Flow level models of DiffServ packet level mechanisms

We present a simple flow level model that helps in understanding the effects of various DiffServ traffic conditioning methods on bandwidth sharing among elastic TCP flows. We concentrate on a single backbone link and consider a static case, where the number of TCP flows is fixed. In a Relative Service approach, the bandwidth share θ of a flow should be proportional to a weight φ assigned to the flow. It turns out that per packet marking (e.g. Leaky Bucket) yields a slightly better approximation compared to per flow marking (e.g. Exponentially Weighted Moving Average). For both marking schemes, however, the approximation is the better, the more there are precedence levels.