Providing end-to-end performance guarantees using non-work-conserving disciplines

A non-work-conserving server is one that may be idle even when there are packets available to be sent. Since they do not provide the optimal average performance, non-work-conserving disciplines were seldom studied in the past. For the guaranteed performance service class in integrated services networks, the main performance index is the end-to-end delay bound, instead of the average delay. Providing end-to-end delay bounds in a networking environment is difficult. While bounding delay in any server requires a bound on the input traffic, complex interactions among traffic streams usually distort the traffic pattern so that traffic inside the network is different from the source traffic. Previous techniques of bounding end-to-end delay in a networking environment usually start from the source traffic characterizations and iteratively “push” the traffic characterizations through the network. This does not only require non trivial analysis, but also has several important limitations. In this paper, we show that non-work-conserving disciplines greatly simplify the analysis in a networking environment and overcome most of the limitations of previous work by controlling traffic distortion inside the network and thus allowing a single node analysis to be extended to arbitrary topology networks.