Improving flow level fairness and interactivity in WLANs using size-based scheduling policies

While 802.11-based Wireless LANs see increasing public deployment, their performances remain way below the ones obtained with a traditional Ethernet access. With 802.11 WLANs access, the bandwidth available at a given station can heavily fluctuate over time. In addition, the access point, with its limited resources and a probability to access the medium similar to any other wireless stations, is a source of major unfairness for TCP connections. This is especially the case when data flows in both directions, i.e., from and to the wireless stations. The problem stems from the firece competition that takes place at the access point between the non responsive TCP ack streams and the responsive TCP data streams. In this paper, we propose the use of size based scheduling policies at the IP layer of the access point to both enforce fairness among TCP connections and to improve the interactivity perceived by the end user. The latter is defined as the ability of the network to maintain small response times to the short flows that are generated by the interactive applications of the users, e.g., mail, or web browsing, when the overall load increases. We first demonstrate that the popular Least Attained Service (LAS) policy, which has been proposed in the context of the wired Internet, is able to work properly only when data flows in a single direction. To deal with the general case when uploads compete with downloads, we propose a new flavor of LAS, called LASACK. LASACK mitigates the impact of the non responsivness of TCP ack streams by assigning a priority to a TCP ack packet that is a function of the number of bytes sent by the corresponding data stream. We demonstrate using simulations under a variety of realistic workloads, that LASACK is able to enforce fairness and maintain a good interactivity even when the load increases. Another contribution of this work is to investigate the use of simple queuing models to capture the dynamics of the queue at the access point. In particular, we develop a queueing model for the LASACK policy that turns out to be accurate under a variety of loads and flow size distributions.