SMARTA

Wireless data networks (in particular ‘Wifi’) are proliferating at a rapid pace. Due to this out-growth, there is a concomitant interest in retrofitting wireless technology into the fabric of everyday human life. We are gradually seeing a blurring of the line between the types of applications that wireless data networks can support and the purpose for which they were originally built, i.e. providing last-mile Internet connectivity. Indeed, city-wide wireless mesh deployments across the globe are an example of these trends, all contributing to the race towards wireless broadband. Emerging technologies such as IEEE 802.11n and WiMax are also expected to be at the forefront of the wireless movement.

Resulting from this exponential growth, two important observations can be made. Firstly, multiple wireless technologies are simultaneously competing for market share. Over the years, many wireless technologies have established niche areas in different markets. For instance, Bluetooth is more commonly referred to as a Personal Area Network (PAN) technology, with dominant applications in-the-home. On the other hand, WiFi technology has found application both in-home as well as in outdoor environments. As a result of the growth of different technologies, it is not unreasonable to expect that future wireless networks will comprise a multi-tiered architecture, with users roaming across different technologies. Multi-mode phones sporting Bluetooth, WiFi, and cellular capabilities are already available in the market. Due to the overlaid nature of wireless infrastructure, it will become important to implement policies and mechanisms to switch between different technologies, as well as properly understanding and inferring system context. Thus, mobile clients will need to incorporate the necessary intelligence in order to decide the appropriate action to take in each type of situation.

The second observation is that most of these wireless technologies also compete for a common shared wireless spectrum. Conventional wireless systems don’t coordinate spectrum usage, causing interference and contention to drastically reduce network performance. To prevent these problems from occurring, spectrum usage needs to be properly coordinated, to prevent different wireless technologies from ‘stepping on each others toes’. This leads us to our second requirement. Infrastructure devices (i.e. APs) will also need to be smart about using spectrum and will also need to properly infer system context. They will need to detect the presence of other technologies and take appropriate measures to avoid RF interference.

Research Goals

Given these observations, we outline the following research directions:

• Intelligent Self-Managing Access Points: Traditionally, management of wireless infrastructure has been a matter of manually configuring radios (e.g. frequency, power, radio sensitivity) to optimize performance. This is typically a one-time process, where client workloads, interference patterns, and other environmental patterns are assumed to be fixed. Real-world deployments, however, have shown this not to be true. Furthermore, existing techniques to model wireless propagation are overly-simplified and fail to accurately capture the multi-dimensional impact of the RF environment on network performance. To address this problem, we advocate a dynamic self-configuring¹ approach to network deployment and maintenance. As the RF environment evolves, our system periodically re-tunes itself to maximize performance 1 (http://blizzard.cs.uwaterloo.ca/keshav/home/Papers/data/05/succ-refinement.pdf).
• Building Intelligent Mobile Clients: User-specific policies (e.g. cost, performance, security / privacy) may dictate how a user accesses different networks. These policies constitute an intelligent software system running at the client end that can dynamically make such decisions, based on the sensed context. Our group has consceived an initial design (http://blizzard.cs.uwaterloo.ca/tetherless/index.php/OCMP) of such an architecture and we are currently evaluating different policies that might prove useful in a heterogeneous network environment.

Our current research directions are focused on supporting infrastructure intelligence, augmenting primitives of intelligence within the access point infrastructure, to dynamically control AP parameters. The architecture we propose is called SMARTA, and is described further.

SMARTA: A Self Managing Architecture for Thin Access Points

We propose a solution that dynamically adjusts WLAN parameters to optimize performance. Our approach is centralized. In other words, all deployed access points are physically connected to a centralized controller, running on a desktop PC. The controller coordinates the radio parameters of the access points. Because access points don’t autonomously decide their respective radio configurations, they are termed thin APs. The access points themselves are layer 3 devices, implemented using boxes from Soekris Engineering, running the 2.6 Linux kernel. Each Soekris houses an IEEE 802.11b/a/g compatible radio. The overall layout of the architecture is shown above. Our solution is novel because of the following features:

• No End-client Modifications: We do not require feedback from clients, and in fact, can have clients remain completely unmodified. Thus, our architecture is able to support legacy clients.

• Efficient Interference Detection: Most empirical methods to measure RF interference require injecting control traffic into the network, to gauge its impact on data traffic. This incurs overhead (i.e. bandwidth, delay, etc). Instead, we propose conducting a series of short tests that take only a few tens of milliseconds, and which can consistently detect interference in a WLAN environment. Because these tests operate on such short timescales, we periodically run them to update our interference model (i.e. conflict graph).

• Realistic Channel Considerations: Due to the empirical nature of our approach, we are not forced to use abstract simplistic wireless propagation models as part of our system. This is a departure from existing work that bases itself on simplified propagation models.

Performance Modeling

Fundamentally, the goal of WLAN tuning is to maximize performance. Therefore, we must build a model which can be used to properly characterize performance. Because there are many diverse use cases (and applications) even for enterprise WLAN deployments, the performance metrics we would potentially like to capture may be quite diverse. Thus, to cater to any type of performance parameter, we use a utility-based model that conveniently decouples the process of WLAN tuning from the parameters that characterize performance. We extend this utility model to a graph-theoretic model called a conflict graph, to realize an annotated conflict graph that represents the performance parameters of the system. These parameters may capture features such as the impact of RF interference on overall network throughput. The figure immediately above shows an example of such a conflict graph.

Parameter Estimation: Active Probing

An integral part of our approach is detecting RF interference. Interference detection is done by sending active probes from access points. Because these probes are initiated by access points themselves, no additional hardware is needed to support this functionality. Simultaneous wireless access and interference-based probing with a single radio is possible with the use of two virtual channels, multiplexed onto a single physical channel. The first virtual channel acts as the control plane, while the second channel acts as the data plane. The control plane carries probe (or signalling) traffic, while the data plane carries normal data traffic. Switching between these channels occurs at fixed time intervals, and the frequency of switching is governed by a tuning parameter provided at the access point.

Performance Tuning

We use the model constructed above as a basis on which to build WLAN tuning algorithms. Modern-day wireless access points have a number of tuning parameters. We focus on channel assignment and transmit power control. We are currently working on refining our algorithms. For details on the algorithms, please request our paper (under submission).

Open Problems

There are a number of avenues for future work, in the context of infrastructure intelligence:

• Extensive Parameter Tuning: Apart from channel assignment and power control, other tuning knobs can also prove useful in optimizing WLAN performance. We are exploring ways to control and adapt CCA thresholds, user association, and data rates, in our model.
• System Scalability: Our experiments validate the efficiency of our approach for moderately sized deployments (e.g. 40-50 APs). However, we are currently exploring the scalability of our techniques to very large-scale deployments that scale up to tens of thousands of access points. This becomes important in dense network deployments that must support applications such as voice.
• Better Optimization Methods: We are currently also in the process of enhancing existing algorithms for access point channel assignment and power control. Potential future enhancements include constructing multi-objective optimization algorithms that quickly converge to the optimal configuration.
• Performance Model Extension: We are also exploring ways to enhance our performance model for the general case where different radio technologies interfere with each other. These radios may be based on different access link technologies, all operating in the same unlicensed band.
• Application to Mesh: Mesh networks are becoming increasingly popular for providing blanket city-wide coverage. There are also popular for rapid network deployment in trade-show scenarios and other events that require short-term wireless coverage. We are exploring ways to extend our ideas and models for such mesh networks. In particular, we are interested in seeing how quickly our proposed architecture can adapt to the RF environment, which is especially useful for short-term mesh deployments.

For more details on our architecture and techniques, please request our paper (under submission)

Waterloo Wireless Testbed Deployment

We are in the process of deploying a 40 node wireless testbed in the Davis Research Centre (DC), at the University of Waterloo. A link to our online monitoring site is given here (http://breeze.cs.uwaterloo.ca/swannet/).