Revisiting the fair queuing paradigm for end-to-end congestion control

Abstract

ongestion control has been a central research topic since the early days of computer networks. Nagle first identified the problems of congestion in the Internet [1]. The fundamental turning point in Internet congestion control took place in the 1980s. Nagle proposed a strategy based on roundrobin scheduling [2], whereas Jacobson proposed a strategy based on slow start (SS) and congestion avoidance (CA) [3]. Each of these solutions has its drawbacks: Nagle's solution has high computational complexity and requires modifications to the routers; Jacobson's solution requires the collaboration of all end users. The modest performance of the routers and small size of the Internet community at that time led to the adoption of Jacobson's proposal. SS and CA were incorporated into TCP, and more than 10 years later the Internet still uses these mechanisms in a somewhat improved form [4].We define the notion of a paradigm for congestion control as a model to be used in designing congestion control protocols that have the same set of properties. Practically, when one designs a congestion control protocol with a paradigm, one has the guarantee that this protocol will have the same set of properties as all the other congestion control protocols designed with this paradigm. The benefits of the paradigm come from the set of properties it guarantees. However, the trade-off is that the paradigm imposes some constraints that must be respected. This notion of a paradigm is not obvious in the Internet. A TCP-friendly paradigm was implicitly defined and introduced only a few years ago, when new applications that cannot use TCP had already been developed.Since TCP relies heavily on the collaboration of all the end users (collaboration in the sense of the common mechanism used to achieve congestion control), the TCP-friendly paradigm was introduced (see [5,6]) to design congestion control protocols compatible with TCP.A TCP-friendly flow must adapt its throughput T according to the equation:where C is a constant, MTU is the packet size used, and RTT and loss are the round-trip time and loss rates experienced by that flow. To compute the throughput T, one needs to measure the loss rates and the RTT. The TCP-friendly equation models the TCP long-term behavior for low loss rate. Padhye et al. [7] introduced an improved TCP-friendly equation that is a good approximation of TCP long-term behavior, even for high loss rate.The throughput T for a TCP-friendly flow must decrease when its loss increases. However, this behavior does not suit the needs of many applications. For example, audio and video applications are loss-tolerant and the degree of loss tolerance Revisiting the Fair Queuing Paradigm for End-to-End Congestion ControlArnaud Legout and Ernst W. Biersack, Institut EURECOM Abstract Today, the dominant paradigm for congestion control in the Internet is based on the notion of TCP friendliness. To be TCP-friendly, a source must behave in such a way as to achieve a bandwidth that is similar to the bandwidth obtained by a TCP flow that wo...

An Internet friendly transport protocol for continuous media over best effort networks

“…Another approach suggests the use of per-flow scheduling mechanisms that separately regulate the bandwidth used by each best-effort flow [6]. Legout and Biersack in Reference [7] suggested a solution to the problem by using a fair scheduler network. They also proposed a TCP-friendly protocol [8] to multicast audio/video content to a large number of heterogeneous receivers.…”

PLM

“…In the context of a formal study of the congestion control problem as a whole, we defined the Fair Scheduler (FS) paradigm (see [8]). We define a Fair Scheduler to be a Packet Generalized Processor Sharing scheduler with longest queue drop buffer management(see [14], [20], [4], and [3] for some examples).…”

“…The main constraint of the FS-paradigm is the deployment of FS routers. However, we explained in [8] how and why this deployment is feasible per ISP. The only assumption that the paradigm makes on the end-user is its selfish and non-collaborative behavior (we do not require these properties, we just do not need anything else to achieve the properties of an ideal congestion control protocol).…”

“…PLM does not require any signaling nor feedback. The burden of PLM is distributed among the receivers, vi) Feasibility: PLM requires a FS network, this issue is discussed in [8] and it is shown to be feasible per ISP (Intemet Service Provider). PLM simply requires an IP-multicast network.…”

“…We want a congestion control protocol for multimedia and file transfer applications that guarantees fast convergence and does not induce losses. We introduce in [8] aparadigm to devise end-to-end congestion control protocols only by taking into account the requirements of the application (congestion control protocols tailormade to the application needs). Our paradigm is based on the assumption of a Fair Scheduler network i.e.…”