Queue mechanisms for differentiation in the Internet

Abstract: This thesis addresses loss-rate differentiation in the Internet. Loss-rate differentiation can be offered by tagging packets with different levels of drop precedence. Multiple drop precedence levels can be used to provide relative service levels and to assure forwarding capacity in the Internet. Assured services offer forwarding qualities known beforehand. This makes them more predictable than relative services. Users are assumed to explicitly request an assured service from their network provider for traffic up to a specified rate. With relative service levels, users switch between these levels to find one that provides an appealing balance between forwarding quality and cost. The policy for drop precedence probabilities defines the type of loss-rate differentiation pro-vided. Sheltered loss-rate (SLR) differentiation is offered by strictly giving drops to traffic at high drop precedence levels. Sheltering means that traffic at a low drop precedence level is protected from losses caused by traffic at higher levels. Such protection is required for assured services. Relative loss-rate (RLR) differentiation is offered when drop precedence probabilities are rela-tively distributed between drop precedence levels. Offering fixed relations in these probabilities further refines RLR differentiation, resulting in proportional loss-rate (PLR) differentiation. This thesis defines three recommendations associated with providing loss-rate differentiation. Such differentiation can be created with queue mechanisms. We specify and evaluate the proper-ties of differentiating queue mechanisms that make them capable of meeting the recommenda-tions defined. These evaluations are made with simulations. Firstly, the total forwarding quality at congested links should not be degraded due to actions taken to create loss-rate differentiation. The total forwarding quality includes packet loss patterns and delay variations. These quality metrics are high when packet drops are delayed. When pro-viding loss-rate differentiation, drops can be delayed through only dropping packets as they ar-rive. Dropping packets from queues enables immediate drops. This thesis shows that with imme-diate drops less bursty loss patterns and lower delay variations are achieved than with delayed drops. Secondly, traffic at high drop precedence levels should always be given a useful share of avail-able forwarding resources. Such traffic may experience high loss-rates, but should not become starved. Traffic at high levels may become starved due to overloading of traffic at lower levels when creating SLR differentiation. Starvation can be avoided with proper control of low drop precedence traffic. This control may, however, fail due to changes in the network routing topol-ogy, inaccurate admission control, etc. To avoid starvation without relying on proper traffic con-trol, this thesis presents a new queuing mechanism that falls back from offering sheltering to providing RLR differentiation during overloading of low drop precedence traffic. Thirdly, PLR differentiation should be predictable. Users should be able to predict the change in loss-rates when switching between drop precedence levels. PLR differentiation requires relations in loss-rates to be fixed to pre-configured target ratios. Running estimates of loss-rate ratios can be used as feedback to adjust towards these targets if the actual loss-rate ratios deviate from the target ratios. To provide predictable PLR differentiation, these estimates need to be accurate and stable at varying traffic loads. Moreover, they need to detect traffic load variations rapidly. This thesis presents a loss-rate estimator that provides accurate, stable and rapid detection of loss-rate ratios at varying traffic loads.