A quality of service (QoS) mechanism for multimedia applications that can tolerate packet loss is described in this paper. The proposed scheme extends the previous mechanisms by trading QoS for scalability.
The authors classify real-time traffic from server to client (a stream) mainly based on its packet loss tolerance. Loss tolerance is defined as the maximum acceptable loss for a given number of packets within a stream. The traffic classification is done in two steps. First, a class is identified for a stream. Then, within each class, a group with the fewest members is chosen for the stream, if one exists. Otherwise, a new group is created for the new stream, and one of the previous groups with the highest packet loss joins this new group to avoid an imbalance of stream distribution. Conversely, as streams leave, multiple groups are combined to reduce the number of groups. While the number of classes is fixed, the number of available groups within a class varies dynamically. As the authors point out, this reduces the amount of state information from per stream to per group, and, thus, makes their solution scalable.
Using simulations, the authors show that the proposed mechanism outperforms previous schemes with respect to QoS. However, the paper does not provide enough details on the simulation environment, and it is not clear if the authors’ workload assumption is representative of a multimedia application. The paper ends with a lengthy description of future improvements, which I found to be expensive and complicated. I believe a more detailed description of the simulation environment and workload, and a quantitative comparison of the implementation cost and complexity, would have been more helpful.