Introduction

There has been an enormous growth in the adoption of IEEE 802.11 wireless networks in the last few years. The ease of installation and the low infrastructure cost of 802.11 networks makes them ideal for network access in offices, malls, airports, cafes, hotels and so on. The widespread deployment of IEEE 802.11 networks means that a wireless client is often in the vicinity of multiple APs with which to affiliate. The selection of the AP that the client decides to affiliate with needs to be done carefully since it will dictate the client's eventual performance.

The conventional approach to access point selection is based on received signal strength measurements from the access points within range. However, it has been pointed out in several papers [1,2,4] that affiliation based on signal strength can lead to very bad performance for the end-host, since the signal-strength metric does not convey information regarding other attributes that affect end-host performance, such as the AP load and the amount of contention on the wireless medium.

In this paper, we describe how an end-host can take the aforementioned attributes into account while choosing an access point to affiliate with. We identify a metric that can capture the bandwidth that an end-host is likely to receive if it were to affiliate with a given access point, which we call potential bandwidth. The MAC-layer bandwidth offered by different wireless networks in the vicinity of the wireless client is a desirable metric as it takes into account the AP load, the contention on the wireless medium, as well as the signal strength.

In designing an affiliation algorithm based on potential bandwidth estimation, several constraints must be taken into consideration. The algorithm needs to be non-intrusive, i.e. it should not introduce additional overhead to the APs or their affiliated users. The algorithm should not require any changes at the AP side, if possible. More importantly, such an algorithm should be able to estimate potential bandwidth without previous affiliation with the several APs within range. Such a constraint minimizes the amount of time a client spends in the evaluation of the several choices it may have (since it does not associate and dis-associate with the different APs), while allowing for the continuous evaluation of AP performance even when an affiliation has taken place¹. The latter implies that a wireless client implementing the proposed functionality will be able to make more informed and efficient roaming decisions, continuously quantifying the performance of all APs in range.

In this paper, we propose a methodology for the estimation of potential bandwidth between a given AP and an end-host that fulfills the aforementioned requirements. The proposed methodology does not require the end-host to change its current affiliation and introduces very little overhead. Unlike [1,2], the affiliation algorithm proposed in this paper is end-host initiated and therefore, does not necessitate changes at the AP.

In a nutshell, our approach to potential bandwidth estimation relies on passive measurements of the timings of beacon frames sent out by an AP. Beacon frames are broadcast by APs periodically, and are used by APs to announce their identity as well as for the synchronization of the entire network. The delay between the time when a beacon frame is scheduled for transmission and its eventual transmission captures the load of the AP and the contention inside the network, conditions that the client would face if affiliated with that AP. The corresponding delay of data frames provides an estimate for the bandwidth a client will receive from the AP downstream. Upstream potential bandwidth estimation relies on frames sent by the client to the AP in the unaffiliated state and is based on a similar methodology that quantifies the respective delays.

Our technique can be used as part of an AP selection mechanism or for the evaluation of a wireless network's health. We evaluate its accuracy using controlled experiments in a low-noise environment. Preliminary experiments indicate that our approach yields fairly accurate estimates of the actual bandwidth from the AP to end-host, indicating that our approach looks promising.

The rest of the paper is structured as follows. In the next section, we describe related work. In Section 3, we describe our potential bandwidth estimation scheme. We discuss experimental results in Section 4. Finally, we conclude and describe in detail future directions in Section 5.