Burst scheduling networks

Most application-level data units are too large to be carried in a single packet and must be segmented for network delivery. To an application, the end-to-end delays and loss rate of its data units are more relevant measures of performance than ones specified for individual packets. From this observation, we introduced the concept of a burst (which subsumes the concept of a block in the ATM literature). A flow is modeled as a sequence of bursts, each of which models a sequence of packets that encapsulate an application data unit. We describe an approach towards designing integrated services packet-switching networks that provide QoS guarantees to bursts. We present a burst-based flow specification, an architecture and algorithms for packet scheduling, and tight bounds on end-to-end burst delays. In particular, we illustrate how to exploit the burst-based flow specification to improve implementation efficiency. We describe how burst scheduling networks can be designed to provide a real-time VBR service with no loss and, with burst-based admission control, a real-time VBR service at a specified loss rate.