Traffic classification and shaping in ATM networks

Abstract

The integration of services is the driving force behind the design of the high speed data networks of today and tomorrow. These networks must be able to deliver a broad range of services and be capable of carrying diverse classes of traffic with very different source characteristics. In the case of data traffic, some delay is tolerable, however the loss of information is not. At the other extreme, some loss is tolerable for voice traffic, however delay is not. In the middle is broadcast quality video traffic, which is sensitive to both delay and loss.

In the case of Asynchronous Transfer Mode networks, the solution of these conflicting requirements is to negotiate a Traffic Contract at the User-Network Interface, which specifies a Quality of Service level and the characteristics of the source. These characteristics are used by Call Admission Control and Usage Parameter Control to protect existing connections.

Unfortunately, the determination of source characteristics by either the user or network provider is difficult, or impossible in some situations. The usual statistical methods of identifying traffic sources do not scale well to high speed networks, nor are they applicable to all traffic types. In addition, they cannot be used to identify the traffic streams emerging from applications not envisioned when these identification techniques were developed. Thus, there is a need for a method that can accurately provide a description of traffic streams in a timely manner. Three contributions are presented in order a satisfy these needs.

The proposed traffic primitive classifier can be used to classify unknown traffic streams. This is accomplished by defining simple, deterministic characteristics of traffic streams which are collectively called traffic primitives. These traffic primitives are used to define training vectors in order for a neural network to learn the classification problem. The traffic classification results show that the neural networks not only can classify deterministic sources from which they are trained, but also they can classify a wide range of random sources, such as the class of on-off sources. With the additional functionality of Traffic Primitive Histogram Identification and Stream Transition Tracking, the primitive classifier can be applied to characterize sources which are not on-off in nature. As well, the primitive classifier can be integrated into a policer to perform more complex policing actions, and to monitor traffic streams for a given set of occurrences.

In addition to the traffic primitive classifier, two additional contributions come in the form of two traffic shapers, the Minimized Variance shaper and the Burst-oriented shaper. Both shapers have the ability to produce near deterministic streams, given appropriate sources are shaped, at fairly low costs in delay and buffer size at the shapers. In the case of the Minimized Variance shaper, source information is utilized in order to find an optimal shaping parameter that has the effect of minimizing the interdeparture time variance of the stream exiting the shaper. For the case of the Burst-oriented shaper, source information is not required since it assumes that bursts and silences emerge from the ATM Adaptation Layer, and so it attempts to spread a burst into the immediately following silence period. By doing so, it has the ability to define an unshaping parameter, which when embedded into the traffic stream, can be used to unshape the source at the destination User-Network Interface. This has the dual benefits of offering the network provider an ability to characterize sources and hence improve network efficiency, and also to allow users to treat the network as a transparent connection.