Fairness of Traffic Controls for Inelastic Flows In the Internet ⋆ Dah Ming Chiu ∗ Adrian Sai-Wah Tam Department of Information Engineering, The Chinese University of Hong Kong, Shatin N.T., Hong Kong Abstract In best-effort networks, fairness has been used as a criterion to guide the design of traffic controls. The notion of fairness has evolved over time, from simple equality to a form of equality modulated by the user’s need (e.g. max-min and proportional fairness). However, fairness has always been defined on a per-user basis for a deterministic workload. In this paper, we argue that we must redefine the notion of fairness when we study traffic controls for the co-existence of elastic and inelastic traffics. Our results indicate that subjecting inelastic flows to fairness congestion control on a per-flow basis does not necessarily maximize the network’s utility. Instead, inelastic flows may follow their own form of traffic control, such as admission control (without congestion control). At the aggregate level, our results indicate that it still makes sense to maintain a balance between elastic and inelastic traffic. In order to support our arguments, we develop a methodology for comparing different traffic controls for given utility functions and different workloads, both deterministic and stochastic. Key words: congestion control, admission control, fairness, utility maximization, non-convex utility function, stochastic traffic model 1. Introduction Internet is a connectionless network. It relies on congestion control implemented in the end-systems to prevent offered load exceeding network capacity, as well as evenly allocate network resources to different users and applications. In the past, the applications (such as email, file transfer) were predominantly elastic, or in other words flexible in their bandwidth requirements. The Internet architecture served these applications well. As the Internet shifts to support increasing volume of multimedia applications, there has been continuous debate on the next generation Internet architecture. Various proposals have been made for building a multi-services network to support different types of applications, for example ATM [1], Integrated Service [2], Differentiated Service [3] and MPLS-based traffic engineering [4]. Despite such efforts, a prevalent belief is that the Internet’s connectionless service model does not need to change, as long as there is adequate provisioning of network bandwidth. What is needed instead is a relaxed end-system ⋆ This work was partially supported by University Grants Committee, Hong Kong SAR under RGC grant 4232/04E (project 2150420) and AoE grant AoE/E-01/99 (project 4801312). ∗ Correspondance author. Email addresses: dmchiu@ie.cuhk.edu.hk (Dah Ming Chiu), swtam3@ie.cuhk.edu.hk (Adrian Sai-Wah Tam). congestion control that co-exists with the widely used TCP congestion control. So what is a suitable alternative congestion control for multimedia applications to practice? The orthodox solution requires all applications to share network bandwidth fairly, as existing TCP flows do. If an application needs more bandwidth than the prevailing fair share, then it should adapt its bandwidth (down) in favor of fairness. It is recognized that multimedia flows need more gradual adaptation to fair bandwidth share, so the effort of designing the alternative control is focused on a smooth transient response in bandwidth adaption. Proposals of such congestion control schemes are generally referred to as TCP-friendly congestion controls in the literature [5–7]. The thesis of this paper is to argue for abandoning the traditional notion of fairness in designing end-system traffic controls for different types of applications. Unlike the case with elastic traffic, the best way to deal with congestion for inelastic traffic should be some form of admission control, which is by definition unfair in the traditional sense. Furthermore, it is important to consider the stochastic nature of bandwidth allocation, rather than the allocation of bandwidth to a fixed number of flows. The fact that flows arrive at different times and have different demands for the network should also be taken into account in fair bandwidth allocation. While we argue against insisting on per-flow TCP- Preprint submitted to Elsevier 25 January 2007