IEEE TRANSACTIONS ON MOBILE COMPUTING,
VOL. 12,
NO. 11,
NOVEMBER 2013
2105
SpiderRadio: A Cognitive Radio Implementation Using IEEE 802.11 Components Kai Hong, Student Member, IEEE, Shamik Sengupta, Member, IEEE, and R. Chandramouli, Senior Member, IEEE Abstract—In this paper, we present SpiderRadio, a software defined cognitive radio (CR) prototype for dynamic spectrum access (DSA) networking. The medium access control (MAC) layer of SpiderRadio is implemented in software on top of commodity IEEE 802.11a/b/g hardware. However, the proposed architecture and implementation are applicable to other spectrum bands as well. We also present a dynamic spectrum sensing methodology for primary incumbent detection. The proposed method is based on observing the PHY errors, received signal strength and statistical model building. For coordination among radio nodes, synchronization and fast channel switching, we present new communication protocols, design extended management frame structure and modify the hardware abstraction layer. Several fundamental tradeoffs (e.g., complexity versus network performance) to be considered during a dynamic spectrum access radio network prototype implementation are also discussed in detail. To demonstrate the practical capabilities of the proposed SpiderRadio prototype, we also present various testbed experimental measurement results. Index Terms—Cognitive radio, dynamic spectrum access, prototype, wireless network
Ç 1
INTRODUCTION
W
ITH spectrum usage being both space and time dependent, it is shown through various experimental studies that a fixed, static allocation often leads to over utilization in some bands and under utilization in others [1]. To eliminate the risk of such artificial spectrum scarcity, spectrum owners in most countries passed new amendments to define provisions for dynamic spectrum access (DSA). (e.g., Federal Communications Commission (FCC) in the US of America passed an important resolution entitled NPRM 04-186 [2]). In DSA, provisions are made to allow unlicensed (secondary) users to operate in the unused or underutilized licensed bands (primary bands) opportunistically. However, such an opportunistic spectrum sharing must not interfere with the licensed service operations (primary incumbents) who are the primary owners of these bands. The success of this policy depends on the ability of secondary users to dynamically identify and access unused spectrum bands, detect the return of primary users, and switch to a different band promptly upon sensing the primary user. The cognitive radio (CR) paradigm [3] is anticipated to make dynamic spectrum access a reality. The basic operating principle of a cognitive radio relies on a radio being able to sense a band, detect if it is used, and, if not used by a primary user, decide whether to initiate communication in that band. The physical layer of a cognitive radio includes sensing (scanning the frequency
. K. Hong and R. Chandramouli are with the Department of Electrical and Computer Engineering, Stevens Institute of Technology, Hoboken, NJ 07031. E-mail: {khong, mouli}@stevens.edu. . S. Sengupta is with the Department of Computer Science and Engineering, University of Nevada, Reno, 1664 N. Virginia Street, Reno, NV 89557. E-mail:
[email protected]. Manuscript received 28 June 2011; revised 1 Aug. 2012; accepted 16 Aug. 2012; published online 12 Sept. 2012. For information on obtaining reprints of this article, please send e-mail to:
[email protected], and reference IEEECS Log Number TMC-2011-06-0350. Digital Object Identifier no. 10.1109/TMC.2012.192. 1536-1233/13/$31.00 ß 2013 IEEE
spectrum and processing the sensed data), cognition (detecting the presence of other users), and adaptation (optimizing parameters such as power, band, and modulation). The medium access layer utilizes the sensing decision while executing the communication protocol. One of the important regulatory aspects is that a cognitive radio enabled secondary device communicating in primary band(s) upon detecting primary incumbents in that band(s) must automatically switch to another channel or mode within a certain time threshold synchronizing with the communicating nodes. More details and regulatory aspects can be found in [4]. Unlike the existing radio devices, CR with dynamic spectrum access capability faces several additional challenges. Apart from the typical functionality of transmitting/receiving data to ensure quality of service in the wireless environment, the additional challenges of cognitive radios in the DSA environment can be divided into three major categories: Primary/incumbent sensing. Preswitching synchronization with communicating node upon successful incumbent detection. 3. Fast switching for successful rendezvous. A great deal of research focus continues to be on spectrum sensing (e.g., [5], [6], [7], [8] and references therein), most of which falls into three categories: 1) matched filter detection, 2) cyclostationary feature detection, and 3) energy detection. While matched filter detection and feature detection provide better accuracy than energy detection, the fact remains that both the mechanisms incur tremendous computation complexity with additional energy consumption and perform much slower than the anticipated requirement [9], [10]. Energy detection although faster, cannot distinguish between secondary transmission and primary transmission. As a result, for energy detection to be successful, it is necessary to forcefully quiet down all 1. 2.
Published by the IEEE CS, CASS, ComSoc, IES, & SPS
2106
IEEE TRANSACTIONS ON MOBILE COMPUTING,
the secondary transmissions periodically which degrades the secondary network performance. In addition to the primary incumbent sensing, when a cognitive radio device moves from one spectrum band to another, it must restart the hardware to reset and adapt to the particular transmission or reception parameters in the new spectrum band. This process of hardware resetting configures the cognitive radio medium access control (MAC) accordingly, which introduces delay during channel switching. Moreover, when two communicating cognitive radio nodes switch to a new spectrum band, they must successfully synchronize with each other to resume communication. Note that, in DSA, there may not be a fixed predefined channel for the cognitive radio enabled secondary devices to move to. Thus, the challenges for the cognitive radio devices are exchanging accurate synchronization messages (available channel information) with each other, dynamic channel switching with minimum switching delay, and resynchronizing with communicating peers as fast as possible in the new channel. During the entire process of frequency switching and resynchronization, unless some remedial actions are taken, data from upper layers may be lost which would again adversely affect the data throughput performance. These practical implementation challenges are critical to demonstrate an efficient and fully functional cognitive radio network, but have not been explored comprehensively in the literature. In this paper, we address some of the implementation challenges and propose solutions. The proposed solutions are implemented in a cognitive radio prototype (called, SpiderRadio) based on off-the-shelf IEEE 802.11a/b/g wireless cards supported by Atheros chipsets. With a programmable software abstraction layer, the CR prototype can adapt the transmission/reception parameters automatically to operate at any unused frequency in the allowable spectrum bands unlike legacy radios which can only operate statically at any one frequency channel. For our testbed setup, the primary user bands were emulated using the 900 MHz, 2.4 GHz, and 5.1 GHz Wi-Fi spectrum bands. The primary user communication was emulated in two different ways: cordless phones communicating with each other using the intercom feature and Agilent signal generator operating in the Wi-fi bands. The SpiderRadio node acts as the secondary user device for the experiments (i.e., upon detecting primary device(s) in a particular band, the CR prototype must vacate that band and move to another unused band). This helps us in developing and testing our prototype without the complexity of buying and managing several licensed spectrum bands. Note that the proposed techniques can be implemented in other bands as well. The key contributions of this paper are as follows: .
A new incumbent sensing algorithm that exploits some PHY layer information from the wireless card (that comes for free) and a synchronization protocol for dynamic spectrum access. The proposed incumbent sensing is adaptive in nature and driven by three parameters: observed PHY errors, received energy strengths, and n-moving window sensing. Detailed discussions about each of these terms are presented later in the paper. The most important aspect in our cognitive radio network framework is, we no longer assume the notion of periodic sensing
VOL. 12,
NO. 11,
NOVEMBER 2013
by forcefully making all the radio nodes quiet because of the degraded throughput outcome. Rather, the CR nodes are allowed to carry out communications even at the time of sensing. The sensing algorithm and the synchronization protocol are implemented in an IEEE 802.11a/b/g network for extensive experiments. . Modifications to the IEEE 802.11 media access control layer and hardware abstraction layer (HAL). The HAL is modified to build two special hardware queues (sync queue and data buffer queue) for synchronization and data buffering. With these hardware queues, dynamic channel switching in the PHY/MAC layer is kept hidden from the upper layers, not affecting the upper layer functionalities. . Design of synchronization management frames which are special purpose frames to help the communicating CR nodes synchronize by exchanging destination channel information (candidate channel) at the time of switching. . Addressing the issue of hidden incumbent (defined later) avoidance by CR nodes. We propose and implement a bitmap channel vector for embedding candidate channel(s) information in the synchronization messages. . Extensive testbed experiments (indoor and outdoor) to demonstrate that the proposed techniques result in fast synchronization and switching. High effective TCP throughput (average 25 Mbps) is achieved even under frequent switching. The remainder of the paper is organized as follows: We discuss the existing literature that relates to this research in Section 2. In Section 3, CR prototype implementation challenges are discussed. The enhanced system architecture and design issues are discussed in Section 4. Adaptive spectrum sensing/detection is discussed in Section 5. Dynamic channel switching and synchronization mechanism are presented in Section 6. In Section 7, we present the testbed setup and experimental results along with detailed discussions. Conclusions are drawn in the last section.
2
RELATED WORK
There are several ongoing research that deal with cognitive radio (e.g., [11], [12], [13], [14], [15], [16], [17], [18] and the references therein). The IEEE 802.22 working group is also investigating a new standard, wireless regional area networking (WRAN) [9], [19] based on cognitive radios. Recent surveys on cognitive radio networks can be found in [20] and [21]. Complementary efforts to WRAN are also recently being investigated in [22] and [23], where the researchers are studying the design of Wi-Fi like network over white spaces. The study mostly concentrates on forming an AP-based network while reusing the Wi-Fi MAC and propose a new adaptive spectrum assignment algorithm to handle spectrum variation and fragmentation. White-space networking maintains a separate 5 MHz backup channel for control information exchange. In contrast to their approach, our proposed mechanism is not dependent on common control channel. This reduces the complexity of channel switching process. The researchers investigate opportunistic channel scavenging mechanisms
HONG ET AL.: SPIDERRADIO: A COGNITIVE RADIO IMPLEMENTATION USING IEEE 802.11 COMPONENTS
for secondary CRs while focusing on the impact of such access mechanisms on primary users’ performance in [24] and [25]. In [26], the channel allocation problem is investigated coupled with power allocation using a twophase mixed distributed/centralized control algorithms. While their first phase concentrate on a distributed power allocation mechanism, the second phase delves into the channel allocation problem from a centralized standpoint. In [27], a spectrum decision framework is proposed to determine a set of spectrum bands by considering the application requirements as well as the dynamic nature of heterogeneous spectrum bands. However, these research mostly investigate the multichannel MAC protocols for dynamic spectrum access without providing much attention to the spectrum sensing and incumbent detection. The researchers investigate the impact of erroneous detection of spectrum holes and missing communication opportunities on the capacity of the secondary channel in [28]. The present literature on spectrum sensing is still evolving. Among the recently proposed incumbent detection techniques, cyclostationary feature detection is a method that exploits the cyclostationary feature of a signal [29]. However, this method has significant computational overhead, consistent bandwidth loss and high time-consumption [10]. Moreover, some knowledge of primary users’ signal is necessary in this approach which may not be available in the first place. Spectral redundancy that is a huge factor in the computation might be corrupted by strong signals in adjacent bands. Alternative to cyclostationary detection, energy detection is another spectrum sensing technique that has low computational overhead and implementation complexity [30]. Energy detectors do not need any knowledge about primary user’s signal. The primary incumbents are detected by comparing the output of the energy detector with a threshold which depends on the noise floor. However, a drawback of the energy detector is that it cannot distinguish between secondary transmission from the primary. Therefore, for energy detection to be successful, it is necessary to forcefully quiet down all the secondary transmissions periodically which degrades the secondary network performance. Other alternative spectrum sensing methods include multitaper spectral estimation [31] and wavelet transform-based estimation. Multitaper spectral estimation is a heuristic approximation to maximum likelihood power spectral density estimator. While the method is fairly applicable for wide-band sensing, it is still computationally demanding. In [32], an FPGA driven prototype to sense spectrum in the UHF band based on waveform analysis is presented. However, 6.1 and 6.9 s are needed for the actual sensing time for W-CDMA and WLAN, respectively. No dynamic synchronization/channel switching upon the detection of primary devices is explored. A feature detector design for TV bands with emphasis on the physical layer is presented in [33]. In [34], incumbent detection method based on IEEE 802.11 PHY/CRC error counts is introduced. However, dynamic frequency switching upon the detection of primary devices is not considered here. Though the research in [34] shows how the PHY/CRC error counts can be an indicator of the existence of incumbents, it relies on a static threshold and lacks adaptive incumbent detection capability.
3
2107
CR PROTOTYPE IMPLEMENTATION CHALLENGES
3.1
Coordination Challenges in Neighbor Discovery and Synchronization In DSA, when a CR node is switched on, it may follow the listen before talk protocol by scanning all the channels and builds a spectrum usage report of vacant and used channels. Unlike the existing single frequency radio devices (which operates using only one static frequency), cognitive radio nodes need to discover its communication peers through extensive channel scanning and beacon broadcasting [35], [36]. This is defined as neighbor discovery phase. Once the cognitive radio nodes locate broadcasts from communicating peers, they then tune to the frequency and transmit back in the uplink direction with the radio node identifier. Authentication and connection registrations are then done gradually to complete the neighbor discovery phase. Due to such extensive connection establishment procedure at the beginning, with the number of dynamic frequency channels increasing, the initial neighbor discovery process is likely to become highly time consuming in DSA environment [4], [35], [36]. However, in a DSA environment, the channel availability may change radically due to the presence of primary users in the bands and unless communicating CR nodes synchronize proactively within a certain time threshold to move to a new band, this might result in synchronization failure. Once the communicating CR nodes suffer synchronization failure, the communication stops resulting in loss of data (degradation in data throughput) and the nodes must go through the highly time-consuming neighbor discovering process repeatedly. An efficient and robust synchronization mechanism is, thus, crucial under such requirements for DSA environment. 3.2 Exchange of List of Current Available Channels Another challenge for the nodes is the method and implementation to exchange the list of currently available channels and the channel to which they will move to resume communication upon detection of a primary in the current operating channel. 3.3 Impact of Switching Channel on Upper Layers Upon successful synchronization and movement from one frequency band to another, the CR node must reconfigure itself to the new frequency channel. For the reconfiguration process, the CR node must restart the hardware to reset and adapt to the particular transmission or reception parameters in the new spectrum band. With the legacy hardware, however, this process is highly inefficient and consumes a lot of time (of the order of seconds [37], [38]). This is a significant delay overhead for channel switching where the aim is to switch as fast as possible. Moreover, at the channel move time, when the CR node reconfigures transmission/reception parameters to the new frequency channel, it needs to stop the data flow from the upper layers. This operation will adversely affect the performance at the higher layers degrading the data throughput performance unless some remedial actions are performed to enhance the MAC for cognitive radio network under DSA requirements.
2108
IEEE TRANSACTIONS ON MOBILE COMPUTING,
VOL. 12,
NO. 11,
NOVEMBER 2013
Fig. 2. Proposed protocol stack for SpiderRadio. Fig. 1. Hidden incumbent scenario.
3.4 Hidden Incumbent Problem Another major concern in this regard is the hidden incumbent problem. Consider two CR nodes (A and B as depicted in Fig. 1) communicating in a specific frequency channel (blue band). Now, suppose, an incumbent D starts transmission near A in the same frequency channel (blue band) as A (refer Fig. 1). Upon detecting the primary device in the blue band, A initiates the channel switch process by dynamically choosing a new available channel from its local spectrum usage report and moves to the new channel. The problem with such a mechanism is that the receiver may have a primary device C already operating in the new frequency channel in its vicinity but outside the sensing region of A (hidden incumbent scenario). The initiator A does not have any information about this frequency usage. As a result, A would end up being in the new channel whereas the node B will still remain in the old channel thereby resulting in synchronization failure and thus the loss of communication. Note that, despite the challenges, dynamic spectrum access must still be simple with a goal toward: 1. 2. 3. 4. 5.
4
efficient sensing, fast channel switching, reduced synchronization failure, reduced synchronization overhead, and increased effective throughput.
SPIDERRADIO ARCHITECTURE AND DESIGN
In this section, we present the proposed SpiderRadio architecture and how it addresses some of the discussed issues.
4.1 Modification to the IEEE 802.11 MAC Layer The proposed software driven protocol stack for SpiderRadio is shown in Fig. 2. The modified Madwifi [39] driver works in a Linux platform. Madwifi is an open source Linux device driver supported by Atheros chipsets. The Atheros wireless card we use supplies raw data to the upper layer (TCP/IP). As depicted in Fig. 2, Madwifi contains three sublayers: IEEE 802.11 media access control layer, the Wrapper (an interface to lower layer) of Atheros Hardware Abstraction Layer, and Atheros Hardware Abstraction Layer. In our implementation, we modified both the IEEE 802.11 Media
Access Control Layer and the Wrapper of Atheros HAL. The IEEE 802.11 Media Access Control Layer is modified to enable faster and more reliable channel switching, while Wrapper of Atheros HAL is modified to build special hardware queues. The details of these modification are presented next. The standard IEEE 802.11 MAC layer is divided into several submodules as shown in Fig. 3 [40]. Our modifications to these modules to implement SpiderRadio are as follows: 1.
2.
Modification to IEEE 80211_Wireless module. This module is responsible for command executions from the user space. In this module, we added an extended user space command. Users could send a switching signal to SpiderRadio for triggering a channel switching. This function is mainly used for testing purposes. With the support of this command, the SpiderRadio software stack developer could test the channel switching scheme easily. Modification to IEEE 80211_Proto module. IEEE 80211_proto controls a state machine that follows the IEEE 802.11a/b/g protocol. It takes care of all the timer functions in Madwifi for timeout events. In the IEEE 802.11 a/b/g protocol, when a node restarts itself, its state machine needs to go to the scan mode, and the neighbor table will be rebuilt. The beacon interval in IEEE 802.11 family of protocols is 100 ms; and sometimes beacon frame will be lost. Under this condition, the process will need more than hundreds of milliseconds. However, in the channel switching function, we have to restart the hardware. To solve this issue and make this process as quick as possible, a modified state machine was implemented in this module. At the bootstrap step of an ad hoc network, every node of this network will just follow the SSID and ESSID that are set up manually; and avoid the method of election process for ESSID. Under this modified module, bootstrapping takes no more than 1 millisecond in the IEEE 802.11 media access layer.
Fig. 3. IEEE 802.11 media access control functional module.
HONG ET AL.: SPIDERRADIO: A COGNITIVE RADIO IMPLEMENTATION USING IEEE 802.11 COMPONENTS
Fig. 4. MAC frame structure of IEEE 802.11 media access control layer.
3.
4.
Modification to IEEE80211_Input module. This module is used to process all the raw data, which comes from the Atheros hardware, including data frame, management frame, and control frame. A SpiderRadio extended management frame process function (called ieee80211_recv_mgmt) was implemented here. Modification to IEEE80211_Output module. As the name suggests, this module takes care of transmit data frame, management frame, and control frame. This module wraps the data from the upper layer into the suitable IEEE 802.11 format, then sends it to the Atheros hardware. The function (called ieee80211_send_mgmt), which is used to create SpiderRadio extended management frame, is implemented in this module.
4.2 Modification to Hardware Abstraction Layer We propose and implement two special hardware queues that become active whenever any dynamic channel switching action needs to be triggered. The first hardware queue is the synchronization queue (sync queue) which is used for transmitting synchronization management frames only. The synchronization management frames are special purpose frames and are used for synchronization between the communicating nodes at the time of switching. Whenever any of the two communicating CR nodes decide to switch channel then the initiator node enables the sync queue and triggers the channel switching request management frame along with the ongoing data communication. Embedded inside the synchronization management frames is the destination channel information (candidate frequency channel(s), i.e., to which the CR nodes desire to switch to upon vacating the current channel). The second hardware queue, data buffer queue, is enabled when the communicating CR nodes are physically switching channels and MAC for both the nodes are being configured with the transmission and reception parameters in the new frequency band. After the data buffer queue is enabled, a local memory is allocated for buffering the data temporarily from upper layers so that no higher data is lost and channel switching does not create any adverse effect. With these hardware queues, dynamic channel switching in the PHY/MAC layer is kept hidden from the upper layers, not affecting the upper layer functionalities thereby creating smooth, seamless channel switching.
Fig. 5. Two-byte frame control field in MAC frame header structure of IEEE 802.11 media access control layer.
2109
Fig. 6. Structure of switching request frame.
4.3
SpiderRadio Extended Management Frame Structure Before we proceed with the discussion of the special management frame structure for the proposed SpiderRadio prototype, we first present the IEEE 802.11 MAC frame structure and MAC header as shown in Figs. 4 and 5. In the MAC header, the 2-bit “Type” field indicates whether the frame is control, management or data frame, while the 4-bit “Sub Type” field indicates different subtypes of frames under one particular type of a frame. For example, with type field set for management frame, there can be 16 different subtypes of management frames. Ten different subtypes of management frames are already defined in IEEE 802.11 [40]. Six more subtypes for 802.11 MAC management frames could be defined out of which we use one subtype of management frame for channel switching and synchronization. Under this subtype, four more extended subtypes are defined. Identification for these extended subtypes is built in the first two bytes of the frame body. The detailed usage of these four different extended management frames is explained in Section 6. In the following, the structures of these four different types of extended management frames are shown. Fig. 6 depicts the structure of the switching request frame. The 2-byte SubType Identification field indicates this as a channel switch request frame. In US, there are three nonoverlapping channels in IEEE 802.11g, and 13 nonoverlapping channels in IEEE 802.11a. The 2-byte destination channel bitmap is sufficient to create a bitmap for all these 16 channels. For bitmapping more number of channels, the 2-byte destination channel bitmap can be extended. The 8-byte timestamp indicates the time when this request frame is prepared for transmission. Fig. 7 depicts the structure of the switching response frame where the 2-byte destination channel field indicates the destination channel information. 4.4
Candidate Frequencies with Bitmap Channel Vector To address the hidden incumbent problem, information regarding multiple available frequencies (candidate frequencies) are embedded inside the channel switching request management frame, instead of initiator CR node attempting to convey channel switching request using only one frequency channel information. The number of candidate frequencies is updated dynamically by the initiator node depending on the feedback received from the receiving communicating CR node. The reason behind embedding
Fig. 7. Structure of switching response frame.
2110
IEEE TRANSACTIONS ON MOBILE COMPUTING,
VOL. 12,
NO. 11,
NOVEMBER 2013
Fig. 9. Nonoverlapping channels and their positions in the bitmap channel vector.
Fig. 8. IEEE 802.11 channel switching request MAC management frame structure.
the available frequencies information in the synchronization message is that even if the receiving CR node encounters a licensed incumbent transmission (hidden to the initiator CR node), it still has ways to choose another candidate channel and report this incumbent transmission to the initiator using a similar management frame called the channel switch response management frame. With this mechanism, even with the presence of hidden node incumbent, risk of synchronization failure is reduced significantly. Fig. 8 shows the proposed channel switching request management frame structure in detail. Recall that, there are three nonoverlapping channels in 802.11g, and 13 nonoverlapping channels in 802.11a which are emulated as primary bands in this work. Obviously, with primary devices dynamically accessing the bands, the availability of the spectrum bands for cognitive radio nodes changes dynamically. Since we use multiple candidate frequency channels sent by initiator nodes, embedding absolute information (spectrum band frequency) of candidate frequency channels would invoke the challenge of variable length management frame. This would result in additional delay to decode the management frame as more processing of header information would be necessary for variable length frames. To solve this issue, we propose to use a bitmap channel vector for sending candidate channel(s) information. In this paper, since we have 16 nonoverlapping channels, we implement the length of this bitmap channel vector as 2 bytes in the MAC payload as shown in Fig. 8, thus mapping each of the channel information to a single bit. When a channel is available (candidate), the corresponding bit is set to 1; otherwise, it is set to 0. In Fig. 9, we present each of the bits in the bitmap channel vector and the corresponding frequency band it signifies. For illustration purposes, let us consider one simple example and refer back to Fig. 1. Assume that the two CR nodes (A and B) are communicating using a specific frequency channel and the incumbent D starts using the same frequency. Assume the candidate channels are 1,
6, 52, and 149; thus, the destination channel bitmap in the channel switching request management frame initiated by A would be 0000; 1000; 1000; 0011. If the hidden incumbent C is already operating using channel 1, B would simply discard the bit corresponding to channel 1 (flip it to zero) and reconsider the modified bitmap vector to decide on the desired candidate channel. The information would then be responded back to A with the channel switch response management frame. Note, the primary advantage of using the bitmap channel vector for transmitting candidate frequency information is that a fixed length management frame can be used even for the case of dynamic availability of candidate frequency information. Moreover, the bitmap channel vector makes the management frame easily scalable. If there are more than 16 nonoverlapping channels, we only need to expand the programmable bitmap vector field for that system. Following the destination channel bitmap vector field, the next field signifies the final timeout switching time from initiator’s perspective. This is designed to indicate when the initiator will timeout from the current synchronization mechanism (if no synchronization could be established; i.e., even after multiple switching requests, no response frame is received from the other communicating CR node), will vacate the current channel and start the resynchronization attempt through quick probing following the destination (candidate) channel bitmap vector.
5
SENSING/DETECTION OF PRIMARY INCUMBENTS
Incumbent sensing/detection is one of the important features of a cognitive radio network. However, as mere energy, noise and interference detection is not sufficient to distinguish a primary incumbent communication from other cognitive radio (secondary) communications, thus it is important to develop a modified approach for accurate sensing/detection. Note that, the preamble of the packets transmitted by primary incumbents are different than that of the secondary cognitive radio nodes thus providing the basis for our sensing/detection technique. PHY errors are reported from the wireless physical layer interface to the upper layers if packets/signal without the intended 802.11 PHY preamble are observed by the physical layer interface [34]. Thus, whenever, the primary incumbents are present and transmitting, cognitive radios present in that channel will observe packets with different packet preambles or corrupted packet preambles (known as observed PHY errors) upon sensing of the channel. That is why our
HONG ET AL.: SPIDERRADIO: A COGNITIVE RADIO IMPLEMENTATION USING IEEE 802.11 COMPONENTS
2111
Fig. 10. CDF and (1-CDF) of PHY errors for one window case at received signal strength (a) 82 dBm and (b) 87 dBm.
proposed PU detection mechanism is applicable to intended TV bands or any other primary incumbents’ bands as well. Another advantage of such a PHY error-based sensing method is that unlike energy detection, the radio nodes need not forcefully quiet down periodically. However, there are several major challenges in a PHY error-based sensing approach. First, the number of observed PHY errors is not fixed due to the unreliable nature of the wireless medium. Second, even for a stable wireless channel quality, the number of observed PHY errors varies drastically based on the received energy strength. Third, depending on the duration of sensing (sensing window), the observed PHY errors vary. Finally, and probably the most important, in our cognitive radio network framework, we no longer assume the notion of periodic sensing by forcefully making all the radio nodes quiet because of the degraded throughput outcome. Rather, the CR nodes are allowed to carry out communications even at the time of sensing; however, the challenge with such approach is that the number of observed PHY errors may actually be suppressed. Clearly, since the wireless environment varies dynamically a simple static detection threshold of observed PHY error is not sufficient to sense/detect incumbents. The detection threshold is defined as the number of observed PHY errors (in a certain sensing duration) exceeding the value of which signifies the presence of primary incumbent. If the threshold is set too low, the cognitive radio may incur a high number of false alarms (i.e., detecting the existence of primary incumbent while the incumbent is actually not present). On the other hand, if the threshold is set too high, there may be a high probability of misdetection (missing to detect the primary when it is present). Therefore, we need to investigate a dynamic thresholding scheme that adapts to the time-varying wireless environment. We begin with modeling the probability distribution of the observed PHY errors both in the presence and absence of primary incumbents (cordless phones and agilent spectrum signal generator are configured and used as primary devices for the experiment purpose) and under different received energy strengths. The experimental setup is such that the primary devices transmit with different transmitting powers and are located at various distances from the SpiderRadio nodes such that the received energy strength at the SpiderRadio nodes varies from 82 to
102 dBm. Note that the reason behind choosing such low received energy for the experiment is that the observed PHY errors has a lot of uncertainty at these low received power levels. For basic sensing, a window duration of 20 ms is used. The number of observed PHY errors in each observation window is reported to the upper layers. Note that, any other sensing duration could also be chosen as a basic sensing window duration as long as the PHY errors can be accurately reported through interrupt signaling from the wireless card physical layer to the upper layer. In Figs. 10a and 10b, we present the cumulative distribution function (CDF) and the complementary cumulative distribution function (1-CDF) of the experimentally observed PHY errors for different received energy strengths for two cases: 1) when primary incumbent is present and 2) when the primary incumbent is not present. The sensing duration is 20 ms, i.e., 1-sensing window. It can then be stated that, the CDF of the case when primary incumbent is present is actually the representation of the probability of misdetection given a certain PHY error threshold. Similarly, (1-CDF) of the case when primary incumbent is NOT present is actually the representation of the probability of false alarm given a certain PHY error threshold. From Figs. 10a and 10b, it is clear that a low detection threshold on the observed PHY errors will reduce the misdetection probability, however will increase the false alarm probability. Increasing the detection threshold has the opposite effect. It is easy to see that if the threshold is chosen to be the point on the X-axis where the two curves intersect it minimizes the average error probability. From Figs. 10a and 10b, we see that the minimum average error probability is 0.11 when the received energy strength is 82 dBm and 0.25 for 87 dBm. This clearly indicates that with diminishing received energy strength, the performance of 1-sensing window degrades steadily. Therefore, we investigate an n-moving sensing window, i.e., PHY errors are observed over n-sensing window durations and the window is moved continuously. Note that, n is a finite positive integer. For illustration purposes, we present the testbed experimental results for a fivemoving window in Figs. 11a and 11b. For a five-moving window sensing strategy, the simultaneous minimization of both the probabilities give us the optimal threshold and as the Figs. 11a and 11b indicate the minimum error probability is 0.01 for 82 dBm and 0.12 for 87 dBm.
2112
IEEE TRANSACTIONS ON MOBILE COMPUTING,
VOL. 12,
NO. 11,
NOVEMBER 2013
Fig. 11. CDF and (1-CDF) of PHY errors for five window case at received signal strength (a) 82 dBm and (b) 87 dBm.
Using the experimentally observed data, we investigate an analytical model for misdetection and false alarm probability. To compute the correct model for the probability distributions of the observed experimental data, we use the multi-variate regression analysis with curve fitting.1 Then, if we denote Pm as the misdetection probability, using multivariate regression analysis, Pm can be expressed as Pm ¼
1 ; 1 þ eðrÞ
ð1Þ
given an observed normalized number of PHY error r is taken as the detection threshold. The parameters, and are parameters that depend on the sensing window length and the received energy strength and are given by ¼ 1 þ 2 S þ 3 en þ 4 S 2 þ 5 ðen Þ2 þ 6 Sen ;
ð2Þ
where S denotes the received energy strength and i ’s are the regression variables. For example, the values of i ’s as derived in this work using minitab software [42] are as follows: 1 ¼ 55:1; 2 ¼ 40:7; 3 ¼ 405; 4 ¼ 9:23; 5 ¼ 342; 6 ¼ 59:4:
ð3Þ
where 0 and 0 are also the parameters that depend on the sensing window length and the received signal strength. These can be computed similar to and as mentioned above. Note that, at any stage of the sensing, when a cognitive radio makes a wrong decision about a primary incumbent (sensing failure), it faces one of two possible costs. If the incumbent sensing decision results in misdetection the cost (penalty) to the cognitive radio is primary network policy specific. For example, the cognitive radio may be blocked from accessing the spectrum access for a certain amount of time. In this paper, we assume this cost to be C1 (blocked) time units.2 On the other hand, due to a false alarm, the cognitive radio switches to some other clear band and it faces a cost of finding a clear spectrum band. We assume this cost to be C2 time units. Once the probabilities of misdetection and false alarm probabilities are calculated, the aim is to find out the dynamic optimized threshold such that the penalty due to misdetection and false alarm is minimized. The optimization problem then becomes the minimization of the average expected cost given by EðCÞ ¼ C1 Pm þ C2 Pf :
ð7Þ
Expanding the above equations, we have
Similarly, is given by ¼ 1 þ 2 n þ 3 eS þ 4 ðeS Þ2 þ 5 neS ;
ð4Þ
EðCÞ ¼
C1 C2 þ : 1 þ eðrÞ 1 þ e0 ð0 rÞ
ð8Þ
where i ’s are the regression variables. The values of i ’s as derived in this work are as follows:
To compute the optimal threshold, r , the first-order condition gives
1 ¼ 0:0717; 2 ¼ 0:0197; 3 ¼ 0:0432; 4 ¼ 0:00171; 5 ¼ 0:00225:
C2 0 e ð rÞ ¼ 0: 2 ðrÞ ½1 þ e ½1 þ e0 ð0 rÞ 2
ð5Þ
Similar to the misdetection probability expression, the false alarm probability can be expressed as Pf ¼
1 1þ
e0 ð0 rÞ
;
ð6Þ
0
C1 eðrÞ
0
ð9Þ
By solving this, we find the optimal threshold, r ¼ r , to satisfy ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s Þ cosh ðr C1 2 : ð10Þ 0 ð0 r Þ ¼ 0 C2 cosh 2
1. Note that, multivariate regression analysis is used as a tool to compute the correct model for the probability distributions from the observed experimental data. Since we are observing three parameters, i.e., observed PHY errors, received energy strengths and n-moving window for sensing, based on which the model/decision is being made, multivariate regression is a natural choice [41].
2. The real-world significance and enforcement of this cost (penalty) C1 is not yet finalized and IEEE working group of WRAN is still working on this issue. However, some of the options being investigated are: monetary penalty, restricting spectrum access for a certain time, limiting data rate, and so on.
HONG ET AL.: SPIDERRADIO: A COGNITIVE RADIO IMPLEMENTATION USING IEEE 802.11 COMPONENTS
2113
Fig. 12. Dynamic frequency switching with bitmap vector enabled channel switching request and response management frame. 2
EðCÞ The second-order differentiation, d dr , can be shown to be 2 positive (with simple derivations), proving the function EðCÞ to be convex with respect to r. Since E(C) is convex with respect to r, r is the desired minimizer. Note that r will be computed periodically when the radio environment changes dynamically.
6
DYNAMIC SYNCHRONIZATION AND CHANNEL SWITCHING
In Fig. 12, the schematic diagram for dynamic channel switching is shown. The channel switching procedure is divided into two phases. The initial phase establishes the synchronization through multiple attempts (if needed) of management frame handshaking and enables the communicating nodes to switch channel successfully. If the initial phase fails, the second phase acts as a quick rescue process, which introduces a channel probing mechanism that takes charge of resynchronization of both the nodes.
6.1 Initial Synchronization Phase When channel switching is required (sensed by one of the two communicating CR nodes), the initiating node will send a switch request frame, including the destination channel bitmap vector in the payload. However, there is a possibility that both the communicating nodes detect the arrival of a primary device, become initiators and transmit channel switch request management frames with information about the candidate channel(s) from their local spectrum usage reports. This might cause a clash in the synchronization mechanism. However, this problem is alleviated by the timestamp field in the request management frame structure. This field indicates when the request was generated. If both the CR nodes initiate channel switch request, the one with earlier timestamp will win. The 8-byte timestamp field from the switching request management frame will let both the nodes decide on the winner (initiator) and the other node will automatically follow the role of receiver. Note that, clock synchronization is not necessary in our case for the use of timestamp in resolving the winner, because only a unique winner is needed to be determined through time-stamp in case both the communicating nodes detect the arrival of a primary device. Upon reception of the request frame, the receiver then decides on the new channel to switch to (deciding from the candidate channel bitmap vector) and transmits the channel switching response management frame back to the initiator. Using the information in the response frame, the initiator and the
Fig. 13. Switching request flow chart from initiator’s perspective.
receiver are now mutually agreed upon the destination channel to which they would move to. As the last step of the handshaking procedure, the initiator sends a quick sync. frame acknowledgment to the receiver and moves to the new channel. Receiver upon receiving the sync. frame also immediately moves to the new channel. Fig. 13 depicts the work flow of this management handshake system from the initiator’s perspective. As shown in the Fig. 13, after the initial channel switch request frame is transmitted a response timer is set up. When the initiator node receives the switch response frame it sends a sync message frame and then switches into the destination channel immediately. However, if no response frame is received within the response time, the initiator will check the number of attempts and the final timeout switching time. In the implementation, we fix the maximum number of attempts to be three and the final timeout switching time to be five times the estimated single round trip timeout. Thus, if the timing permits and if the number of attempts made is less than 3, the initiator node will repeat the switch request, setup response timer, wait for timer expired and look for the response frame again. In this regard, we implement and compare two management frame transmission schemes like ð1; 1; 1Þ and ð1; 2; 3Þ in the proposed prototype. In terms of notation, a ð1; 1; 1Þ scheme uses one management frame in each of the three synchronization attempt(s). A ð1; 2; 3Þ scheme uses three attempts of synchronization with one frame transmitted in the first attempt, two frames in the second attempt, and three frames in the third and final attempt. Note that, the ð1; 2; 3Þ scheme presents more robustness due to redundancy compared to the ð1; 1; 1Þ scheme, however, at the same time, it incurs additional overhead in terms of a higher number of management frames compared to the ð1; 1; 1Þ scheme. Thus, it is useful to investigate this tradeoff. In contrast, if even after three attempts no response frame is received and the response timer expired this node will go into the second phase of channel switching, i.e., the quick probing-based resynchronization.
6.2
Resynchronization Phase through Quick Probing Fig. 14 shows the progress of how to resynchronize two nodes if initial phase synchronization fails. Note that, even if the synchronization failure happens (initial phase failure), it is not desired for the communicating CR nodes to go to the neighbor discovery process straightaway
2114
IEEE TRANSACTIONS ON MOBILE COMPUTING,
VOL. 12,
NO. 11,
NOVEMBER 2013
Fig. 15. (a) IEEE 802.11 wireless access card used in CR prototype implementation and (b) cordless phones acting as primary devices.
Fig. 14. Resynchronization phase through probing.
because of the high time consuming nature of the process. Thus, we intend to devise a middle path rescue process through this phase. Note, initial synchronization phase may fail (even after multiple attempts) due to one of three following reasons: 1) All the switching request frames themselves are lost (though we have not encountered this case in any of our experiments; however, for comprehensiveness of the research we mention this scenario) resulting in a fatal scenario of synchronization failure in which the CR nodes may end up in different spectrum bands resulting in loss of communication. Receiver node does not have any information about the candidate channel(s). In such scenario, the CR nodes have no option but to repeat the neighbor discovering process again. 2) The receiver may have received the switching request frame(s) from the initiator, but the response frame got lost in all the response attempts or 3) the receiver may have received the switching request frame(s) from the initiator, sent out the response frame(s) to the initiator, initiator received them successfully and sent out the sync ack but the sync ack got lost. In either case 2 or 3, the receiver node would understand this if it does not receive any sync ack message from the initiator. However, note that, the receiver now has the knowledge of the candidate channel(s) and the final timeout switching time from the initiator’s perspective. The receiver, after final timeout switching time, then moves to the first available channel from the bitmap (in terms of channel number) and so on. From the initiator’s perspective, if the initiator has already received the response frame, the synchronization is successful; so after sending out the sync ack messages, the initiator will move to the first available candidate channel as synchronized. However, if the initiator did not receive any of the response frames (assuming all the response frames got lost), Initiator does not have any knowledge of whether the receiver has received the desired information; thus assumes this as a synchronization failure, times out, vacate the current channel and starts the resynchronization attempt through quick probing. At the beginning of this phase, the initiator selects the candidate channel(s) from the destination channel bitmap vector for the purpose of probing. The initiator will go to first candidate channel, transmit a quick probe management frame and try to sense if there is any probe response from
the receiver within the round trip timeout. If it receives probe response, both nodes will be resynchronized; otherwise, initiator moves onto the next candidate channel. This process will be repeated until probe response is received in one of the candidate channels or all the candidate channels are scanned; after which the CR nodes will go to the neighbor discovery process again.
7
TESTBED SETUP AND EXPERIMENTAL RESULTS
For evaluating, the implementation of proposed CR prototype system with adaptive threshold-based sensing, dynamic synchronization, and channel switching policy, we conducted extensive experiments and present the results. In this regard, we also compare both the proposed management frame transmission schemes, ð1; 1; 1Þ and ð1; 2; 3Þ. For performance evaluation purpose, we primarily focus on the following metrics: . . . . .
probability of sensing failure, penalty induced due to wrong decision outcomes, average synchronization attempt, average time to synchronize, and effective throughout.
7.1 Testbed Setup For conducting extensive experiments with SpiderRadio enabled nodes, we built two groups of SpiderRadio prototypes: one for indoor testing and the other for outdoor testing. Each node of the indoor group is a standard desktop PC running Linux 2.6 operating system. They were all equipped with Orinoco 802.11 a/b/g PCMCIA wireless card (Fig. 15a). There is no PCMCIA slot for the desktop PC; so we make use of an ENE-CB1410 PCMCIA to PCI adapter card for converting the PCMCIA devices to operate on the desktop PC. In the outdoor group, SpiderRadio is deployed on two laptops running Linux 2.6 operating system: Compaq NC4010 and Dell Inspiron 700 m. Both of them were equipped with Orinoco 802.11 a/b/g PCMCIA wireless card. The TX powers of these wireless devices were set to 100 mW. Another laptop running Windows VISTA and equipped with Wi-Spy 2.4x, acted as a monitor in the testbed. These Orinoco devices are equipped with Atheros 5212 (802.11 a/b/g) chipsets. For our testbed setup, the primary user bands were emulated using the 900 MHz, 2.4 GHz, and 5.1 GHz Wi-Fi spectrum bands. The primary user communication was emulated in two different ways: two cordless phones communicating with each other using the intercom feature
HONG ET AL.: SPIDERRADIO: A COGNITIVE RADIO IMPLEMENTATION USING IEEE 802.11 COMPONENTS
2115
Fig. 16. Probability of sensing failure with one-sensing window mechanism.
Fig. 17. Probability of sensing failure with five-sensing window mechanism.
(Fig. 15b) and Agilent signal generator (e4437b) operating in the Wi-fi bands. The SpiderRadio node was configured to be the secondary user device for the experiments. For the purpose of sensing and detecting the arrival of primary user, we implemented the spectrum sensing methodology based on observed PHY errors, received signal strengths and n-moving window strategy as proposed. We placed SpiderRadio nodes at a distance of 5-20 m from each other, communicating with TCP data streams. We carried out experiments under five different network traffic load scenarios: 3 MB/s, 2 MB/s, 1 MB/s, 0.5 MB/s, and 10 KB/s during day and night times. Note that since the testbed is located in Hoboken (in close proximity to Manhattan, New York City) the radio interference is significantly different during day and night. Interference due to students using the Stevens campus wireless network also varies significantly between night and day.
proposed dynamic threshold clearly produces better results in terms of reduced number of wrong decisions; probability of sensing failure is as low as 0.018. Fig. 17 shows similar results with five-sensing window with probability of sensing failure even more reduced (0.01376) thus clearly demonstrating the effectiveness of proposed n-moving window sensing methodology. The average penalty (cost in terms of time units) induced due to the wrong decisions (either misdetection or false alarm) made by the CR nodes are presented in Figs. 18a and 19a. Note that, at any stage of the sensing, when a CR node makes a wrong decision about a primary incumbent (sensing failure), it faces one of two possible costs in terms of time units (spectrum access denial penalty for time C1 in case of misdetection or unnecessary switching cost penalty of time C2 in case of false alarm). In our experiments, we assume, C1 ¼ 100 ms, while the value of C2 is the actual time taken to switch and mostly varies in between 5 and 20 ms. It is seen from the figures that dynamic threshold based on received energy and n-moving window sensing clearly outperforms the static threshold mechanisms in inducing minimized penalty. For comprehensiveness, we also demonstrate the corresponding dynamic threshold values from a snapshot of the experiment in Figs. 18b and 19b.
7.2 Experimental Results To evaluate the effectiveness of the proposed primary sensing/detection methodology, Fig. 16 shows the probability of sensing failure with one-sensing window and compares the dynamic threshold scheme with other static threshold schemes. As evident from the figure, the
Fig. 18. (a) Average penalty induced due to sensing failure; (b) a snapshot of the corresponding dynamic thresholds as adapted by SpiderRadio based on received energy strength and one-sensing window mechanism.
2116
IEEE TRANSACTIONS ON MOBILE COMPUTING,
VOL. 12,
NO. 11,
NOVEMBER 2013
Fig. 19. (a) Average penalty induced due to sensing failure; (b) a snapshot of the corresponding dynamic thresholds as adapted by SpiderRadio based on received energy strength and five-sensing window mechanism.
Next, in Figs. 20 and 21, we present the time to switch and synchronize (in milliseconds) for CR nodes with ð1; 1; 1Þ and ð1; 2; 3Þ schemes, respectively. We define the time to switch and synchronize as the total-time consumed by the CR nodes to switch from old channel and resynchronize successfully on the new channel. Each of the experiment is carried out for an hour in daytime under high network traffic load where time to switch and synchronize is recorded for 2,000 instances of switching. It is clear from the figures that even under high network traffic load (approximately
Fig. 20. Time to synchronize under high network traffic (approximately 3,000 KBps) with ð1; 1; 1Þ scheme.
Fig. 21. Time to synchronize under high network traffic (approximately 3,000 KBps) with ð1; 2; 3Þ scheme.
3,000 KBps), the proposed CR prototype successfully proves its robustness with very low time to synchronize (the maximum recorded time to switch and synchronize never exceeded 80 ms for ð1; 1; 1Þ mechanism, while for ð1; 2; 3Þ, it never exceeded 70 ms). To provide a better insight into the results, we next compare the average (averaged over 2,000 instances of switching for each of the experiments) switch and synchronization times for different network traffic load states and under both ð1; 1; 1Þ and ð1; 2; 3Þ schemes. In Fig. 22, we present the results. Seven different network traffic loads are considered here. It is evident from the results that with network traffic load decreasing, average time to synchronize also decreases for both the mechanisms as less number of synchronization attempts are required with low network traffic load. Moreover, it is seen from the comparisons that ð1; 2; 3Þ scheme produces better results than ð1; 1; 1Þ in terms of average time required. However, another interesting observation from the comparison is that, when the network traffic load is very high (approximately 3 or 2 MB/s), there is not much difference between ð1; 2; 3Þ and ð1; 1; 1Þ scheme. This is because req./responses are lost in both the cases almost to the similar level leading to almost equal synchronization time. The difference increases as the network traffic load is medium (approximately 1.5-0.5 MB/s), where the ð1; 2; 3Þ scheme produces significantly better
Fig. 22. Average synchronization time under various network traffic with ð1; 1; 1Þ and ð1; 2; 3Þ scheme.
HONG ET AL.: SPIDERRADIO: A COGNITIVE RADIO IMPLEMENTATION USING IEEE 802.11 COMPONENTS
Fig. 23. (a) Average channel switching request management frames required, and (b) Average synchronization attempts; under various network traffic with ð1; 1; 1Þ and ð1; 2; 3Þ scheme.
results than ð1; 1; 1Þ. Again the difference decreases when the network traffic load is low (approximately 100 or 10 KB/s). Next, in Figs. 23a and 23b, we show the comparison results of average number of channel switching request management frames required and average number of synchronization attempts made for each of the dynamic channel switching attempts under both ð1; 1; 1Þ and ð1; 2; 3Þ scheme. The comparison is made under all seven different types of network traffic loads specified earlier. It is observed from the result that, though average number of switching request management frames (overhead) required to synchronize in each switching is slightly greater in case of ð1; 2; 3Þ, but average number of synchronization attempts made under ð1; 2; 3Þ scheme is lesser than ð1; 1; 1Þ thus resulting in less average time to switch and synchronize as seen earlier in Fig. 22. Last, but not the least, we present the effective throughput achieved in the proposed CR prototype with the proposed dynamic switching policy in Fig. 24. The experiment is conducted to evaluate the performance of the proposed system and its effect on the data throughput based on different arrival rate of the primary incumbents and accordingly CR node switching from the current channel. To conduct this experiment, we controlled the incumbent arrival such that the “desired” switching interval is achieved. The results are plotted against different switching intervals (1, 2, 3, 5, and 10 s). For benchmarking purpose, we calculate the ideal maximum throughout achieved under the same environment and conditions without any frequency switching. The dotted line in the figure depicts the maximum possible throughput (3.353 MB/s—benchmark). As evident from the figure, the proposed CR system demonstrates high throughput even with highly frequent switching and as is obvious, with less frequent switching (switching every 5 or 10 s), the throughput achieved is almost same as the benchmark throughput; proving the effectiveness of the proposed CR prototype.
8
CONCLUSIONS
SpiderRadio demonstrates that software abstraction of MAC layer implemented on commodity hardware is a feasible option for dynamic spectrum access networking. By treating the commodity wireless network interface card as a black box, it is possible to design and implement quick and reliable spectrum sensing algorithms that exploit the inherent characteristics of the interface. Fast channel
2117
Fig. 24. Average effective throughput with various frequency intervals.
switching can be implemented with negligible synchronization failure probability using the software abstraction. High effective throughput is achievable using the proposed implementation as seen in the experimental results.
ACKNOWLEDGMENTS This research was partially funded by US National Science Foundation (NSF) grants CNS #0917008, CCF #0916180, CNS #1149920, and NSF #1256996.
REFERENCES [1] [2] [3] [4]
[5]
[6]
[7] [8]
[9] [10] [11]
[12]
[13]
http://www.sharedspectrum.com/inc/content/measurements/ nsf/NYC_report.pdf, 2013. Federal Communications Commission (FCC), “Notice of Proposed Rule Making,”ET Docket No. 04-113, May 2004. J. Mitola and G. Maguire Jr., “Cognitive Radio: Making Software Radios More Personal,” IEEE Personal Comm., vol. 6, no. 4, pp. 1318, Aug. 1999. I.F. Akyildiz, W.-Y. Lee, M.C. Vuran, and S. Mohanty, “Next Generation/Dynamic Spectrum Access/Cognitive Radio Wireless Networks: A Survey,” Computer Network, vol. 50, no. 13, pp. 21272159, 2006. G. Ganesan and Y. Li, “Cooperative Spectrum Sensing in Cognitive Radio Networks, Part I: Two User Networks and Part II: Multiuser Networks,” IEEE Trans. Wireless Comm., vol. 6, no. 6, pp. 2204-2222, June 2007. Z. Quan, S. Cui, A.H. Sayed, and H.V. Poor, “Optimal Multiband Joint Detection for Spectrum Sensing in Cognitive Radio Networks,” IEEE Trans. Signal Processing, vol. 57, no. 3, pp. 1128-1140, Mar. 2009. A. Ghasemi and E.S. Sousa, “Spectrum Sensing in Cognitive Radio Networks: The Cooperation-Processing Tradeoff,” Wireless Comm. and Mobile Computing, vol. 7, no. 9, pp. 1049-1060, 2007. Z. Ye, G. Memik, and J. Grosspietsch, “Energy Detection Using Estimated Noise Variance for Spectrum Sensing in Cognitive Radio Networks,” Proc. IEEE Wireless Comm. and Networking Conf. (WCNC), pp. 711-716, Apr. 2008. IEEE 802.22 Working Group on Wireless Regional Area Networks, http://www.ieee802.org/22/, 2013. T. Yucek and H. Arslan, “A Survey of Spectrum Sensing Algorithms for Cognitive Radio Applications,” IEEE Comm. Surveys and Tutorial, vol. 11, no. 1, pp. 116-130, Mar. 2009. N. Choi, M. Patel, and S. Venkatesan, “A Full Duplex MultiChannel MAC Protocol for Multi-Hop Cognitive Radio Networks,” Proc. First Int’l Conf. Cognitive Radio Oriented Wireless Networks and Comm., June 2006. C. Cordeiro and K. Challapali, “C-MAC: A Cognitive MAC Protocol for Multi-Channel Wireless Networks,” Proc. IEEE Second Int’l Symp. New Frontiers in Dynamic Spectrum Access Networks (DySPAN), pp. 147-157, Apr. 2007. L. Lai, H.E. Gamal, H. Jiang, and H. Poor, “Cognitive Medium Access: Exploration, Exploitation and Competition,” IEEE Trans. Mobile Computing, vol. 10, no. 2, pp. 239-253, Feb. 2011.
2118
IEEE TRANSACTIONS ON MOBILE COMPUTING,
[14] B. Lee and S. Rhee, “Adaptive MAC Protocol for Throughput Enhancement in Cognitive Radio Networks,” Proc. Int’l Conf. Information Networking (ICOIN), pp. 1-5, Jan. 2008. [15] H. Salameh, M. Krunz, and O. Younis, “MAC Protocol for Opportunistic Cognitive Radio Networks with Soft Guarantees,” IEEE Trans. Mobile Computing, vol. 8, no. 10, pp. 1339-1352, Oct. 2009. [16] H. Su and X. Zhang, “Cross-Layer Based Opportunistic MAC Protocols for QoS Provisionings over Cognitive Radio Wireless Networks,” IEEE J. Selected Areas in Comm., vol. 26, no. 1, pp. 118129, Jan. 2008. [17] Q. Zhao, L. Tong, and A. Swami, “Decentralized Cognitive MAC for Dynamic Spectrum Access,” Proc. IEEE Int’l Symp. New Frontiers in Dynamic Spectrum Access Networks, pp. 224-232, Nov. 2005. [18] S. Sengupta, R. Chandramouli, S. Brahma, and M. Chatterjee, “A Game Theoretic Framework for Distributed Self-Coexistence among IEEE 802.22 Networks,” Proc. IEEE Global Telecomm. Conf. (GlobeCom), pp. 1-6, 2008. [19] IEEE P802.22 Standard for Wireless Regional Area Networks Part 22: Cognitive Wireless RAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications: Policies and Procedures for Operation in the TV Bands, IEEE, 2006. [20] I.F. Akyildiz, W.-Y. Lee, and K.R. Chowdhury, “CRAHNs: Cognitive Radio Ad Hoc Networks,” Ad Hoc Networks, vol. 7, pp. 810-836, http://portal.acm.org/citation.cfm?id=1508337. 1508834, July 2009. [21] C. Cormio and K. Chowdhury, “A Survey on MAC Protocols for Cognitive Radio Networks,” Ad Hoc Networks, vol. 7, no. 7, pp. 1315-1329, 2009. [22] Y. Yuan, P. Bahl, R. Chandra, P. Chou, J. Ferrell, T. Moscibroda, S. Narlanka, and Y. Wu, “Knows: Cognitive Radio Networks over White Spaces,” Proc. IEEE Second Int’l Symp. New Frontiers in Dynamic Spectrum Access Networks (DySPAN), pp. 416-427, 2007. [23] P. Bahl, R. Chandra, T. Moscibroda, R. Murty, and M. Welsh, “White Space Networking with Wi-Fi Like Connectivity,” SIGCOMM Computer Comm. Rev., vol. 39, pp. 27-38, http:// doi.acm.org/10.1145/1594977.1592573, Aug. 2009. [24] M. Sharma and A. Sahoo, “Opportunistic Channel Access Scheme for Cognitive Radio System Based on Residual White Space Distribution,” Proc. IEEE Int’l Symp. Personal, Indoor and Mobile Radio Comm., pp. 1842-1847, 2010. [25] A. Plummer, M. Taghizadeh, and S. Biswas, “Measurement-Based Bandwidth Scavenging in Wireless Networks,” IEEE Trans. Mobile Computing, vol. 11, no. 1, pp. 19-32, Jan. 2012. [26] A.T. Hoang, Y.-C. Liang, and M. Islam, “Power Control and Channel Allocation in Cognitive Radio Networks with Primary Users’ Cooperation,” IEEE Trans. Mobile Computing, vol. 9, no. 3, pp. 348-360, Mar. 2010. [27] W.-Y. Lee and I. Akyldiz, “A Spectrum Decision Framework for Cognitive Radio Networks,” IEEE Trans. Mobile Computing, vol. 10, no. 2, pp. 161-174, Feb. 2011. [28] A. Tajer and X. Wang, “Beacon-Assisted Spectrum Access with Cooperative Cognitive Transmitter and Receiver,” IEEE Trans. Mobile Computing, vol. 9, no. 1, pp. 112-126, Jan. 2010. [29] N. Khambekar, L. Dong, and V. Chaudhary, “Utilizing OFDM Guard Interval for Spectrum Sensing,” Proc. IEEE Wireless Comm. and Networking Conf., pp. 38-42, Mar. 2007. [30] F. Digham, M. Alouini, and M. Simon, “On the Energy Detection of Unknown Signals over Fading Channels,” Proc. IEEE Int’l Conf. Comm., vol. 5, pp. 3575-3579, May 2003. [31] S. Haykin, “Cognitive Radio: Brain-Empowered Wireless Communications,” IEEE J. Selected Areas in Comm., vol. 3, no. 2, pp. 201220, Feb. 2005. [32] H. Harada, “A Software Defined Cognitive Radio Prototype,” Proc. IEEE 18th Int’l Symp. Personal, Indoor and Mobile Radio Comm. (PIMRC), pp. 1-5, 2007. [33] R. DeGroot, D. Gurney, K. Hutchinson, M. Johnson, S. Kuffner, A. Schooler, S. Silk, and E. Visotsky, “A Cognitive-Enabled Experimental System,” Proc. IEEE First Int’l Symp. New Frontiers in Dynamic Spectrum Access Networks (DySPAN), pp. 556-561, 2005. [34] K. Shin, H. Kim, C. Cordeiro, and K. Challapali, “An Experimental Approach to Spectrum Sensing in Cognitive Radio Networks with off-the-Shelf IEEE 802.11 Devices,” Proc. IEEE Fourth Consumer Comm. and Networking Conf. (CCNC), pp. 1154-1158, Jan. 2007. [35] C. Cordeiro, K. Challapali, D. Birru, and S. Shankar, “IEEE 802.22: The First Worldwide Wireless Standard Based on Cognitive Radios,” Proc. IEEE First Int’l Symp. New Frontiers in Dynamic Spectrum Access Networks (DySPAN), pp. 328-337, 2005.
VOL. 12,
NO. 11,
NOVEMBER 2013
[36] S. Sengupta, S. Brahma, M. Chatterjee, and S. Shankar, “Enhancements to Cognitive Radio Based IEEE 802.22 Air-Interface,” Proc. IEEE Int’l Conf. Comm. (ICC), pp. 5155-5160, 2007. [37] IEEE Standard for Information Technology - Telecommunications and Information Exchange between Systems - Local and Metropolitan Area Networks-Specific requirements - Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, IEEE, Mar. 2007. [38] T. Hamdan, H. Sigiuk, and Y. Omar, “Reduction of Handoff Search Phase Time in IEEE 802.11 WLAN to Fulfill Real Time Services Requirements,” Proc. Int’l Conf. Telecomm., pp. 346-351, May 2009. [39] http://madwifi-project.org, 2013. [40] IEEE Standard for Information Technology - Telecommunications and Information Exchange between Systems - Local and Metropolitan Area Networks - Specific Requirements - Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, IEEE, Mar. 2007. [41] J. Jobson, Applied Multivariate Data Analysis: Regression and Experimental Design. Springer Verlag, 1991. [42] http://www.minitab.com, 2013. Kai Hong received the BS and MS degrees in automatic control from the Beijing Institute of Technology, China, in 2004 and 2007, respectively. He is currently working toward the PhD degree in computer engineering at the Stevens Institute of Technology, where he is a member of the Media Security, Networking, and Communications (MSyNC) Lab. His research interests include the area of cognitive radios, dynamic spectrum access networks, and wireless security. He is a student member of the IEEE. Shamik Sengupta received the BE degree (first class hons.) in computer science from Jadavpur University, India, in 2002 and the PhD degree from the School of Electrical Engineering and Computer Science, University of Central Florida, Orlando, in 2007. He is an assistant professor in the Department of Computer Science and Engineering, University of Nevada, Reno. His research interests include cognitive radio, dynamic spectrum access, game theory, and security in wireless networking. He serves on the organizing and technical program committees of several IEEE conferences. He is the recipient of an IEEE GlobeCom 2008 best paper award. He is a member of the IEEE. R. Chandramouli is the Thomas Hattrick Chair Professor of Information Systems in the Department of Electrical and Computer Engineering at the Stevens Institute of Technology, cofounder and chief strategist of Dynamic Spectrum, LLC (a technology company offering dynamic spectrum management solutions), and cofounder of jaasuz.com (offering advanced text mining technologies). His research spans the areas of cognitive wireless networking, text mining, social media security, and analytics. His projects in these areas have been supported by the US National Science Foundation (NSF), US National Institute of Justice, US Department of Defense agencies, and industry. He was an invited member of the White House Communications Roundtable to give input on the National Wireless Initiative, an IEEE ComSoc Distinguished Lecturer, founding chair of the IEEE ComSoc Technical Committee on Cognitive Networks (TCCN), TCCN’s representative to the IEEE ComSoc Standards Board, founding editor of the IEEE Journal on Selected Areas in Communications Cognitive Radio Series, founding editor of the Advances in Multimedia Journal, associate editor of the IEEE Transactions on Circuits and Systems for Video Technology, and a member of the international advisory boards for several international conferences and journals. He received an IEEE GlobeCom 2008 Best Paper Award, IEEE CCNC 2006 Best Student Paper Award, NSF CAREER Award, and IEEE Richard E. Merwin Scholarship. He is a senior member of the IEEE.
