Comparing Different Approaches to the use of DiffServ in the Internet

The existing Internet appears to provide good quality service for a very wide range of services, possibly because the TCP protocols aim to achieve fair queueing, or processor sharing. The DiffServ architecture aims to do better than this by providing different performance standards for different classes of service. The natural way to apply DiffServ is to allocate classes in accordance with the urgency or priority of the requests. However, another approach is to use DiffServ to allocate service classes according to the "size" of the requests, where the concept of "size" can be defined in a variety of ways: total bytes in a flow, rate of a flow, or by a series of token buckets. We use simple queueing models to investigate how much improvement in performance could be obtained by implementing this service discipline, and whether, as a consequence, it is unnecessary and perhaps even dangerous to assign classes of service in accordance with the type of service requested. The results suggest that shortest job first offers considerable advantages over processor sharing. Thus in spite of the difficulties of identifying the size of flows it may be worthwhile to consider how something like shortest job first can be implemented. On the other hand it appears that other priority queue strategies, not based on job size, are risky, in that the marginal advantages gained by favoured jobs are very small, and the majority of jobs can expect to suffer worse response times.