How to achieve fair differentiation

Abstract

We present a simple packet level model to show how marking at the DiffServ boundary node and scheduling and discarding inside a DiffServ node affect the division of bandwidth between two delay classes: elastic TCP flows and streaming non-TCP flows. We conclude that only per flow marking together with dependent discarding thresholds across both delay classes is able to divide bandwidth fairly, according to the load of the network, and in a TCP friendly way.