Project name: Measuring, modeling and cost allocation for quality of service (M2C-QoS)
Research Group: Probability and Stochastic Networks (PNA2) 
Coordinator: Prof.dr. M.R.H. Mandjes
Members: Michel Mandjes, Rudesindo Nunez Queija.
Partners: Telematica Instituut (M. Bijlsma and H. Eertink), Universiteit Twente (A. Pras, R. van de Meent, B. Nieuwenhuis).

To ensure Quality of Service in IP networks, roughly speaking two approaches seem viable. The first approach uses QoS-enabling mechanisms for protecting traffic streams, and builds on numerous initiatives developed within the Internet Engineering Task Force. The initially proposed architecture, Intserv, relied on QoS guarantees made on a per flow basis, and has consequently scalability problems. Therefore Intserv may be applied in the edge of the network (where the number of flows is relatively low), but not in the core. Another mechanism is Diffserv, in which agreements are made for aggregates of flows rather than micro-flows, hence solving the scalability problems. Despite considerable research efforts, however, is still hardly used in operational environments. The above-described problems with the implementation of 'active' QoS-enabling techniques led to an approach that is of a more passive nature, usually referred to as overprovisioning. Overprovisioning does not rely on QoS mechanisms in the network, but rather on proper dimensioning of the network resources. This dimensioning relates in particular to link rates: these have to be chosen such that the load imposed very rarely exceeds the link capacity; in this approach buffers are mainly used to absorb packet-level queueing effects. All traffic streams are treated in the same way (i.e., no prioritization mechanisms are involved) -- consequently overprovisioning is particularly appropriate if there is no serious heterogeneity among the sources with respect to the QoS requirements. The current practice in IP backbone links rough provisioning procedures are the common way to deliver QoS. The required capacity is usually based on 5-minute average values of the traffic rate on the network link, adding some safety and growth margins.

Although the idea of overprovisioning is simple, it poses interesting questions. For the cases a traffic model of the network is available, standard methods from queueing theory can be used to find an appropriate value for the link capacity. However when only coarse traffic measurements are available (rather than a full traffic model), no standard machinery has been developed yet.

Clearly, without sufficient provisioning, the performance of the network (in particular from a user's perspective) will drop below tolerable levels. By overprovisioning too much, the QoS levels hardly improve anymore, and are consequently not cost-efficient. This leads to the question what the lowest capacity figure is, at which additional capacity does not improve the service level. This level of overprovisioning seems a reasonable compromise between simplicity (no complicated QoS-enabling techniques) and cost-efficiency, and is therefore of crucial interest for network operators. We coin the term `efficient overprovisioning' to denote provioning of resources at the above described `threshold level'.

As mentioned above, operators ususally base dimensioning decisions on coarse measurements, of the time-scale of 5 minutes or even longer. Evidently, this is a relatively long time-interval when speaking of QoS. Hence, the question is whether these measurements give any useful information on the capacity needed: QoS degradation experienced by the users may be caused by fluctuations of the offered traffic on a much smaller time scale, e.g., seconds or even milliseconds.

In the light of the discussion above, the goals of the present project are twofold. (i) In the first place we investigate the relation between traffic fluctuations on small and large time scales. (ii) In the second place, using the relations derived under (i), we develop appropriate, theoretically supported methods/guidelines for efficient overprovisioning. These dimensioning guidelines are based on standard (coarse) traffic measurements (i.e. measurements on a relatively large time scale, say, 5 minutes), but give guarantees on considerably shorter time-scales (order of seconds). Hence, the procedures proposed require relatively low measurement effort.