TUF: tag-based unified fairness

Abstract

Abstract-Finding an appropriate end-to-end congestion control scheme for each type of flow, such as real-time or multicast flows, may be difficult. But it becomes even more complex to have these schemes be friendly among themselves and with TCP. The assistance of routers within the network for fair bandwidth sharing among the flows is therefore helpful. However, most of the existing mechanisms that provide this fair sharing imply complex buffer management and maintaining flow state in the routers. In this paper, we propose to realize this fair bandwidth sharing without perflow state in the routers, using only a trivial queueing discipline. Packets are tagged near the source, depending on the nature of the flow. In the core of the network, routers use FIFO queues, and simply drop the packet with the highest tag value in case of congestion. Contrarily to other stateless fair queueing algorithms in the core routers, we do not try to maintain instantaneous flow rates equal. Instead, we take into account the responsiveness nature of the flows, and adjust loss rates such that average rates are equal. The novel approach of our scheme, called ¡ UF , Tag-Based Unified Fairness, not only improves the overall fairness but enables us to maintain it in realistic environments, with non-negligible round trip times or bursty traffic, where other schemes fail. The corresponding cost is the need for models of the end-to-end responsive natures of the flows.

A Control-Theoretical Approach for Fair Share Computation in Core-Stateless Networks

“…The method used in CSFQ for estimating fair share α is simple but does not function well when the aggregate incoming traffic is bursty [6], [7]. In fact, in order to minimize the negative effects of buffer overflow, Stoica included a simple heuristic whereby each time the buffer overflows, α is decreased by a small fixed percentage, taken to be 1% in his simulations [3].…”

“…Through extensive simulations, CSFQ was found to approach the fairness of Deficit Round Robin (DRR) [2] and offer significant improvements over FIFO and Random Early Drop (RED) [5]. Other similar algorithms based on the same SCORE/DPS architecture, like Rainbow Fair Queueing (RFQ) [6] and Tag-based Unified Fairness (TUF) [7], were subsequently proposed.…”

“…Unfortunately, CSFQ is known to function poorly under bursty traffic [6], [7], which can be attributed to the ad-hoc method used in its fair share estimation. Our main contribution in this paper is the use of a control-theoretical approach that can provide a fast and robust computation of the fair share value.…”

Controlling bursts in best-effort routers for flow isolation

“…In CSFQ, dropping is function of the flow rate information of the packet and of an estimated fair share rate. Other examples of similar schemes are [1,3]. The main improvement of this kind of approaches is that the multiplexing function of buffers is protected by avoiding that flows using less than their fair share suffer packet losses.…”

Assuring fair allocation of excess bandwidth in reservation based core-stateless networks

“…Therefore, both per-flow queuing mechanisms [7] [8] [9] and core-stateless [6] [10] [11] [12] approaches were developed for provisioning fair bandwidth distribution.…”