We then present a channel width adaptation algorithm, called SampleWidth that takes into account the properties of different channel widths. This algorithm is based on a simple search process that builds on top of existing techniques for adapting modulation. Per specified policy, it can maximize throughput or minimize power consumption. Evaluation using a prototype implementation shows that SampleWidth correctly identies the optimal width under a range of scenarios and can provide up to 60% more throughput compared to the best fixed-width configuration.
Much wireless communication today involves the use of channels with preset widths. A wireless channel is the frequency spectrum block over which nodes transmit; it is uniquely specified by its center frequency and width. The use of preset channel widths is a direct result of how the available spectrum is divided by existing wireless technologies. For example, in 802.11 (WiFi) b/g, the spectrum block is divided into 11 overlapping channels that are 20 MHz each and are separated by 5 MHz. WiFi nodes communicate over one of these channels.
In this paper, we argue that nodes in WiFi networks should adapt the width of the communication channel based on their current needs and environmental conditions. To our knowledge, such adaptation has not been proposed or explored before. We find it surprising that WiFi nodes dynamically change many variables today to improve communication, such as center frequency, transmission power, and modulation, except one of the most fundamental variable -- the channel width.
We make our case in three steps. First, using measurements from controlled and live environments, we study properties of different channel widths. We use commodity WiFi hardware manufactured by Atheros for this study. With software modifications alone, we get these NICs to communicate at 5, 10, and 40 MHz channels in addition to the standard 20 MHz. We find that different widths perform differently on many measures of interest. Narrower channels have lower throughput but they have longer range, are more resillient to multipath delay spread, and consume less power.
Figure 1: Indoor range for two modulations as a function of channel width.
Figure 1 illustrates this range benefit of using narrow channel widths. The experiment involved a UDP transfer between two laptops with Atheros WiFi cards. The -axis represents the communication range of the two laptops using different channel widths. We use an office as unit of distance and define range as the minimum number of offices crossed at which the loss rate between two nodes is 100%. The offices are of identical size, and there are 8 offices in a straight line. This experiment clearly demonstrates that by using narrow channels we can communicate over farther ranges. Systems can now be designed to use narrow channels when communication over a narrow channel is not possible. Narrow channel widths offer several such exciting possibilities, which we discuss in detail in Section 4.
In the second step, based on the findings, we identify several unique benefits of dynamically changing channel width that are otherwise not available today. For instance, in times of low throughput requirement, nodes can simultaneously increase range and reduce power; in fixed-width systems, these two highly desirable properties are perenially in conflict. Another example is that total network capacity can be increased without increasing spectrum usage, by splitting multiple flows that share a wide channel into narrower channels. Yet another example is that nodes can substantially improve throughput by adapting channel width, because different widths offer the best throughput in different conditions.
Realizing these benefits requires practical channel width adaptation algorithms. In the third and final step, we design a channel width adaptation algorithm, called SampleWidth, for the base case of two communicating nodes. For efficient search and sampling, SampleWidth builds on top of existing techniques for adapting modulation. We have prototyped SampleWidth on top of the same Atheros WiFi NICs. Our experiments show that its simple adaptation scheme correctly identifies the optimal width in a range of distances between the sender and receiver. In our mobile experiment, SampleWidth improves throughput by more than 60% compared to the best fixed-width system.
We varied the channel width by changing the frequency of the reference clock that drives the RF front end and baseband. Wi-Fi chipset designs incorporate a RF transceiver and a baseband/MAC that operate with a common reference clock. The baseband/MAC uses the reference clock to control access to the wireless network by regulating timing, encryption, encoding/decoding, and data transmission. The RF transceiver uses the reference clock to drive the radio's Phase Locked Loop (PLL). Therefore, one way to change the width of the transmitted signal (i.e, the channel width) is by modifying the clock frequency. For another node to successfully receive the transmitted signal, its PLL clock needs to be tuned to the same frequency (and hence its channel to the same width).
We implemented this technique on off-the-shelf Atheros-based NICs. These cards use a clock frequency of 20 MHz to generate a 20 MHz wide signal. The value of the clock frequency can be configured in multiples of 2 using a hardware register. We changed the register values to generate signals on four channel widths of 5, 10, 20, and 40 MHz.1
Figure 2: A few 802.11 timing parameters for different channel widths.
|Since the reference clock is shared by other components in the wireless card, including the baseband processor [12,4], slowing or increasing the clock rate affects various 802.11 timing parameters. For example, a 4 s OFDM symbol in 20 MHz channel width gives symbols of length 2 s in 40 MHz, and 16 s in 5 MHz. Similarly, a 400 ns OFDM guard interval at 40 MHz is 3.2 s at 5 MHz. We list a few important parameters that have different values at different channel widths in Figure 2.|
Figure 3: Screenshot of the spectrum analyzer showing 20MHz, 10MHz and 5MHz signals.
Figure 4: Properties of Channel Widths
Figure 3 shows the effect of our changing the PLL clock rate on a spectrum analyzer screenshot on which different widths have been overlaid. It can be seen that while the center frequency for all widths during this measurement was 2412 MHz (corresponding to Channel 1 of IEEE 802.11 b/g), the channel width changes.
In this section, we characterize the impact of channel widths on three of the key metrics of wireless communication: flow throughput, packet reception range, and power consumption. In all cases, we explain the underlying reason for the observed behavior and how it differs from what may be expected. The findings of this section form the basis of our arguments for dynamic adaptation of channel width.
Figure 4(a) shows throughput obtained by a UDP flow when using different channel widths and modulations. As expected, the throughput increases as the channel width or the modulation rate is increased. According to Shannon's capacity formula the theoretical capacity of a communication channel is proportional to the channel width. Our measurements on commodity Atheros cards follow this relationship approximately but not exactly. The increase in throughput from doubling the channel width is less than a factor of two. This less-than-doubling behavior is due to overheads in the 802.11 MAC, such as various inter-frame spacings. Since some of these overheads are fixed in terms of absolute time, e.g., the slot-time is 20 s, their relative overhead for wider channels is higher.
Figure 4(b) shows the loss rate as a function of the attenuation for different channel widths. The modulation is fixed to 6 in this graph. We see that narrower widths are able to withstand greater attenuation, which implies that they can reach further. We define the range threshold of the signal as the minimum attenuation at which the loss rate is less than 10%. Then, we can see this threshold is 74 dB for 40 MHz and 81 dB for 5 MHz. As we discuss below, this 7 dB difference when going from 40 MHz to 5 MHz can be substantial because dB is a logarithmic unit.
The longer range of narrower widths can be explained as follows. The FCC limits the total transmission power of Wi-Fi radios. Transmission power of a signal depends on the channel width, which is measured in Hz, and power per unit Hz. Thus, at narrower widths, the radio can transmit with higher power per unit Hz without changing the total transmission power. Given equivalent noise per unit Hz across various widths, the SNR (signal-to-noise ratio) is higher for narrower widths, giving them a longer range.
We now quantify the effect of channel width on power consumption using a setup similar to the one used in . We connect a 0.1 ohm resistor in series with the wireless card, and measure the current drawn through the resistor using a data acquisition system. We compute the power consumed by the wireless card by multiplying the current drawn through the resistor with the voltage supply of the wireless card (5 Volts).
igure 4(c) shows the power consumed by different channel widths while idling, receiving, and sending packets. We present results for modulation 6, although, for the same channel width, the numbers were same across different modulations. The figure indicates a linear relationship between the channel width and the power consumption. We see that wider channels consume more power. The additional consumption from 5 to 40 MHz is around 40% while idling and receiving packets and is 20% while sending packets. Thus, substantial powers savings can accrue from switching to narrower channels when appropriate.
The previous subsection shows that substantial benefits can be had by dynamically adapting channel width. But realizing those benefits relies on practical adaptation algorithm. In this section, we present such an algorithm.
Figure:5 Outline of SampleWidth algorithm
Our algorithm is called SampleWidth and it enables two nodes to dynamically select a channel width according to their workload and optimization criterion (e.g., throughput or energy consumption). This scenario forms the base case for channel width adaptation. It is of interest by itself in several settings: two personal mobile devices (e.g., an iPod) sharing media content; a link in a multi-hop mesh network where the two nodes have a dedicated radio to talk to one another; and in 802.11 infrastructure networks where the AP has multiple radios on different widths and the client dynamically selects the best width.
Consider two nodes, and . They have at their disposal different channel widths . The goal of the algorithm is to select a channel width according to a given objective. SampleWidth uses a state-of-the-art SampleRate autorate algorithm to find an efficient data rate on a specifc width and then searches across widths. In addition to reducing the dimensionality of the search, this process enables us to search across widths less frequently and across rates more frequently. SampleWidth is based on sampling only adjacent (i.e., the next narrower or wider) widths. It samples adjacent widths and switches if the sampled throughput is higher than the current throughput. Further, it probes the adjacent wider channel only if the probability of disconnection is low, i.e., if the average data rate on the current width is high. The algorithm is outlined below in Figure 5. The optimality and convergence of the algorithm, as well as optimization for energy are described in depth in the full version of this paper 
In summary, we showed the following properties of channel widths:
Figure:6 Comparison of throughput achieved using SampleWidth with that of static width schemes in indoor settings.
Fixed channel width systems face a hard choice between increasing
range and reducing power consumption. They can increase range only
by increasing transmission power and reducing transmission power
reduces range. Adaptive channel width systems can have both!
Narrower channels have both lower power consumption and longer
range. Reducing channel width may come at the cost of reduced
throughput, however, and so the width must be reduced when the
additional throughput of the wider channel is not desired. Though,
as our results above show, in some case narrower channels can
improve throughput as well.
Even without the range advantage, it is well known that Wi-Fi drains significant battery power . Research has looked at numerous techniques to reduce battery power, such as using a low power radio , aggressive power saving modes , or transmission power control . We propose another powerful mechanism - channel width adaptation - to reduce the power consumption of wireless devices.
Figure 7: Instantaneous and cumulative energy usage for different channel width configurations.
Based on these findings, we make a case that wireless networks should dynamically adapt the width of the channel. We propose such an adaptation algorithm, SampleWidth, that builds on existing rate adaptation techniques in order to select the width that can provide the best throughput (or most power savings). We perform a representative evaluation, where in two laptops equipped with commodity Atheros cards are enabled to operate on 5, 10 and 40 MHz in addition to the standard 20 MHz capability. One of the laptops was configured to be the sender, and using this we transfered a 25MB file to the receiver. The sender was also configured to use the SampleWidth in the last case, and we compare the throughput and power savings obtained in comparison to fixed-width configurations. Figure 6 shows the evaluation results of SampleWidth for indoor settings, averaged over three runs. There are two important take-aways from this graph: a) there is no one channel width that can provide the best throughput at all distances and b) The proposed algorithm SampleWidth achieves optimum throughput within an error of 8% owing to the switching overheads.
Not only does adaptation achieve the best throughput, but we obtain power savings as well. Figure 7(a) show the power consumption behavior in detail for all configurations at the sender. The fixed width systems start out at their idle mode power consumption, move to their send mode consumption level, and then come back to their idle mode levels. SampleWidth starts out at the idle mode level for 5 MHz, because that is least costly. When the transfer starts, it moves to the the power consumption level of 40 MHz, because that yields the least power-per-byte ratio. When the transfer finishes, it comes back to the 5 MHz level. Figure 7(b) shows that through this adaptation, SampleWidth is able to consume the least amount of energy, approximately 25% less than the fixed width schemes.
Having explored the basic capabilities provided by adaptive channel width, we now discuss how adapting channel width can be used to improve wireless networks.
For instance, one way to reduce load on an AP with many close clients is to reduce its power so that it serves fewer clients; but this can potentially cause clients close to this AP to associate to distant APs, which hurts performance [3,10,8]. Similarly, while adapting center frequency helps minimize interference [7,9], it does not reduce load on APs with many close clients. In a sense, these solutions try to alleviate the symptoms, rather than solving the cause of the problem. In contrast, adapting channel width can provide a more direct and conceptually cleaner solution. APs could be dynamically allocated channels of different widths (centered on different frequencies) where the width of an AP's channel is determined based on traffic demand of clients near it and the number of interfering APs in its vicinity. We have proposed a spectrum allocation scheme for this problem in . Preliminary results show that this scheme significantly improves total network throughput and prevents starvation for clients that connect to popular APs.
Early experiences with city-wide wireless mesh networks suggest that their total capacity is an important limitation, e.g. [16,17,18]. Adapting channel widths can help alleviate the capacity problem. When different channel-widths are allocated to ``links'' based on the traffic they carry, many links may operate on a narrower channel, leaving more spectrum for the heavily-loaded links. Also, as multiple flows on narrow channels provide higher overall throughput than a single wide channel the total capacity of the mesh backhaul can be improved.
In this paper, we demonstrate for the first time how--using standard, off-the-shelf hardware--the channel-width of IEEE 802.11-based network communication channels can be changed adaptively in software. We also show how it is beneficial to adapt the channel width and propose algorithms that enable adaptation.
While we show that signficant gains can be obtained from adapting channel widths, Several hardware and software challenges must be met to fully realize the benefits of adapting channel width that we uncover in this paper. On the hardware side, the most useful capability would be for radios to be able to decode packets at different widths (on the same center frequency). This capability would eliminate the coordination cost from channel width adaptation and allow nodes to unilaterally adjust width. We are currently working on various aspects that would make adaptive width systems deployable in existing networks.
The complete version and demonstration of this work has appeared in ACM SIGCOMM 2008 and Microsoft TechFest 2008. This work has been featured in slashdot and online press such as seattlepi.