Network quality of service is traditionally thought to be provided by a combination of scheduling in the network nodes to enforce a capacity sharing policy and traffic controls to prevent congestion that could annihilate that policy. The work presented herein is instead based on an end-to-end argument: A capacity sharing policy is enforced by traffic controls in the hosts at the edges of the network, without any scheduling support in the network. Our proposal is to add a feed-forward control at

more »

... the transport layer to provide a service that is better suited to conversational and streaming applications than the batchoriented transfer mode provided by TCP. The paper presents the control and its evaluation: We compare the sharing of capacity between traffic classes and study the loss rate seen by admitted streams. The outcome is that the new control adds a distinctly different service to the service offered by TCP for the Internet.