We analyze the operation of two scheduling algorithms, class based queuing (CBQ) and delay bounded hybrid proportional delay (DBHPD). We compare results obtained from simulations and measurements in a reasonably equivalent setup and state that considerable differences can be observed in the results produced by these two methods. We also perform kernel profiling to investigate the resource overhead caused by these algorithms. Kernel profiling results support the observation that deviations between simulations and measurements are mainly caused by complex measurement and calculation procedures that generate extra delays in real implementations. As a conclusion we suggest criticism to be used when making judgements about algorithm performance based on simulation results.

In this paper we show the first working implementation of an adaptive, measurement-based scheduling algorithm called delay-bounded hybrid proportional delay (DBHPD) in a FreeBSD based ALTQ prototype router. We describe how we have implemented DBHPD and discuss what kind of difficulties were encountered in the implementation. We present measurement results of the DBHPD implementation with FTP, HTTP, Video Streaming and VoIP traffic in underload, overload and heavy overload conditions. We show that the algorithm operates well according to the theoretical model and preserves the desired delay-bound as well as the delay ratios between the classes. We also compare DBHPD to an existing Class-Based-Queueing (CBQ) algorithm that is widely used both in research and in the industry. We show that DBHPD is able to achieve at least as good link utilization as CBQ and in addition results in much better and predictable differentiation in terms of delays and more controlled packet losses.

In this paper, we study how well packet and flow level measurements can differentiate application types from the traffic mix. The results of this work show that we are able to divide traffic into three classes without any a priori knowledge of the traffic types. These classes closely resemble the three categories of traffic that are found in the Internet today. The externally controlled unidirectional traffic flows, the long held adaptive reliable traffic flows and the short duration transactions. Results indicate that it may be feasible to build a system where the network is given responsibility to classify traffic independently.