An Architecture for Extensible Wireless LANs

Today’s wireless LANs are a mess. The current 802.11 family of WLANs involves a jumble of competing standards, and a slew of implementations with varying degrees of interoperability and conformance to those standards. This situation has arisen in part from the need to innovate and evolve these networks over time, driven by new applications and increasing load. In the future we expect this situation to get worse, given the shift towards wireless as the dominant access network. Driving these changes are new devices such as Wi-Fi enabled VOIP handsets and audiovisual equipment, as well as new services such as Apple’s Time Capsule which performs background backups over wireless. In the longer term, we anticipate WLANs will become the default access network and will support filesystem and server traffic as well. Currently, it is not unusual for wireless LAN users to experience performance and reliability problems. A significant factor is the scarcity and poor utilization of the wireless spectrum, which suffers from a “tragedy of the commons”. Scaling up WLANs to meet new traffic demands, especially time-critical applications involving audio, video, or sensor telemetry data, is proving to be difficult. This is in spite of underlying innovations at the PHY layer, which largely address the need for more throughput, but not how that throughput is managed. Moreover, enterprises often have an interest in imposing customized policies on WLAN traffic, for example, prioritizing timeand safety-critical traffic over large file downloads. Existing wireless LANs make poor use of the wireless spectrum, largely due to the “intelligence” which is hard-coded into the vendor-specific software and firmware of wireless LAN clients. For example, WLAN clients control the decisions for AP associations, transmit power control, and physical data rate adaptation. The 802.11 standards specify the mechanisms, yet the policy is left entirely up to vendor-specific implementations. As a result, vendors view these areas as an opportunity to innovate and compete with each other. However, the end result of these attempts to innovate are limited, and we argue that is primarily an architectural limitation: by viewing the client as a stand-alone entity that solely uses local information about the devices it is interacting with, many important opportunities for improving the behavior of these algorithms cannot be realized. The current approach to innovation and evolution in wireless LANs is primarily through standardization, which has resulted in an alphabet soup of protocols within the 802.11 family. Certain Wi-Fi vendors offer vendor-specific WLAN extensions such channel bonding, or non-standard data rates to support communication with weak signals. Such extensions only work when both the AP and the client are using the same brand of Wi-Fi chipset and software drivers, which prevents widespread adoption. The downside to standardization is primarily the glacial progress in deploying new protocols. The standards process takes a very long time to reach agreement, and even after standards are ratified it takes a long time to replace and/or upgrade the wide variety of client equipment utilizing the infrastructure. We argue that to move away from the current mess, we need to rethink the basic architecture of wireless LANs. Our focus is not on changing the fundamental building blocks such as PHY-layer coding schemes or the CSMA nature of the MAC. Rather, we are interested in a developing an architecture that allows for extensibility, to ensure WLANs can adapt to meet future needs. We are guided by two key design principles: • Whatever we design today will be wrong in five years. If history is any guide, we cannot anticipate all future uses for wireless LANs. Furthermore, the best way to evaluate innovations is through actual deployments with real users. With current WLANs, deploying new hardware and upgrading NICs and drivers for all of the affected clients is an expensive proposition, not to mention the management and personnel costs involved. We argue that an extensible WLAN can adapt to new uses, and can allow rapid deployment and evaluation of experimental designs. • Infrastructure should manage the wireless spectrum. Networks can make the best use of resources by shifting much of the responsibility for managing the wireless spectrum (such as associations, power control, channel assignment, and physical layer rates) to the infrastructure, away from the individual clients. This has the additional benefit of making it easier to evolve the system because clients take much less of the responsibility for spectrum management. This approach also allows administrators to customize the network’s policies for handling different traffic demands. This paper describes Trantor1, a new architecture for wireless LANs. Trantor’s architecture is based on global management of channel resources, taking this responsibility explicitly away from clients and moving it into the infrastructure. To provide extensibility, the interface between the infrastructure and clients is simple and relatively low-level. Clients implement a small set of relatively simple commands which allows the complicated logic of the algorithms to exist primarily within the infrastructure. The commands fall into two categories: measurement commands allow the infrastructure to instruct clients to gather local information on channel conditions, such as RSSI from visible APs, and to report this information periodically; and control commands allow the infrastructure to control the behavior of clients, such as setting the transmit power or instructing a client to associate with a specific AP. Each client still implements a basic CSMA MAC for individual packet transmissions, but is otherwise not responsi-