Early Fair Drop : A New Bu er Management

In this paper, we investigate the eeectiveness of packet-drop mechanisms in conjunction with fair queuing link scheduling and hierarchical link sharing. Under fair queuing, the link share of a ow changes dynamically due to the arrivals and departures of ows and their bursts. This phenomenon becomes more pronounced in the case of hierarchical link sharing. Packet-drop mechanisms play an integral role for bandwidth-adaptive ows, such as TCP, that are expected to adjust their rates to the ows' changing fair share of the link bandwidth. We show experimentally that, under the existing drop policies (including random early detection and per-ow schemes such as the longest-queue drop) implemented with fair scheduling policies, TCP ows are slow to adjust their rates to their changing share of the link bandwidth. To overcome this problem, we propose a new packet-drop policy that simultaneously exploits two dimensions| when to drop and what to drop. We demonstrate that, under our packet-drop policy and with fair queuing, TCP ows adapt to their changing fair share of the link bandwidth when competing with diierent types of cross-traac (e.g., bandwidth adaptive, rate controlled, greedy and on-oo). We also illustrate that our drop policy provides isolation and fairness to ows other than TCP (e.g., rate controlled, on-oo).