White Space Networking with Wi-Fi like Connectivity
and detailed experimental and
simulation results. Categories and Subject Descriptors: ABSTRACT:
Networking over UHF white
spaces is fundamentally different from conventional Wi-Fi along three axes: spatial variation, C.2.1 [Computer-Communication temporal Network]: Wireless variation, and fragmentation of communication the UHF spectrum. Each of General Terms: Algorithms, these differences gives rise to Design, Experimentation new challenges for implementing Keywords: white spaces, channel a width,Wi-Fi, dynamic spectrum wireless network in this band. access, cognitive radios We present the design and implementation INTRODUCTION of WhiteFi, the firstWi-Fi like system constructed on top The unused portions of the UHF of UHF white spaces. WhiteFi spectrum, popularly referred to incorporates a new adaptive as “white spaces”, represent a spectrum new frontier for wireless assignment algorithm to handle networks, spectrum variation and offering the potential for fragmentation, substantial bandwidth and long and proposes a low overhead transmission protocol to handle temporal ranges. These white spaces variation. WhiteFi builds on a include, but are not limited to, simple technique, called SIFT, 180 MHz of available bandwidth that reduces the time to detect from channel 21 (512 MHz) to transmissions in variable 51 (698 MHz), with the channel exception of channel 37. On width systems by analyzing raw November signals in the time domain. We 4, 2008, the FCC issued a provide an extensive system historic ruling permitting the use evaluation in terms of a of unlicensed prototype implementation devices in these white spaces [10]. In its ruling the FCC imposed an important Most of the prior research in requirement that white space UHF white spaces has focused wireless devices on accurately detecting the must not interfere with presence of incumbent RF incumbents, including TV signals broadcasts [14, 17, 18]. Recently, and wireless microphone researchers have mentioned that transmissions. This landmark they are ruling was beginning to look at the problem a result of extensive tests of establishing a wireless link performed by the FCC on white between space white space devices [8,12]. Our hardware prototypes that were research pushes the state-ofart submitted by Adaptrum, to the next level by going Microsoft, beyond a single link. We Phillips and Motorola. These identify the prototypes demonstrated feasible challenges of forming a UHF solutions white space network and show for an accurate and agile sensing how of incumbent signals [9]. to overcome them by presenting Permission to make digital or techniques, algorithms, and hard copies of all or part of this protocols work for backed up by extensive personal or classroom use is evaluation over a prototype granted without fee provided network that copies are as well as in simulations. We not made or distributed for focus primarily on the problem profit or commercial advantage of and that copies setting up a Wi-Fi like network bear this notice and the full consisting of an Access Point citation on the first page. To (AP) copy otherwise, to with multiple associated clients. republish, to post on servers or We leave the case of evaluating to redistribute to lists, requires multiple APs with multiple prior specific clients as follow-on work. Our permission and/or a fee. solutions SIGCOMM’09, August 17–21, are complementary to the 2009, Barcelona, Spain. ongoing work in the IEEE 802.22 Copyright 2009 ACM 978-1- Working Group [1], as we 60558-594-9/09/08 ...$10.00. discuss in Section 7. To appreciate the networking show that even a single packet problem, it is important to transmission causes audible understand interference during wireless the differences between white microphone transmissions. spaces and the popular ISM Consequently, bands where Wi-Fi devices both the AP and its clients must operate. First, in both bands disconnect and then there is rapidly reconnect using a spatial variation in spectrum different available channel availability, but the impact of (Section 2.3). this variation We have built WhiteFi, a UHF is higher in white spaces than in white space wireless network ISM bands. This is because that the FCC ruling requires non- adaptively configures itself to interference with wireless operate in the most efficient part transmissions of of primary users (incumbents) the available white spaces. In (Section 2.1). Second, since the following sections we the incumbents can operate in describe any portion of the white spaces, three major innovations that the network must be designed to allowed us to overcome the handle spectrum fragmentation, challenges with the possibility of each in networking white space fragment being of different devices. Briefly, our width. A contributions are: UHF channel is narrow (6 MHz _ A novel spectrum assignment wide in the US), and prior algorithm for managing variable research bandwidth communications. Our has shown that aggregating algorithm is unique in the contiguous channels improves way it addresses the dual throughput [15, 21]. challenges of spatial variation of Consequently, the network must available support variable spectrum and spectrum width channels (Section 2.2). fragmentation. We introduce a Third, RF transmissions in white new metric that leverages the spaces are subject to temporal available airtime measurements variations because wireless from each available UHF channel microphones to predict the available airtime can become active at any time when using multiple channels. without warning. Our 27821 m 187 m experiments _ A novel AP discovery a hardware prototype for white mechanism. These APs could be space networking. This platform using incorporates a Wi-Fi card, a any available channel width and UHF band converter, and a could be operating in any softwaredefined portion of the 180 MHz wide radio (SDR) [5]. We use the white spaces. We have designed KNOWS platform to extensively a new technique called SIFT, evaluate the quality and which is short for Signal performance of our innovations Interpretation and before Fourier Transform. SIFT design in WhiteFi. To the best of analyzes incoming our knowledge, WhiteFi is the signals in the time domain to first network prototype that detect transmissions over demonstrates the feasbility of different Wi-Fi channel widths without changing like networking over UHF white the channel width of the spaces. wireless card. It thus overcomes the core limitation with previous CHARACTERIZING WHITE approaches [15] that can only SPACES detect packets sent at the In this section, we discuss the same channel width. differences between the UHF _ A novel method for handling white space spectrum and the disconnections. Unexpected ISM bands where current Wi-Fi disconnections systems are a direct result of the operate. To understand the temporal variations described differences, we performed a set above. We leverage SIFT to of real-world measurements in significantly reduce the the UHF bands in several time to discover APs that have different switched to a different part of settings to characterize spatial the spectrum and we have and temporal variation. We also designed a new signaling analyzed mechanism publicly-available TV spectrum that allows clients to signal allocation data to understand disconnections to the AP without the distribution of UHF spectrum interfering with ongoing usage in urban, suburban, and transmissions over wireless rural microphones. settings in the United States. We We have implemented WhiteFi describe how each of these on the KNOWS [20] platform, characteristics has a substantial impact on the design of a UHF white space wireless network. much lower degree of 2.1 Spatial Variation fragmentation and more Television stations represent the contiguous spectrum largest incumbent use of the than urban areas. UHF spectrum. Across a wide white space AP, since area, the set of occupied TV communication ranges are channels expected to exceed depends on the location of TV 1 km [2]. We computed the transmitters as well as the Hamming distance, defined as number the number of channels available of stations operating in an area. at one location but unavailable However, spatial variation exists at on smaller scales as well, based another, across all pairwise on obstructions and construction buildings. Our results showed material. Wireless microphones, that the which are used in settings median number of channels ranging from small-scale lecture available at one point but rooms to large-scale music and unavailable sporting events, have typical at another is close to 7. This transmission ranges of a few statistic reveals significant hundred variation in meters [16]. For these reasons, spectrum availability within we expect significant spatial nearby buildings. While most variation incumbents in spectrum availability for detected in these measurements wireless network were TV channels, we also communications. found a few wireless To quantify this variation, we microphones. performed measurements of the The implication of this spatial UHF spectrum inside 9 buildings variation for a white space on our campus spanning an area wireless of approximately 0.9 km _ 0.2 network is that an AP (a home km, as shown in Figure 1. Note wireless router for example) this must not naively select entire area could be covered by a channel(s) to operate on based single suitably positioned UHF solely on its Figure 2: Expected spectrum own local observation of fragmentation after the US DTV spectrum availability. The AP transition in June 2009. Rural must take and suburban regions exhibit a into account the availability of spectrum at its clients as well. 2.2 Spectrum fragmentation dataset, we estimate UHF While the ISM bands are a spectrum fragmentation in 3 contiguous chunk of spectrum, settings: urban UHF (top 10 populated cities), white spaces are fragmented due suburban (10 fastest growing to the presence of incumbents. suburbs The size of each fragment can based on the 2007 Forbes list) vary from 1 channel to several and rural (10 random towns in channels. the The amount of fragmentation in US with a population less than the UHF bands depends to a 6000). Figure 2 shows a large extent on the density of TV histogram stations, which varies of the contiguous spectrum considerably widths that will be available in with population density. Rural each of (and suburban) areas, are likely these settings. As we see in the to have larger chunks of figure, in all 3 settings there is available UHF spectrum than at urban areas. least one locale in which there is In addition, the US digital a fragment of 4 contiguous television transition [7], channels scheduled to available, that is, 24 MHz of be completed in June 2009, will spectrum. In rural areas open up much more of the UHF fragments spectrum, as a number of analog of up to 16 channels are TV stations will stop operating expected. in A consequence of this these bands. fragmentation is that radios need To quantify the spectrum to tune fragmentation after the DTV the spectrum that they occupy to transition, fit within available fragments. we analyzed TV station data This implies the need for radios from TV Fool [4], a website that to use variable channel widths uses [15] sophisticated signal and terrain or channel bonding. Compared to modeling to estimate the Wi-Fi, the use of variable availability channel of TV channels at a given widths introduces two new latitude and longitude. Based on challenges. First, it makes this channel assignment more challenging, In both cases, we detected the since APs now occupy a range use of wireless mics at different of channels, rather than just one. times of day and for different Second, it increases the the durations. time taken for nodes to discover Wireless mics can be turned on APs. This is due to a limitation at any time. Since in its initial 28 ruling the FCC requires that of techniques that can achieve white space devices avoid variable channel widths on Wi-Fi interfering cards [15]. Using this technique with mic transmissions, both a radio can only decode packets clients and APs should detect the that are sent at the same channel presence of a mic on a channel width and same center and move away from that frequency. channel. An expensive switch of the PLL Furthermore, if only a client or clock frequency is required to an AP detects a mic, each must decode have a means of informing the packets at other channel widths. other of the channel switch In Section 4 we show how without WhiteFi overcomes these inducing interference. problems. Unfortunately, simple solutions 2.3 Temporal Variation to this problem are not feasible Finally, the UHF white spaces in practice. For example, one also suffer from temporal approach is for an AP to avoid variation, using in particular due to the channels where wireless mics widespread use of wireless might be used. However, simply microphones blacklisting known wireless mic (mics) – from lecture rooms in channels is overly conservative campuses to musicians at and makes inefficient use of the home, and from sporting events spectrum, since mics tend to to churches. We performed be used intermittently, for measurements limited durations, and on any of the UHF spectrum in two UHF white settings: the campus setting space channel. A more described earlier and a sophisticated approach would University dormitory, over build a historical several days. We database of mic usage patterns used the prototype described in that APs can query to determine Section 3 to determine the incumbents. the channels that are used at any an anechoic chamber. We instant in time. However, our measured the audio quality of measurements recorded show that mic use is highly speech transmitted over the unpredictable. For example, wireless mic with and without although use of wireless mics in UHF campus lecture rooms might transmissions. For UHF follow transmissions, we sent 70-byte a predictable schedule, each packets room tends to be over- every 100 ms on the same UHF provisioned channel as the mic. The with multiple mics on different transmission channels, and A/V operators power level was -30 dBm, which choose is below the FCC-permitted only a few of those mics for an maximum of 40 mW (16dBm). event. Furthermore, it is We found that the perceived impractical audio to predict many other uses of quality degrades with data wireless mics, such as for transmissions. The Mean special Opinion Score events, musical performances, or (MOS) of the received audio, rehearsals; each again may use a computed using Perceptual multitude of mics on various Evaluation channels. of Speech Quality (PESQ), The second possibility, which is decreased by 0.9 during the UHF being considered by the IEEE packet transmissions. Other 802.22 working group [1], researchers have shown that a involves an explicit channel MOS renegotiation reduction of only 0.1 is protocol between clients and APs noticeable by the human ear when they detect a wireless [22]. We note mic. This approach assumes that that these results may be worse control messages will not than what one would observe in induce audible interference on practice. In our experiments the the wireless mic. To test this antenna of the data transmitter assumption, and we performed an experiment by the mic receiver were within a placing a wireless mic few feet (in the same anechoic receiver along with our chamber). prototype WhiteFi device (Section 3) in We are actively working on 3. KNOWS HARDWARE acquiring an experimental PLATFORM license UHF white space networking Processor currently requires specialized Scanner hardware support, and several UHF hardware prototypes have been Translator reported UHF Antenna in the literature [8, 12, 20]. All Figure 3: Photograph of the these devices have a KNOWS hardware prototype. transceiver radio and a separate from the FCC and repeating scanner radio. The need for a these experiments in more separate realistic and scanner stems from the normal settings. requirement to quickly and One could argue that accurately interference in the beginning of detect the presence of primary a mic users. recording might not be cause for To support WhiteFi, we concern. However, when clients developed the KNOWS hardware are mobile, a mic may be sensed prototype only in the middle of a The hardware consists of three recording. components: a PC, a Furthermore, such a naive scanner, and a UHF translator. approach relies on all nodes in The PC is used both to control the vicinity the of a mic detecting its appearance scanner and to transmit and at the same time. If not all receive packets over the UHF detect the mic synchronously, bands. then each node transmits one The PC comes equipped with a after standard 2.4 GHz Wi-Fi card, the the other, thereby inducing antenna port of which is further interference with the connected to the UHF translator, mic. which These results demonstrate the downconverts the outgoing 2.4 need for a protocol that can GHz signal to the 512–698 MHz signal band. Incoming signals are the presence of a wireless mic to likewise upconverted and passed the network without interfering to the with the mic. We present such a Wi-Fi card. The center frequency protocol in Section 4.3. of the UHF translator is set from the PC via a serial control to solve the classic hidden interface. To ensure that the terminal problem, in which a TV outgoing is within transmission range of signal fits within a 6 MHz UHF the TV tower but the channel, we use the technique transmitting presented device is not. in [15] of changing the PLL Sensing of incumbents, clock frequency to reduce the especially microphones, is an Wi-Fi transmission bandwidth to actively 5 MHz. researched problem. Recent The scanner samples the UHF proposals use energy detection spectrum to detect the presence to detect of the primary users [17,18]. TV broadcasts and wireless However, such an approach is microphone signals. It is prone implemented to false positives, especially using the USRP [5] software- given the extremely low defined radio board coupled with detection a thresholds that have been set 50–800 MHz TVRX receiver- forth in the FCC report [9]. only daughter board. The False positives scanner reduce the amount of available scans UHF TV channels 21–51 in white spaces to form a white 6 MHz increments. Due to the space network. Furthermore, USRP bandwidth constraint [6], false alarms might cause WhiteFi the frequency span for each scan to is 8 MHz. We perform the FFT vacate the channels (Section on the PC, and using the feature 4.3). This switching overhead detection algorithms described might in [20], our scanner is able to affect the performance of detect associated clients. TV signals at signal strengths as Several solutions attempt to low as -114 dBm, and wireless avoid false positives. One microphones at -110 dBm. We approach note that this is much below the proposed from Motorola requires TV microphones to beacon at high Temporal Analysis power when they are being used (SIFT) [13]. Researchers from Berkeley This 30 dB detection buffer is have proposed collaborative required sensing to improve sensing accuracy [14]. The FCC is looking at the throughput than single-channel use of a geo-location systems 5. database to regulate and inform Signal Inspection before Fourier clients about the presence of Transform (SIFT): Existing primary systems detect signatures of users [9]. primary users in the frequency In this paper, we do not address domain, the problem of accurate after performing a Fast Fourier incumbent Transform (FFT) on the time detection, which remains an series active research area. Instead, we signal. However, such scanners focus on the networking cannot detect data transmissions challenges that arise assuming a for two reasons. First, in reasonably contrast to TV and microphone accurate incumbent detection transmissions, technique. We expect WhiteFi data transmissions are to benefit from future advances intermittent. Therefore, it is in incumbent detection. diffi- To enable efficient networking cult to distinguish intermittent over white spaces, our platform data from noise using prior has two key features unavailable detection in previous systems [8, 12, 20]: techniques. Second, data Variable Channel Widths: transmissions in our system can Existing systems can only use be sent one over multiple channel widths. UHF channel, even when Unless the entire signal is multiple contiguous UHF received, channels are including all subcarriers, data unoccupied. This is because the packets cannot be decoded. To bandwidth of the outgoing signal address is fixed to be 5 MHz. To support these concerns, we propose multiple contiguous channels, we SIFT, which processes raw modified the Atheros Wi-Fi signals driver using the techniques in the time domain and extracts presented data information from them. in [15] to transmit and receive We describe this technique in signals of bandwidth 5, 10 and detail in Section 4.2.1. 20 4. WhiteFi DESIGN MHz. As we show in Section 5, In this section, we describe the this provides substantially WhiteFi design in detail. greater WhiteFi is an implementation of a Wi-Fi the UHF spectrum on which a like protocol on top of the UHF WhiteFi AP or client white spaces that addresses the communicates. key challenges described earlier. A channel is represented as a We design our system on the tuple (F;W), where F is the hardware described in the center previous frequency, andW is the width of section, with one transceiver and the channel. In our current one scanner. Also, we focus on implementation, systems with a single data rate W can be either 5 MHz, 10 MHz, (since rate adaptation itself is an or 20 MHz, but our open problem in white spaces). hardware is generally capable of Our network architecture is using more channel width based on three key components. options. First, WhiteFi incorporates a In contrast, the term UHF novel spectrum assignment channel indicates one of the 30 algorithm segments that is able to handle spatial of the UHF spectrum, which are variation of the spectrum as well each 6 MHz wide. Note that in as spectrum fragmentation. our Second, WhiteFi uses an current hardware efficient, implementation, channels are time-domain signal analysis always centered at technique, called SIFT (Signal a UHF channel’s center Interpretation frequency. Hence, a 5 MHz before Fourier Transform), that WhiteFi channel allows clients to rapidly can fit within a single UHF discover APs transmitting on a channel, a 10 MHz channel spans range of channel widths. Third, 3 UHF channels, and a 20 MHz WhiteFi provides a chirping channel spans 5 UHF channels. protocol that permits a client to 4.1 Spectrum Assignment indicate As shown in Section 2, the a sudden disconnection from the problem of selecting an AP due to a channel conflict appropriate with an incumbent, such as a transmission channel is wireless microphone, without significantly harder in white interfering spaces than in with the primary user. regular Wi-Fi. Because of In the following, we use the term temporal and spatial variability channel to represent a range of in spectrum availability, the AP must pick a In prior work, SampleWidth [15] channel that is free for all its solves the channel width clients. Moreover, fragmentation assignment leads to different-sized spans of problem for a pair of nodes. In available white spaces, so the WhiteFi, we look at the AP also has to decide on the best broader problem of selecting possible both the center frequency and channel width to use as well. channel Always using the widest channel width when there are more than that is available for all clients two nodes. may not be the right solution, Preliminaries: The AP and each since client maintains a spectrum there could be significant map which is a bit-vector fu0; : : background traffic (from other : ; ukg where each ui represents APs) on whether the corresponding UHF some of the underlying UHF channel is currently in use by an channels. Similarly, always incumbent user (that is, a TV picking channel or wireless microphone). the narrowest channel width ui = 1 if the channel is in use by (i.e., a single UHF channel) may an incumbent, and 0 otherwise. be In the United States, there are 30 wasteful if there are wider UHF channels represented channels available. in the spectrum map. Each node These challenges motivate an also maintains an airtime utiliza- adaptive spectrum-assignment 30 algorithm tion vector fA0; : : : ;Akg, that periodically reevaluates the where Ai represents an estimate assignment based on white of the space availability at the AP and airtime utilization on each UHF clients. The algorithm also has channel. Note that for to incumbentoccupied be client-aware since the AP channels, Ai is undefined. The cannot simply rely on its own spectrum map and airtime observation utilization are measured using of the available UHF channels. the secondary scanning radio, This requires clients to using share information with the AP on the SIFT technique described in their own observed UHF channel Section 4.2.1 below. availability. Triggering new channel selection: An AP decides to probe for a new channel when one of two channel selection without client conditions occurs. The first is input.1 an involuntary channel switch The first step is to take the induced by an incumbent (such bitwise OR of the clients’ and as a AP’s wireless microphone) becoming spectrum maps, u?, to determine active anywhere on the AP’s the set of UHF channels current available channel (F;W). Likewise, if a at all of the nodes. client detects an incumbent, it The second step is to consider will disconnect from the AP and each possible channel (F;W) in cause a channel switch to occur. the available white spaces, and This process is described in estimate the aggregate Section 4.3. The second is a bandwidth voluntary that the AP and clients would channel switch, which is receive if selecting that channel. triggered when the AP detects a The performance challenge is that the probed drop on its current channel. The channel (F;W) might overlap AP periodically probes partially for a new potential channel as or completely with channels well, in case another portion of occupied by other APs; we do the not spectrum has opened up since its disallow channel overlaps last probe that could yield between APs. For this reason, higher estimating performance. Of course, an AP aggregate bandwidth based on also performs channel selection airtime utilization measurements when booting up. in Channel probing: To probe for a UHF white spaces is harder than, new potential channel, the AP say, in the 2.4 GHz Wi-Fi band. must have information about the For a given 6 MHz UHF channel spectrum map and airtime c and a node n, we define _n(c) utilization as the expected share of c that observed at each of the clients. node n will receive if c is Clients periodically transmit contained this information to the AP as within (F;W). This is a function part of a control message. When of the busy airtime An bootstrapping, the AP will not c on the have any clients and will perform channel c as measured at n, as that a node will be able to use well as the estimate of the the channel c on each number transmission of other access points operating opportunity. on c, which we denote Bn 1It is possible that the AP c . This selects a channel that is blocked value can be determined for for all instance by using the scanning or some of its potential clients, a radio case that is handled by the and the SIFT technique (Section disconnection 4.2.1). For every node n and any mechanism described in Section UHF channel c, we define 4.3. _n(c) = max(1 _ An 2Since today’s wireless networks c ; are dominated with downlink 1 traffic [11], and it is difficult to Bn measure the number of c + 1 interfering ): (1) clients, we estimate the number The intuition behind this of contending nodes as Bn definition is as follows. At any c , i.e. instant the number of interfering APs. in time, the probability that a Given a range of UHF channels node will be able to transmit on spanned by the probed channel the (F;W), therefore, we define the channel c is at least the residual multichannel airtime metric, or airtime 1 _ An MChamn(F;W), at node n as c . This is a good MChamn(F;W) = estimate for n’s expected share W when the channel is mostly free. 5MHz _ Y c2(F;W) However, even when the medium _n(c) (2) is completely utilized by Since _n(c) represents the neighboring expected share of a UHF channel APs (An c, the c is 1) a node can still expect to product of these shares across get its “fair share” each UHF channel in (F;W) of the airtime when it is gives contending with them.2 the expected share for the entire Therefore, we take channel. We note that simply the maximum of these two values taking as an estimate of the probability the minimum or the maximum across all channels, instead of the product, will be an possible channel (F;W) in the underestimate since the traffic available white spaces, and on a narrower selects channel contends with traffic on the channel that maximizes this an overlapping wider channel metric. In order to also include [15]. the The product is then scaled by the values measured at the clients optimal capacity of the probed (for upstream traffic), the AP channel, W=5MHz. We use a 5 bases MHz channel as our reference its decision on an average value point because it fits into one of all its clients MChamn(F;W) single UHF channel. value, as well as its own value, Example 1: If there is no MChamAP (F;W). Since most background interference or other traffic in today’s wireless APs occupying any portion of networks is on the downlink (F;W), then MChamn(F;W) [11], the AP simply evaluates to the optimal weights its own MCham channel capacity. That is, proportionally higher. In our MChamn(F;W) = 1 for W=5 implementation, MHz, 2 for W=10 MHz and, and the AP selects a channel that 4 for W=20 MHz. maximizes: N_MChamAP + Example 2: Consider a channel c PnMChamn where N is the = (F; 20MHz). Out of the 5 number of clients attached to the UHF channels that are spanned AP. However, notice that other by c, let three have no metrics (such as metrics background including interference, one has 1 AP and fairness conditions) can easily airtime utilization of 0.9, and be implemented instead of one aggregate has 1 AP with airtime utilization throughput. 0.2. MChamn(F; 20MHz) = The AP broadcasts the new 4 _ 0:5 _ 0:8 = 1:6. That is, the channel to its clients, which metric predicts a throughput on upon this hearing the message, switch to channel that is equivalent to the new channel. (If a client roughly 1.6 times of an empty 5 misses MHz the channel switch message, it channel. will revert to the disconnection Channel selection: The AP protocol evaluates MCham(F;W) for each described in Section 4.3.) In the case of a voluntary channel switch, if the measured for periodic beacons from APs, performance of the new channel which are typically transmitted is less every 100 ms. the previous channel, the AP will re-evaluate its channel An alternative approach would selection, be to leverage the SDR in the possibly switching back to the node’s scanner to capture a trace original channel. In our of the signal across a band, and prototype then apply real time OFDM of WhiteFi, the AP measures the decoding, in software, on aggregate throughput achieved successive by channel center frequencies and all clients as a measure of the widths to detect an AP. effectiveness of a channel However, switch. this would incur substantial To prevent frequent changes in computational overhead; the channel or ping-ponging performing across OFDM decoding in software at two channels, we also add 802.11a PHY rates requires hystersis to our system as done multiple in [19]. cores of a well-provisioned In our evaluation in Section 5.4, server-class machine [23]. we show that the MCham metric Moreover, predicts the best possible since the SDR hardware can only channel to a degree of accuracy sample an 8 MHz range of that is spectrum at a time, multiple such sufficient for the above WhiteFi scans would be required. spectrum assignment algorithm 4.2.1 SIFT: Efficient Variable- to Bandwidth Signal Detection achieve near-optimal throughput We propose a hybrid solution in a wide variety of test cases. that uses the SDR to sample a 4.2 AP Discovery given The use of variable channel 8 MHz band, but performs an widths in WhiteFi presents a new efficient time-domain analysis of challenge when performing AP the discovery. TraditionalWi-Fi raw signal to detect the presence clients of an AP and determine its perform access point discovery channel by scanning each channel and width. This approach avoids the listening high overhead of decoding beacon packets in software, while 3There are a total of 30 5MHz making efficient use of the WhiteFi channels, 28 10MHz SDR’s capabilities. channels, Once the AP’s channel (F;W) has and 26 20MHz channels. been identified, the average increases beyond a radio transceiver is tuned to that certain threshold. Similarly, channel and decodes the beacon when the packets in hardware. average falls below the This approach, which we call threshold, the algorithm marks it Signal Interpretation before as an end Fourier Transform, or SIFT, of a packet. In our current works as follows. For a given implementation this threshold is center fixed at frequency F, the USRP board a low value. We are actively samples a bandwidth of 1 MHz working on techniques to around F at 1 MSamples/sec. dynamically Each sample represents 1.024 _s adjust the threshold based on of background noise levels. raw RF signal as an (I;Q) pair; A key question is, how do we the signal amplitude is computed determine the size of this sliding as pI2 + Q2. The USRP delivers window? Since the 802.11 SIFS blocks of 2048 samples at a duration determines the time time to the PC. between the end of a data packet SIFT uses a simple detection and the start of the subsequent algorithm that determines packet acknowledgement, both of which widths based on signal we want to detect accurately, we amplitudes. To accurately detect limit the size of the sliding the beginning window to less than the and end of a packet minimum possible transmission, we compute a SIFS value in our system. As moving prior work has shown [15], SIFS average over a sliding window of values change across different the signal amplitude values. We channel widths and the lowest do not use instantaneous values, SIFS since the signal amplitude might value in our system is for a 20 fall to very low values even in MHz transmission, which is 10_s the middle of the packet or transmission 10 samples. Hence, we choose a (Figure 5). The start of a packet window size of 5 samples. Once transmission is detected when this the algorithm determines the short packet, such as a CTS-to- start and end time of a packet, self, one SIFS interval after the duration sending of the packet is known. From a beacon packet. this we also glean information We expect SIFT to have very few about the interval between a data false positives since it matches packet and its acknowledgement. both the ACK duration and the Both the packet duration and the interval between the packet and SIFS interval are inversely ACK. However, in extremely proportional noisy environments or in the to the channel width. This presence information can be used to of concurrent transmissions, infer the channel width on which SIFT might have false negatives. the packet was transmitted. For It could fail to accurately detect example, by matching the delay all transmissions. We note that between the data and its although this will add delay to acknowledgement the time for discovering APs packet, and the duration of the (Section acknowledgement packet, 4.2.2) the discovery algorithm we can determine the channel will continue to work as long as width of the unicast we can detect even a single transmission. packet. The reason this technique works When SIFT samples an 8 MHz is twofold. First, the band centered at a frequency Fs, acknowledgement it will be able to detect a packet is the smallest MAC layer WhiteFi transmitter whose packet (14 bytes), and channel overlaps cannot be confused with a data with Fs, even though their center transmission. Also, the duration frequencies may not match. of an acknowledgement packet at For example, when SIFT detects the narrowest width of 5 MHz is a 20 MHz WhiteFi channel at Fs, still much smaller than any data the true center frequency Fc of packet sent at 20 MHz. Second, the WhiteFi transmitter can be the SIFS interval is different on anywhere every width and reduces the in the range Fs_10MHz. probability Therefore, the output of the of any false positives. We use a SIFT similar technique to match algorithm is (F _ E;W) where F against non-data packets such as is the center frequency of the beacons. We require APs to send a transmitter, E is an error term, 2: while AP not detected and j _ and W is the transmitter’s 0 do channel 3: c := 0; width (5, 10 or 20 MHz). Since 4: while AP not detected and cur W can be determined exactly by < NC do SIFT, E = _W=2. 5: if cur =2 S then We demonstrate the accuracy and 6: SIFTscan(cur); performance of the SIFT 7: S := S [ fcurg; algorithm 8: if AP not detected in Section 5.1. 9: cur := cur + wj ; 4.2.2 AP Discovery using SIFT 10: end if SIFT enables clients to discover 11: end if APs without tuning into all 12: end while possible 13: j := j _ 1; (F;W) channel combinations. 14: end while Based on the SIFT primitive, Determining AP’s center we devise two AP discovery frequency: algorithms, as described below. Let cur be channel on which Throughout our discussion, NC SIFT detected an AP denotes the number of UHF LetW be the AP’s channel width channels reported by SIFT (30 in the United States) and NW 15: k := 0; represents the number of 16: while AP beacon not decoded channel widths (three in our do implementation). 17: Listen for AP beacons on Linear SIFT-Discovery channel Algorithm (L-SIFT): This 18: [cur _W + k; cur + k]; algorithm 19: k := k + 1; simply scans each of the 30 UHF 20: end while channels in succession, attempt- ing to detect an AP using the 32 SIFT technique at each one. Algorithm 1 J-SIFT Algorithm: Because UHF channels are numbered L-SIFT scans the spectrum from 0; : : : ;NC. lower frequencies to higher w0; : : : ;wNW : channel width frequencies, options (5, 10, 20 MHz) as soon as a transmitter is S: Set of UHF channels already detected, its center frequency scanned. Fc is known: Fc = Fs + E, where SIFT search: Fs is the frequency that SIFT 1: j := NW; c := 0; S := fg; was scanning and E is the and attempt to decode packets to uncertainty returned by SIFT. exactly determine the center The expected frequency. number of iterations until an AP J-SIFT works as presented in is discovered is NC=2, Algorithm 1. It operates in two and the worst case is NC phases. First, it scans the UHF (compared to roughly NC _ spectrum in a staggered fashion, NW=2 and using SIFT to detect the NC _ NW, respectively, by the presence of a WhiteFi non-SIFT baseline). transmitter. In Jump SIFT-Discovery Algorithm the second phase, it identifies (J-SIFT): We can improve the transmitter’s center upon the expected scan time of frequency L-SIFT by performing a Fc. While the worst-case staggered discovery time of J-SIFT is the search of the spectrum. Since same SIFT is able to detect a WhiteFi as for L-SIFT (NC), the expected transmitter by scanning discovery time can be shown to anywhere within its band, we can be 1 improve NW performance by first scanning (NC + 2NW_1 + (NW _ 1)=2). for 20 MHz WhiteFi channels We elide the derivation (skipping due to lack of space. over 5 UHF channels at a time), In WhiteFi, we expect the then 10 MHz channels (skipping average number of scans over 3 UHF channels at a time as required for well as any UHF channels L-SIFT and J-SIFT to be NC=2 previously scanned), and finally and (NC + 4 + 1)=4, respec- for 5 MHz channels (in the 4Generally, if more widths are remaining available, we would do the unscanned UHF channels).4 staggered One disadvantage to J-SIFT is search starting from the widest that the WhiteFi transmitter’s channel width. center tively. That is, we expect J-SIFT frequency is not immediately to outperform L-SIFT when NC known when it is detected. is greater than about 10 UHF Therefore, channels. For narrower white it is necessary to tune the radio spaces, to each of Fs _ E channels L-SIFT is more efficient. Our measurements in Section 5.2 validate these theoretical findings. channel, in a manner similar to 4.3 Handling Disconnections that being considered by the A key challenge in WhiteFi is 802.22 dealing with the sudden working group [1]. To avoid appearance disrupting communications with of a primary user (such as a stillconnected wireless microphone) on a clients, chirp detection is channel performed using SIFT on the that an AP-client pair is using secondary radio, in the for communications. Note that background. Once a chirp has either been detected, the AP or the client might detect the AP can switch its main radio the primary user, requiring that to the backup channel and a decode the contents of the chirp channel must be vacated. We call packet. As a further this a disconnection. optimization, Our approach is as follows. The we can encode some amount of AP maintains a separate 5 MHz information in the time domain, backup channel that is advertised such as the client’s SSID, for as part of its beacon packets on example by setting the length its main channel. If the AP or a of the chirp packet. (In effect, client detects a primary user on this uses SIFT to implement a the main channel, the node lowbitrate switches to the backup channel OOK-modulated channel.) This and approach avoids switching transmits a series of chirps that the main radio to the backup contain information on the white channel for clients associated spaces available at that node. with a If a client senses that a different AP. disconnection has occurred (e.g., Once a node begins chirping, because after a threshold time interval no data packets have been Tc, received in a given interval), it the collective white space switches availability advertised by each to the backup channel and listens node on for chirps, as well as the backup channel is used to transmitting reassign spectrum to the AP and its own. Access points clients in that SSID, as periodically scan for chirps on described in Section 4.1. Nodes the backup in the SSID switch to the new channel and time taken to switch across resume communication. channels, which is known to be a There is an additional case we few milliseconds. We realize must consider, namely, when a that this overhead can be node (either the AP or the avoided by client) determines that the adding security features to SIFT, previouslyselected so that only an authorized client backup channel is occupied by will cause the AP to switch its another primary user. In main radio. We are actively this case, an arbitrary available investigating channel is selected as a this approach. secondary EVALUATING WhiteFi backup and used for chirping. In this section, we evaluate Therefore, in addition to WhiteFi in detail. Using a scanning combination the backup channel for chirps, of simulations and experiments the AP periodically scans all on our prototype channels implementation, in an attempt to reconnect with we show: “lost” nodes. Note that chirps 33 contend for the channel using 0.125 M 0.25 M 0.5 M 0.75 M 1 CSMA, just like data packets; as M a 5 MHz 0.99 0.98 0.98 0.98 0.97 result, it is unproblematic for a 10 MHz 0.99 0.99 0.99 1.00 0.99 backup channel to overlap with 20 MHz 0.99 1.00 0.99 1.00 0.99 another Table 1: SIFT’s packet detection AP’s main channel. rate, i.e., the median number An attacker can potentially of packets detected by SIFT hijack our system by sending divided by the total sent by the fake wireless card. The values are chirps. However, the impact of measured across different widths this attack is limited. Once the when varying the traffic AP’s main radio switches to the intensity from 125 Kbps to 1 backup channel, it will process Mbps. the _ In Section 5.1, we demonstrate chirp packet only if it is encoded the accuracy of SIFT in with the network’s security key detecting (similar to Wi-Fi). Therefore, packets across different channel the overhead of this attack is the widths and when there is extra high signal attenuation. _ In Section 5.2, we show the the second device to count the effectiveness of WhiteFi’s AP number of packets detected by discovery SIFT. algorithms. Our experiments We repeated this experiment for show that J-SIFT improves 5, 10 and 20 MHz channel the time to discover APs by more widths, than 75% compared to non- and for each width, we varied SIFT based techniques. the traffic intensity. All the _ We demonstrate the reported correctness of WhiteFi’s numbers are over 10 runs. In protocol in handing every run, we sent 110 packets disconnections in Section 5.3. of size _ Finally, in Section 5.4, we 1000 bytes each. show that WhiteFi’s spectrum Accuracy across Channel assignment Widths: Table 1 shows the algorithm adapts quickly to fraction changes in network conditions. of the number of packets Using extensive simulations in detected by SIFT when varying QualNet, we also show the rate that WhiteFi’s performance is at which packets were sent close to optimal under various across different channel widths. conditions. As we 5.1 Accuracy of SIFT see in the table, SIFT detects We evaluate the packet detection nearly all the packets for every accuracy of SIFT. We first channel describe width. The worst case loss our methodology and then show across all widths and rates was its accuracy when varying 2%. two parameters: channel width An interesting observation is and signal attenuation. that the detection rate for 5 MHz Methodology: We used the was following set up for our slightly worse than the detection experiments. rate at other channel widths. We started an iperf session from This one KNOWS device, and was a result of the way 5 MHz measured the number of packets packets are transmitted by our that were received at a second hardware device in the time domain. As we see in using a packet sniffer. Figure 5, the initial portion Simultaneously, we used the scanner of of a packet at 5 MHz channel of SIFT at low signal strengths width is sent at a lower by connecting two KNOWS amplitude devices than the rest of the packet. Figure 6: Accuracy of air time Consequently, our algorithm utilization measurement using sometimes SIFT. Error bars were within 2% fails to accurately match the of the mean. length of the detected packet and Figure 7: Discovery of APs with the distance. SIFT is able to transmitted one. However, SIFT discover always correctly detects the APs until as long as theWi-Fi channel card can decode packets. width of the transmitted packet, Figure 8: Reduction in discovery even when it mis-estimates the times using L-SIFT and JSIFT packet length. when compared to the non-SIFT In addition to detecting the based baseline. appropriate width, we also use through a tunable RF attenuator, SIFT and performing the same to measure the airtime experiment utilization for WhiteFi’s as above. Figure 7 shows the spectrum assignment percentage of packets that were algorithm. We show that SIFT detected by SIFT and the packet performs as expected in Figure sniffer upon varying the 6. The total time occupied by the attenuation. packets doubles on halving At low attenuation, both SIFT the channel width. This stems and the packet sniffer perform from the observation in [15] that very well. However, SIFT halving the channel width also outperforms the packet sniffer, halves the effective transmission as it is rate. Since we send the same even able to detect corrupted number of packets at a given packets. At higher attenuation, width, SIFT the total airtime is constant, continues to detect more packets even when we change the rate of than the sniffer until 96 dB injected attenuation. packets. Since SIFT applies a threshold Accuracy with Signal to the amplitude of the incoming Attenuation: We evaluated the signal, it performs poorly accuracy beyond a certain attenuation. In our experimental setup, this occurs realistic settings, i.e., in at 96 dB. Beyond 96 dB we see a metropolitan, suburban and rural very sharp drop in the settings. percentage of successfully Methodology: We set up two detected packets. KNOWS devices as before, and In contrast, the reception ratio configured one as an AP and the of the packet sniffer falls off other as a client. In the more beginning smoothly, and performs better of the experiment, the AP started than SIFT beyond 98 dB to beacon on a randomly chosen attenuation. UHF channel and channel width. However, at this attenuation the We then measured the time for capture ratio is extremely low the at around 35%. Most client to discover the AP using applications, including TCP, will L-SIFT, J-SIFT and the non- perform SIFT poorly at such high loss rates. baseline. Depending on the Hence, we conclude that in most scenario, we artificially commonly occuring scenarios, specified the SIFT detects almost all packets spectrum at the AP and the that client. The AP did not beacon on are successfully received by a any transceiver radio. of the occupied channels, and 5.2 Time to Discover APs the client did not scan these We now evaluate the channels performance of the L-SIFT and for an AP. J-SIFT discovery Contiguous Channels: In this algorithms in discovering APs. experiment, we set the spectrum We compare them to a non- map to have only one available SIFT baseline that would have to fragment. We varied the number scan every possible center fre- of 34 UHF channels in the fragment Figure 9: Time to discover one from 1 to 30, since 30 is the AP at various locations. total quency and width to discover the number UHF channels that are APs. In this section, we consider available to portable devices. In two scenarios. First, we show Figure 8, we plot the total time the benefit of our algorithms as taken by L-SIFT and J-SIFT to a discover the AP as a fraction of function of contiguous width. the total time taken by the non- Then, we evaluate the benefits in SIFT baseline. When there is only one fewer contiguous channels, J- available UHF channel, the time SIFT is 34% faster than the taken by all the algorithms is the baseline. same. However, when we In rural areas (more contiguous increase channels), we see that J-SIFT the width of the available can fragment of spectrum, L-SIFT discover APs in less than one- and JSIFT third the time taken by the perform much better than the baseline baseline. As expected, L-SIFT algorithm. outperforms J-SIFT initially (for 5.3 Handling Disconnections narrow white-spaces) since it We now quantify the time taken does by WhiteFi to reconnect not require the “endgame” of disconnected trying to find the proper placing clients. We setup a client and an of AP and started a data the AP channel. On the other transfer between them. Then we hand, as exactly predicted by our switched on a wireless analysis in Section 4.2, J-SIFT microphone becomes more efficient for white near the client. This causes the spaces spanning more than 10 client to disconnect, and it UHF channels (60 MHz). starts chirping on the backup Realistic Settings: We also channel. In our experimental measured the time to discover an setup, AP the AP switched to the backup in metropolitan, suburban and channel once every 3 seconds, rural areas in the US. We used and the picks up the chirp in at most 3 methodology described for seconds. Immediately, the AP Figure 2 to obtain the spectrum uses maps the spectrum assignment post-DTV transition. We algorithm to determine the best randomly placed the AP on an available available channel to operate on, and the channel and width and repeated system is operational again after the experiment 10 times for a every lag of at most 4 seconds. locale. As shown in Figure 9, in 5.4 Spectrum Assignment metro areas, where there are We now evaluate WhiteFi’s spectrum assignment algorithm. For a detailed understanding of our mechanism in QualNet such that algorithm, and to evaluate it a node spanning multiple UHF under channels will transmit a packet varied settings, we decided to only if no carrier is sensed on use the QualNet simulator [3]. any of The those channels. We also need to use the simulator arose modified QualNet to support for two reasons. First, we were fragmented constrained by having a limited spectrum. Every node reads its number of prototype devices, initial spectrum map from a and configuration second, we did not have an FCC file. license to transmit packets in the 5.4.1 Simulation Results TV bands. Therefore, we We study the performance of evaluated our system (spectrum WhiteFi’s spectrum assignment assignment, algorithm under various settings. discovery and disconnection First, we microbenchmark the protocols) in a limited setting MCham metric, and show that it – on a testbed spanning one floor is a good estimate of the in our building, and a maximum expected transmit power of 1 mW. throughput on a channel. Then, Modifications to QualNet: We using large scale experiments, modified QualNet to support we variable channel widths by show that WhiteFi performs appropriately scaling the OFDM reasonably well under: (i) symbol varying period, and various MAC layer amounts of background traffic on parameters that were described the channels, (ii) large amounts in [15]. We also adjusted the of spatial variation in spectrum channel noise levels based on availability, and (iii) when there the is a channel width. Furthermore, at lot of churn in background every node, we explicitly drop traffic. In all these experiments, packets that were sent at a WhiteFi different channel width. To performs nearly as well as an ensure that a optimal algorithm. In the node appropriately contends with process, we packets that are sent on also show the need for WhiteFi overlapping to adapt both the center channels of different widths, we frequency modified the carrier sensing and the channel width. Microbenchmark Setup: To the narrower 10 MHz channel verify that MCham correctly surpasses the wider channel predicts thereafter. the channel that will lead to the Similarly, at about 24 ms inter- best throughput, we simulate packet delay, 5 MHz starts a spectrum fragment of 5 achieving the highest adjacent UHF channels (26-30), throughput, which is accurately each having one background predicted by client/AP-pair. There is one AP the MCham metric. We can with conclude that the MCham metric one associated client, yields transmitting a link-saturating a reasonably accurate prediction UDP flow. We of which channel width will vary the traffic intensity of the result background nodes (from 0 to 50 in the highest throughput given a ms certain level of background inter-packet delay) and measure traffic. the effect on the MCham metric Setup of large-scale simulations: and client throughput when To better understand the transmitting on the 5, 10, and 20 behavior MHz of WhiteFi in large-scale channels centered at channel 28. settings, the next three Accuracy of the MCham Metric: simulations The results in Figure 10 show consider the following basic that the MCham metric setup. We place one AP in the accurately predicts which middle channel achieves of an area, and randomly the highest throughput for any distribute clients as well as given background intensity. For background example, 35 selecting a 20 MHz channel Figure 10: MCham value and achieves best throughput until resulting throughput of a 5, 10, a background traffic intensity of and 20 MHz channel as a roughly 18 ms inter-packet function of background traffic delay. intensity. Similarly, the MCham metric The MCham metric accurately predicts that roughly at this predicts which channel level of achieves highest throughput. background traffic, 10 MHz and AP/client-pairs within 20 MHz become equally good, transmission range of this AP and (background clients are always deployed MHz are the algorithms that within transmission range of statically pick the best possible their respective 10 background AP). The AP and and 20 MHz channel, clients are backlogged and respectively. Finally, OPT is an transmit UDP flows (up- and ideal, omniscient downstream). Background nodes algorithm that for every transmit constant-bit-rate (CBR) experiment run picks the channel traffic at a pre-specified with maximum throughput. The intensity. goal of the WhiteFi spectrum All experiments are repeated 5 assignment times with different random algorithm is to approach OPT as placements closely as possible. of nodes, and results are Impact of Background Traffic: averaged. Figure 11 shows how WhiteFi An underlying spectrum map is reacts to varying degrees of shared across all clients (except background traffic. Specifically, in the experiment in which we there focus on the impact of spatial are X background AP/client- variation). pairs in the system, each being Specifically, the spectrum map randomly is taken from our real assigned to one of the free UHF measurements channels, and each sending in Section 2. There are 17 free at a packet interval delay of 30 UHF channels, and the ms. widest contiguous white space is The figure shows that WhiteFi 36 MHz, i.e., there are multiple achieves close to optimal possibilities of selecting even 20 performance MHz wide channels for the AP. for varying degree of In all experiments, we measure background traffic. With little or the per-client throughput of no clients/APs. We consider the background traffic, WhiteFi following baseline algorithms performs as well as picking the for widest comparison with WhiteFi. OPT 5 available channel (OPT 20 MHz denotes the throughput MHz), which is optimal. As the achieved when statically picking traf- the best (across all fic increases, the throughput nonincumbent) achieved by OPT 20 MHz drops, UHF channels. Similarly, OPT and 10 MHz and OPT 20 OPT 10 MHz becomes better (at the AP start with a common about 10 background spectrum map. Then, for each AP/clientpairs). client Even at this point WhiteFi (and AP) and for each UHF performs near-optimally, which channel i, we randomly flip the shows that WhiteFi adaptively entry switches to narrower channels as ui with probability P. In the needed. In fact, our evaluation experiment, we vary P from 0 shows that WhiteFi is always (no within spatial variation) to 0.14 (large 14% of the optimal value spatial variation). throughput OPT. It can be seen in the figure, Figure 11: Impact of background spatial variation reduces traffic on throughput. achievable Figure 12: Impact of spatial aggregate throughput. Because variation on throughput. the AP needs to select a channel An important observation is that that is free at all clients, no due to fragmentation and contiguous free spectrum parts background remain traffic, there is no single best available for P > 0:1, and hence, center frequency and channel the aggregate throughput reduces width that should be used in to the throughput of a single UHF white spaces. WhiteFi is UHF channel (5 MHz). For low capable spatial of adjusting to the appropriate variation, the throughput is width and selects a near-optimal much higher when selecting a 20 channel. MHz Impact of Spatial Variation: wide (e.g. at P = 0:01) or a 10 Figure 12 shows the impact of MHz channel (e.g. at P = 0:05). spatial variation on per-client Generally, the figure highlights throughput. In this experiment, the need for adaptive channel there width are 10 clients connected the AP, in UHF white spaces: no single and one background channel width (OPT 20 MHz, client/APpair OPT per UHF channel, transmitting at 10 MHz, OPT 5 MHz) achieves CBR with 30 ms inter-packet close-to-optimal throughput in delay. Spatial variation is all modeled as follows. Each client cases. On the other hand, and WhiteFi is near-optimal in all cases. Impact of Churn: Finally, we Figure 13: Impact of churn on want to understand the impact of throughput. churn (in terms of background Figure 14: Experimental traffic) on the throughput validation of WhiteFi’s spectrum achieved assignment by WhiteFi and the various algorithm on a testbed with baseline algorithms. There are a variable background total traffic. Top figure shows the of 34 background AP/client- MCham metric for each of the pairs, two per free UHF channel. three channel widths. Bottom In figure shows the throughput order to model churn, we model (averaged background nodes using a simple over 5 sec windows) for WhiteFi discrete Markov chain with two and OPT. states (A=active, P=passive). A Figure 13 shows that WhiteFi background node in the active performs near-optimally for state transmits CBR traffic with varying 60 degree of churn. For low churn ms inter-packet delay. A node in and little background traffic, the passive state does not WhiteFi selects the widest transmit. channel. For high churn (e.g., We simulate this setting for state duration various state transition 45 seconds and passive probabilities, probability 1=3), always picking selecting them to cover the the entire range of (1) likelihood of widest channel (OPT 20 MHz) being in either state and (2) becomes the worst performing average state duration (see x- algorithm. axis in Instead, WhiteFi is better than Figure 13). The extreme cases any static channel width are (i) all nodes are always in choice. In fact, WhiteFi even state outperforms OPT. In this P, (ii) nodes are in each state experiment, with equal likelihood and they this is possible because OPT is remain the optimal static channel in their current state for an selection average of 30 seconds, and (iii) throughout the entire execution all nodes of the simulation. Instead, are always in state A. WhiteFi 36 is adaptive and can adjust to the Initially, when there is no current values of background background traffic, the AP and traf- client fic, changing its channel operate on the 20 MHz spectrum accordingly. chunk between channels 26 and 5.4.2 Results from our Prototype 30. Then at time 50 seconds, we To demonstrate the adaptability introduce background traffic on of WhiteFi’s spectrum channels 26 through 29. assignment Correspondingly, the value of algorithm, we set up an the MCham experiment with an AP and a metric for the 20 MHz fragment client drops sharply, and the AP and its in our building, which is clients move to the 10 MHz Building 5 in Figure 1. The spectrum fragment. As shown in spectrum map the of our building has the following figure, this is also the fragment free UHF channels: 26 to 30, 33 that has the best throughput. to 35, 39 and 48. Therefore, we Then have fragments of size 20 MHz, at time 100 seconds, we 10 introduce background traffic on MHz and two channels of 5 MHz channels to form a network. 33 and 34, and as before the Every client and AP using value of the 10 MHz channel’s WhiteFi spends 1 second on MCham every metric drops, and the system UHF channel to determine the switches to channel 39 (any 5 airtime utilization using SIFT, MHz as chunk could have been chosen). described in Section 5.1. All Then at times 150 and 200 nodes feed their airtime to the seconds, AP, we remove the background which computes the MCham interference from channels 33 metric and decides on the and 34, and from channels 26 channel to through 29, respectively. use for the network. We present Correspondingly, the throughput of our system WhiteFi switches to the fragment with with the best MCham time, and the corresponding value, i.e. to the 10 MHz MCham value on the different fragment at 150 seconds, and to spectrum the 20 chunks in Figure 14. MHz fragment at 200 seconds. chunk of spectrum. An We conclude from the above alternative technique might use a experiments PHY that WhiteFi adaptively operates layer that operates over non- on the best part of the contiguous spectrum chunks. The spectrum. AP DISCUSSION AND can then operate over the entire FUTUREWORK bandwidth, decoding signals White space networking provides from a unique opportunity for the different clients who may be cleanslate using different OFDM network design, owing to the subcarriers. lack of existing standards. Our For AP-to-client decision to build the WhiteFi communications, the PHY layer prototype with a Wi-Fi card was could either motivated by several factors. suppress or send a null signal on Wi-Fi is a mature, well- the subcarrier that the primary understood user is using [21]. In theory this technology that is inexpensive is a reasonable idea but it poses and easily available. Several two practical problems. First, wireless leakage from adjacent card vendors we have spoken subcarriers with are considering pushing causes interference to the some version of Wi-Fi to the primary user. To avoid this IEEE standards body for white interference, space we would require a highly networking. Additionally, Wi-Fi accurate bandpass filter of enabled us to build a prototype appropriate quickly and focus on some of the bandwidth but to the best of our higher layer issues that are knowledge researchers are still somewhat working on developing such agnostic to the existing physical sharp bandpass filters. Second, and MAC protocols. However, and we do realize that alternative more importantly, sending data designs are possible and might over different subcarriers to an be used in future networks. We AP discuss a few of these below. is difficult to implement for WhiteFi leverages the technique uplink traffic. We are not aware described in [15], which requires of any the AP and its clients to operate over the same contiguous system that can decode packets pact slot scheduling and lead to sent simultaneously from poor performance. Furthermore, multiple in UHF white spaces the clients clients over non-overlapping and AP may be over a mile away, subcarriers. This is an active further aggravating the research scheduling problem. Additional area and we are investigating the research is practicality of such a system. needed to understand these Another issue is our choice of issues and is out of scope for CSMA/CA, the medium access this paper. control (MAC) protocol for Wi- Our initial results show that Fi, in WhiteFi. The research CSMA/CA is a reasonable choice literature for has several interesting proposals white space networking. for MAC protocols, which Prior work has proposed the use can be broadly categorized under of control channels to reserve Listen Before Transmit (LBT) bandwidth and spectrum [12, and Time Division Multiple 24]. While there are advantages Access (TDMA). Observing what to is a control channel design, we happening in the ISM bands we believe that control channels can made the decision that WhiteFi be must be able to co-exist with compromised, thus bringing other unlicensed devices. The down the network. Also, control success channel of LBT protocols (e.g., Wi-Fi) in based solutions are prone to the the ISM bands made it a natural range-mismatch problem [24]. choice for white space We overcome these problems by networking. We also believe that not using a dedicated control an alternative channel. WhiteFi uses a backup TDMA like MAC (e.g., channel in the white spaces Bluetooth) will not perform well (instead in of 900 MHz spectrum as white spaces without significant proposed by CMAC [24]) thereby modifications. Local avoiding the range mismatch interference problem. Also, WhiteFi does not from wireless microphones use around the client or the AP a static control channel. It would im- dynamically adapts the backup 37 channel to operate on spectrum that is broadband access to rural areas not occupied by a primary user. and neighborhoods. In contrast, WhiteFi considers a usage model RELATEDWORK similar to Wi-Fi, with one AP providing coverage to several Prior work has mostly focussed possibly mobile users. Despite on the problem of the opportunistically difference in the scenarios, the forming a single link over UHF techniques developed by white spaces [8, 12]. This WhiteFi, involves accurate sensing of the for disconnection, discovery and spectrum [14,17,18], reliable spectrum assignment, are also identification applicable of incumbents, and radio agility in WRANs. For example, the on detecting a primary. 802.22 draft includes support However, to the best of our for variable widths, although it knowledge, no prior work has does not specify how to use it. studied A recent technology that enables the problems of forming a Wi-Fi unlicensed devices to co-exist like network over white spaces. with licensed users is SWIFT WhiteFi builds upon our prior [21]. SWIFT pokes the primary work on KNOWS [24], which user uses to learn about its presence. a similar hardware platform and Unfortunately, this technology proposes a control channel based cannot MAC protocol for ad hoc be used over white spaces networks over white spaces. because the FCC does not allow WhiteFi “testing” looks at the problem of forming the presence of an incumbent by an AP based network while “poking” at it with a reusing transmission. the Wi-Fi MAC and without Also, the incumbents of UHF using a control channel. white spaces do not back off. A complementary effort to WhiteFi is the IEEE 802.22 [1] CONCLUSIONS working In this paper, we have presented group’s proposal for WRANs the design and implementation (Wireless Regional Area of WhiteFi, the first white space Networks) Wi-Fi like wireless network. over UHF white spaces. It is We moved beyond the current intended to provide wireless state-of-art that considers a single link to building a real network with multiple links. In building WhiteFi we identified and described several unique challenges in operating a white space network and showed with extensive experiments how white space networks differ from ISM band Wi-Fi netwoks. WhiteFi contributes a new spectrum assignment algorithm that solves the dual challenges of spatial variation of available spectrum and spectrum fragmentation. We further described a new mechanism that quickly discovers APs operating anywhere in the 180 MHz white space, using any arbitrary channel which quickly analyzes packets width. in the time domain, allowing fast We also described a new AP discovery and managing technique for handling disconnections due to temporal disconnections variations. where clients signal to the AP We demonstrated WhiteFi in the without interfering with ongoing context of our custom built wireless microphone prototype UHF hardware and transmissions. Underlying our QualNet simulations. As part of solutions isa new applicationof a ongoing signal recognition technique work, we are deploying WhiteFi called SIFT, over a campus wide white space network