomogenous, i.e. packets within a flow are of different importance for actual user perception. When transmitting such flows with real-time constraints in networks like the Internet which offer no reliability of transmission, some packet loss is inevitable. The perceptual impact of these losses is then amplified by the arbitrary distribution of packet losses within the flow which impairs the reconstruction of application data units (ADUs) at the receiver, as well as the performance of end-to-end loss recovery mechanisms. To control the loss distribution within a flow (“intra-flow” QoS) typically filtering higher-layer information within the network is proposed, which is both expensive in terms of resources, as well as undesirable with regard to network security. We compare two novel queue management algorithms which improve the intra-flow QoS without higher-layer filtering. The first algorithm is called DiffRED (differential random early detection). It differentiates between packets marked by the sender as either more or less eligible to be dropped in comparison to unmarked packets without keeping any per-flow state. The other algorithm called PLoP (predictive loss pattern) operates without per-packet marking, yet keeping partial per-flow state. We introduce simple metrics to describe the loss process of individual flows and present simulation results with voice as foreground traffic using the proposed methods in a multi-hop topology. We find that both algorithms do not have a significant impact on the background traffic. For the given scenario algorithms using packet marking are found to be superior because for the foreground traffic a high probability for short bursts with potentially high perceptual impact can be traded against a higher probability for isolated losses as well as higher (but acceptable) probability for very long loss bursts