Prioritization of Contention in Modern Packet Switched Networks
In the context of any new Quality of Service (QoS) model for access networks, a prioritization model must be developed for all packet switched traffic in the network. This wiki refers to this as a "contention priority." In addition to contention priority, other prioritization mechanisms may be used within the network such as service admission, traffic shaping, and queue management. These other mechanisms are typically used on the traffic that is isolated within a particular contention priority. In other words, the contention priority creates the service isolation characteristics for any QoS model.
Queuing Points and Contention Points
All packet switched systems share common characteristics in their architecture. Specifically, routers and switches use two primary components for storing and relaying packets. These are known as queuing and contention points, respectively. Queuing points hold the packets in memory as they are classified to determine traffic management policies and output port destination. Contention points act as the switched or routed output path for the packets. Typically, cross-bar switch fabrics or more sophisticated shared memory schemes might be used to realize the routing architecture. If Active Queue Management (AQM) is used, or other explicit packet drop mechanisms are used within the packet switch, ancillary systems within the queuing points typically determine which packets to drop. However, contention points also may drop packets due to congestion. If one output becomes congested due to contention for service and there is not sufficient queuing capacity, packets will be dropped.
Packet Loss and Latency
Therefore, the two primary QoS characteristics of any network are determined by the behavior of queuing and contention points. In the absence of AQM, contention points are responsible for the packet loss and queuing points for the delay. Any network which is tiered to offer differing levels of service quality will first determine the packet loss and delay characteristics it wishes to offer to flows of various types. The overall derived bandwidth characteristics such as average transfer rate for TCP flows, etc. will be largely determined by the underlying packet loss and delay characteristics. In essence, ISPs must follow a decision tree with regard to their QoS model and the decisions made will have implications on both transparency measures for broadband consumers (i.e. "network management policies") as well as the desired types of service they wish to offer their customers. All QoS characteristics are thus derived from the underlying delay and packet loss policies assigned to contention points and queuing points in the network. Contention priority becomes the dominant factor when determining packet loss for a particular classified service type or flow.
Individualized (per flow) Quality of Service versus Class of Service
QoS models come in two forms: per flow and class. In a per flow model, each uniquely classified set of flows on the network may be assigned a contention priority. In a class-based approach, flows are grouped. Flows in each generalized class share the same contention priority. Individualized per flow models typically involve dynamic pricing models for highly granular service. If the network can determine the level of service available within specific time instances, then bandwidth brokers or other dynamic pricing agents and mechanisms can be used to offer service. Currently (in 2012), there are no known commercial offering using highly granular per flow service models. Class of service models are much more popular in terms of implementation and commercial availability. Many IPTV and Voice over IP networks use class of service approaches since they offer bulk, non-granular, service for their applications. In general class of service models are easier to manage while per flow approaches may offer more fine grain bandwidth that can be associated with dynamic pricing applications.
Priority and Packet Loss Economics
Since queuing points dominate the delay and contention points dominate the loss characteristics of the network, economic frameworks for premium services may be built around derived services from this basis.
Implications for Broadband "Erlang Formulas"