TCP throughput guarantee in the DiffServ Assured Forwarding service: what about the results?

TL;DR

This paper reviews the challenges of guaranteeing TCP throughput in DiffServ Assured Forwarding networks, highlighting that current conditioning schemes are inadequate and the problem remains unresolved.

Contribution

It provides an overview and classification of recent proposals for TCP throughput guarantees in DiffServ/AF, emphasizing the limitations of existing conditioning mechanisms.

Findings

Standard DiffServ conditioners are poor TCP traffic descriptors.

Most proposals focus on edge conditioning schemes.

TCP remains incompatible with DiffServ/AF despite various proposals.

Abstract

Since the proposition of Quality of Service architectures by the IETF, the interaction between TCP and the QoS services has been intensively studied. This paper proposes to look forward to the results obtained in terms of TCP throughput guarantee in the DiffServ Assured Forwarding (DiffServ/AF) service and to present an overview of the different proposals to solve the problem. It has been demonstrated that the standardized IETF DiffServ conditioners such as the token bucket color marker and the time sliding window color maker were not good TCP traffic descriptors. Starting with this point, several propositions have been made and most of them presents new marking schemes in order to replace or improve the traditional token bucket color marker. The main problem is that TCP congestion control is not designed to work with the AF service. Indeed, both mechanisms are antagonists. TCP has the property to share in a fair manner the bottleneck bandwidth between flows while DiffServ network provides a level of service controllable and predictable. In this paper, we build a classification of all the propositions made during these last years and compare them. As a result, we will see that these conditioning schemes can be separated in three sets of action level and that the conditioning at the network edge level is the most accepted one. We conclude that the problem is still unsolved and that TCP, conditioned or not conditioned, remains inappropriate to the DiffServ/AF service.