A Flow Table-Based Design to Approximate Fairness

The current Internet architecture relies on congestion avoidance mechanisms implemented in the transport layer protocols, like TCP, to provide good service under heavy load. If routers distribute bandwidth fairly, the Internet would be more robust and could accommodate more diversity of end hosts. Most of the mechanisms proposed to accomplish this can be grouped into two general categories. The first category, which includes Fair Queueing (FQ [4]) and its many variants, uses packet scheduling algorithms that are more difficult to implement compared to FIFO queueing. The algorithms in the second category, active queue management schemes with enhancements for fairness (e.g., FRED [8], SFB [5]), are based on FIFO queueing. They are easy to implement and are much fairer than the original RED [6] design, but they don’t provide max-min fairness among a large population of flows. Recently, a router mechanism, AFD [14] (Approximate Fair Dropping), has been proposed to achieve approximately max-min fair bandwidth allocations with relatively low complexity. In this paper, we propose an implementation of AFD which can mimic the performance of the original design with much less state. I. BACKGROUND Approximate Fair Dropping (AFD) [14] is an active queue management scheme which uses a FIFO queue and drops packets probabilistically upon arrival. However, the decision whether to drop a packet (say from flow i) or not is based not only on the queue size but also on an estimate of the flow’s current sending rate ri. To achieve max-min fairness, the dropping function is defined to be di = (1 rfair ri )+. As a result, the throughput of each flow is bounded by the fair share: ri(1 di) = min(ri; rfair). Hence, drops do not occur evenly across flows but are applied differentially to flows with different rates. The key design aspects of AFD lie in the methods by which ri and rfair are estimated. To estimate ri, AFD uses the observation that, like the distribution of flow sizes, the distribution of flow rates is also longtailed, i.e. most bytes are sent by fast flows and a vast majority of flows are slow. For example, Figure 1 shows the cumulative distributions of the 1-second flow rates for three different traces; in these data sets, 10% of the flows represent 60% 90% of the total bytes transmitted. Therefore, a sample of the recent 0 100 200 300 400 500 RED FRED AFD T h ro u g h p u t (k b p s) Fig. 2. Background: Mixed Traffic throughput