A Simulation Study on the Relation of DiffServ Packet Level Mechanisms and Flow Level QoS Requirements

We deepen our previous analytical studies to include simulations, with special considerations on the effect that the metering mechanisms at the DiffServ boundary nodes have on differentiation. The two metering and marking mechanisms compared are token bucket and exponential weighted moving average (EWMA). The conjecture made in our analytical models is that the EWMA principle marks packets per flow, while the token bucket is a per packet metering and marking mechanism. We consider a single bottleneck link of a DiffServ cloud loaded by a fixed number of elastic TCP sources in congestion avoidance mode and with a similar round trip time (RTT). The sources are divided into L groups according to price paid by flows. At the boundary of the DiffServ cloud, the traffic of these TCP sources is conditioned, i.e. metered and marked to I priorities, i.e. flow aggregates based on price paid. Inside the DiffServ node the packets of the flow aggregates are forwarded or discarded by a scheduling unit that includes a single buffer with I priority levels. As a result, we give further hindsight to the effect that separate DiffServ traffic-handling mechanisms have on the fairness perceived by the user. We use the division of bandwidth and the division of packets into priority levels as the decisive factors for our results.