Evaluation

Figure 5: The network stack on each DenseAP node

We have built a prototype implementation of the DenseAP system. Figure 5 illustrates the network stack on each DAP. The network stack enables AP functionality on ordinary desktop machines. An integral part of the stack is our software AP (SoftAP), a fully programmable AP for the Windows platform. The wired and wireless interfaces are bridged. Each DAP also runs a DenseAP daemon, a user-level service responsible for managing local access point functionality. The service periodically queries the SoftAP driver, and sends summaries of client statistics to the DC. It also receives commands from the DC and sets appropriate parameters in the driver.

Figure 6: The testbed. The area is roughly 32m x 35m. The rooms have full walls, and solid wood doors.

The DAPs are off-the-shelf PCs running Windows Vista, and the networking stack described above. Each machine is equipped with a Netgear JWAG511 wireless NIC. The wireless NICs are based on a chipset from RealTek. They support operation in 802.11 a/b/g modes, with one limitation: in the 802.11a mode, they can operate only on the lower 8 channels (channels 36-64). We also found that the RealTek cards do not work reliably in promiscuous mode, so we use an additional radio on each DAP to simulate promiscuous mode. This second radio is not fundamental to our approach, and is used only to compensate for the shortcomings of the RealTek card. All of the DAPs are connected to the same IP subnet on their wired Ethernet link. The DC is an ordinary desktop-class machine.

The DenseAP testbed is deployed on a portion of our office floor, as shown in Figure 6. The testbed consists of 24 DAPs. The DAPs are deployed roughly in every other office. Within each office, the machine is placed on the floor; the exact location determined by the consent of the occupant. This area of our building is served by a single corporate WLAN AP. The AP is located roughly at the center of the area, and is placed on the ceiling. Note that the DenseAP deployment is 24 times denser than the corporate WLAN. In addition to the DAPs, we have also deployed 24 machines to serve as clients. The clients are a mix of ordinary desktop and laptop machines, equipped with a variety of off-the-shelf wireless NICs.

Most of the the experiments reported in this paper were run in 802.11a mode (5 GHz band). There is very little corporate traffic in the 802.11a band. Thus, for most experiments, the background traffic is negligible, and we did not need to run the load estimation algorithm. The 802.11g band does see a fair amount of usage during normal office hours, but to avoid impacting corporate traffic, we were limited to conducting 802.11g experiments outside of normal work hours. We now turn to evaluating the performance of the DenseAP system using this testbed. We begin by validating the rate-map approach (section 4.1.2), which lies at the heart of our association and load balancing algorithms.