Statistical bounds on the drop probability of assured forwarding services in DiffServ interior nodes under the processor sharing scheduling discipline

Abstract

This paper addresses the problem of modelling and analysing an interior node in IETF's DiffServ services model. Specifically it is concerned with the performance of the DiffServ assured forwarding service category in presence of a premium service class. In this model, the network node shares its outgoing link capacity between a premium service representing the expedited forwarding (EF) per-hop behavior, and two classes of assured service, that represent two classes of the assured forwarding (AF) per-hop behavior. It this paper, the traffic is modelled as Markov modulated fluid sources, and we focus on a system where out of profile traffic is dropped at the edge of the network thus both AF queues support only one drop precedence. Using a decomposition approach, approximations, and spectral analysis, we are able to derive upper and lower bounds on the tail of the distribution of the buffer content for both AF classes given a generalized processor sharing scheduling is used to differentiate the two classes. Such approximate analysis of the interaction between traffic classes can help to achieve a better understanding of this type of networks; enables the provision of throughput differentiation as defined by the AF PHB through the GPS scheduler while quantifying delay; and finally helps simplify greatly the design of bandwidth brokers that do not rely on long term bandwidth (over) provisioning.