Introduction

In a typical office environment, the wired network is generally well-engineered and over-provisioned [12]. In contrast, deploying 802.11 wireless networks (WLANs) in enterprise environments remains a challenging and poorly understood problem. WLAN installers typically focus on ensuring coverage from all locations in the workplace, rather than the more difficult to measure properties such as capacity or quality of service. Thus, WLAN users commonly experience significant performance and reliability problems.

The usage model for enterprise WLANs is currently undergoing a significant transformation as the ``culture of mobility'' takes root. Many employees now prefer to use their laptops as their primary computing platform, both in conference rooms and offices [18]. A plethora of handheld Wi-Fi enabled devices, such as PDAs, cell phones, VoIP-over-Wi-Fi phones, and personal multimedia devices are becoming increasingly popular. In addition to the scalability challenges that arise with increased WLAN usage, the applications for many of these new mobile devices require better QoS and mobility support.

The need to improve enterprise WLAN performance has been recognized by both the research community [20,19,7,10,24] as well as industry [3,5,2,4,1]. Upgrades at the PHY layer, such as the transition from 802.11g to 802.11n, are important steps along the path to increasing WLAN capacity, but they are not enough. Deploying more APs has the potential to improve WLAN capacity, but one must also address issues such as channel assignment, power management, and managing association decisions.

In this paper, we present a new software architecture called DenseAP, that supports a dense deployment of APs to significantly improve the performance of corporate WLANs. A key emphasis in our design of the DenseAP system is on practical deployability. Because of the incredibly wide diversity of existing Wi-Fi devices, DenseAP must provide significant performance benefits without requiring any modifications to existing Wi-Fi clients. Furthermore, we do not consider any changes that require hardware modifications or changes to the 802.11 standard.

The DenseAP architecture and design challenge two fundamental characteristics of most current enterprise WLAN deployments. First, existing WLANs are designed with the assumption that there are far fewer APs than clients active in the network. In the DenseAP architecture however, the APs are deployed densely - in the common case there may be an AP in every office. Second, in conventional WLANs clients decide which AP to associate with, whereas the DenseAP system uses a centralized association control.

The scarcity of APs in conventional enterprise WLANs limits their performance in a variety of ways. For example, with a large number of non-overlapping channels (e.g. 12 in 802.11a) but only a few APs, the WLAN is unable to fully utilize the available spectrum at each location. Because radio signals fade rapidly in indoor environments, adding extra radios to existing APs is not as effective as deploying a larger number of APs in different locations. If APs are densely deployed, each client can associate with a nearby AP, and will see better performance. A dense deployment also reduces the impact of the ``rate anomaly'' problem [13] that hurts the performance of conventional WLANs.

With a dense deployment of APs, clients have many possible APs to choose from, and therefore access point selection policy is critical to achieving good performance. In conventional WLANs, clients select which AP to associate with using only locally available information. Most clients use signal strength as the dominant factor in selecting an AP, yet it is well-known that this behavior can lead to poor performance [14]. For example, when many clients congregate in a conference room, they all tend to select the same AP even when multiple APs operating on different channels are available. To improve performance in this scenario, clients must associate with different APs.

In the DenseAP architecture, a central controller gathers information from all APs, and then determines which AP each client should associate with. Simultaneously, the central controller also decides on the assignment of channels to APs. Even though Wi-Fi clients implement their own association policies and we do not modify these clients, the DenseAP controller effectively bypasses the client association policy by only exposing to each client the particular AP with which it wants the client to associate. Using a similar technique, the DenseAP controller also carries out periodic load balancing by seamlessly moving clients from overloaded APs to nearby APs with significantly less load.

The DenseAP architecture is quite versatile, and capable of improving many aspects of performance of enterprise WLANs. In this paper, we primarily focus on describing how DenseAP significantly improves the capacity of enterprise WLANs. We define capacity simply as the sum total of throughput all active clients in the network can potentially achieve. We will also briefly discuss how the architecture impacts other aspects of performance, such as quality of service for delay and jitter sensitive applications.

One obvious question that arises when considering a dense AP deployment is whether the performance gains justify the costs. One approach to reducing equipment costs is to leverage existing enterprise desktops and convert them to APs, similar to our previous work on DAIR [8]. However, the key concern of enterprise IT departments when deploying any new technology is typically the people costs associated with managing that technology. The DenseAP system is designed to require very little management overhead: the DenseAP nodes are self-configuring, and the redundancy available from the dense deployment means that AP hardware failures do not need to be addressed immediately.

This paper makes the following new contributions. First, our system supports a high density of APs with off-the-shelf, completely unmodified clients. As a result, it provides performance benefits for all clients, including the many different types of handheld Wi-Fi devices that have recently appeared. Second, we demonstrate the performance benefits of our system at a significantly higher density than previous work. Third, we demonstrate that intelligent management of the association process is necessary even when you have a very high density installation of APs. Forth, we present a novel load estimation technique that allows our system to automatically factor in impact of external interference, such as traffic from nearby networks.

We have deployed the DenseAP system with 24 APs in our offices. The testbed can function in both 802.11a and 802.11g modes. Our experiments show that the system provides large improvements in performance over the existing corporate network. In specific cases, the improvement in throughput can be as large as 1250%. We present a series of experiments that show how various aspects of our system work together to provide these gains. We also show that our system is capable of handling nomadic and mobile clients.