Cooperative adaptive spectrum sharing in cognitive radio networks

1 Cooperative Adaptive Spectrum Sharing in Cognitive Radio Networks Haythem Bany Salameh Marwan Krunz Dept. of Telecomm. Eng. Dept. of Electrical and Computer Eng. Yarmouk Univ., Irbid, Jordan Univ. of Arizona, Tucson, AZ Email: haythem@ece.arizona.edu Email: krunz@ece.arizona.edu Abstract— The cognitive radio (CR) paradigm calls for open spectrum access according to a predetermined etiquette. Under this paradigm, CR nodes access the spectrum opportunistically by continuously monitoring the operating channels. A key challenge in this domain is how the nodes in a CR network (CRN) cooperate to access the medium in order to maximize the CRN throughput. Typical multi-channel MAC protocols assume that frequency channels are adjacent and that there are no constraints on the transmission power. However, a CRN may operate over a wide range of frequencies, and a power mask is often enforced on the transmission of a CR user to avoid corrupting the transmissions of spectrum-licensed primary-radio (PR) users. To avoid unnecessary blocking of CR transmissions, we propose a novel distance-dependent MAC protocol for CRNs. Our protocol, called DDMAC, attempts to maximize the CRN throughput. It uses a novel probabilistic channel assignment mechanism that exploits the dependence between the signal’s attenuation model and the transmission distance while considering the traffic profile. DDMAC allows a pair of CR users to communicate on a channel that may not be optimal from one user’s perspective, but that allows more concurrent transmissions to take place, especially under moderate and high traffic loads. Simulation results indicate that compared to typical multi-channel CSMA-based protocols, DDMAC reduces the blocking rate of CR requests by up to 30%, which consequently improves the network throughput. Index Terms— Cognitive radio networks, spectrum access, distance-awareness, traffic-awareness, MAC protocols. I. I NTRODUCTION Spectrum measurements by FCC and other organizations (e.g., XG DARPA initiative) indicate significant temporal and geographical variations in the utilization of the licensed spectrum, ranging from 15% to 85% [1]. These measurements motivated the need for a new technology that improves spectrum utilization without degrading the performance of licensed primary radio networks (PRNs). To cope with the rising demand in unlicensed wireless services, cognitive radio (CR) technology has been proposed. This technology allows an open access to the spectrum subject to a predetermined etiquette. In a cognitive radio network (CRN), users are aware of the radio frequencies used by existing legacy networks, and they opportunistically adapt their communication parameters to be able to communicate without affecting active PR users. A CRN has unique characteristics that distinguish it from traditional multi-channel wireless networks. Unlike traditional wireless networks, which typically occupy contiguous bands [2]–[4], a CRN is expected to operate over a set of widely-separated non-contiguous frequency bands. Communication on such bands exhibits different RF attenuation and interference behaviors. It is well known that signal attenuation An abridged version of this paper was presented at the IEEE SECON’08 Conference, June 2008. This research was supported in part by NSF (under grants CNS-0721935, CNS-0627118, CNS-0325979, and CNS-0313234), Raytheon, and Connection One (an I/UCRC NSF/industry/university consortium). Ossama Younis Applied Research, Telcordia Techn., Inc. Piscataway, NJ Email: oyounis@research.telcordia.com increases with the distance between the two communicating users and also with the carrier frequency used for communication [5]. Therefore, when assigning transmission channels in a CRN, it is necessary to consider the signal attenuation model and the interference conditions to improve spectrum utilization. Another characteristic of a CRN is that users must operate using a relatively low transmission power (i.e., abide by a power mask) to avoid degrading the performance of the PR users [4]. These peculiar characteristics call for new MAC protocols that efficiently utilize the available spectrum while improving the overall network throughput. A. Motivation Channel assignment mechanisms in traditional multi-channel wireless networks typically select the “best” channel, or set of channels, for a given transmission (e.g., [3], [6], [7]). In these mechanisms, the best channel is often defined as the one that supports the highest rate. We refer to this approach as the best multi-channel (BMC) approach. When the BMC approach is employed in a CRN, the blocking probability for CR transmissions, defined as the percentage of CR packet requests that are blocked due to the unavailability of a feasible channel assignment, can increase, leading to a reduction in the network throughput. To illustrate, consider an environment in which two PRNs and one CRN coexist. PRN 1 operates over a low-frequency band (CH1), while PRN 2 operates over a high-frequency band (CH2). Suppose that PRN 2 introduces a higher average PR-to-CR interference. Consequently, a CR receiver experiences a higher average signal-to-interferenceplus-noise ratio (SINR) over CH1 than over CH2. Assume that two CR users A and C need to send data to CR users B and D, respectively (see Figure 1). Also assume that the distance between A and B (dAB ) is less than that between C and D (dCD ). Figure 1(a) shows that when the CR users employ the BMC approach, the transmission A → B uses CH1, whereas the transmission C → D uses CH2. A → B is allowed to proceed because it operates over a low carrier-frequency channel with low PR-to-CR interference for a short transmission distance. On the other hand, C → D requires relatively higher transmission power to overcome the high attenuation associated with the high-frequency/high-interference channel and the long transmission distance. If the required transmission power exceeds the specified power mask, C → D cannot proceed. However, both A → B and C → D have much better chances of proceeding simultaneously if each CR transmitter selects channels while keeping in mind the constraining power mask of the other transmitter (Figure 1(b)). As a numerical example, assume that PRN 1 and PRN 2 operate in the 900 MHz and 2.4 GHz bands, respectively. Assume that dAB = 10 meters and dCD = 50 meters. Also assume that a CR transmission is successful if the received SINR over the selected channel is greater than the SINR threshold. For both channels, we set the SINR threshold and the interference mask to 5 dB and 60 mW, respectively. Assume that CR receivers B and D experience the same level of total interference over both channels (0.05 µW). Given the above parameters and using the propagation model in [8] with path loss exponent of 2, the required transmit powers over CH1 and CH2 for A → B are 2.2 mW and 16 mW, respectively. For C → D, these powers are 56.18 mW and 399.5 mW. According to the BMC scheme (Figure 1(a)), A → B can proceed over CH1 (the power mask is not violated), whereas C → D cannot proceed over CH2 (the required transmit power exceeds the power mask). On the other hand, if A → B uses CH2 and C → D uses CH1, both transmissions can proceed simultaneously (Figure 1(b)). C CH 2 D C CH 1 D X A B CH 1 (a) BMC channel assignment A B CH 2 (b) Distance-dependent channel assignment Fig. 1. Scenarios in which two CR transmissions can/cannot proceed simultaneously. It is worth mentioning that in a given (one-hop) neighborhood, the optimal channel assignment that maximizes the number of simultaneous CR transmissions can be formulated as an integer linear programming (ILP) problem [9], [10]. Since computing the optimal solution for the ILP problem grows exponentially with the size of the network [9], heuristic algorithms with suboptimal performance are needed. Such algorithms should attempt to compute channel assignment with reasonable computational/communication overhead. B. Contributions In this work, we develop a novel CSMA-based MAC protocol that aims at enhancing the throughput of the CRN subject to a power mask constraint. The proposed protocol (DDMAC) employs an intelligent stochastic channel assignment scheme that exploits the dependence between the RF signal attenuation model and the transmission distance while taking into consideration the local traffic conditions. The channel assignment scheme accounts for the interference conditions and the power constraints at different bands. In particular, the scheme assigns channels with lower average SINR to shorter transmission distances, and vice versa. In addition, our scheme associates more preferable channels to the most frequent transmission distances and less preferable channels to the less frequent distances. In other words, the assignment process identifies a “preferable” channel list for each CR user. Such a list indicates which channels are preferable to use depending on the estimated distance between the transmitter and the receiver. We propose two variants for the channel assignment scheme. The first variant is suitable for offline planning of spectrum sharing 2 in networks with known deployment and traffic patterns. In this case, there is no need for distance-traffic pattern prediction. The second variant is suitable for online dynamic network operation with unknown traffic patterns. To estimate the distance-traffic pattern in a given neighborhood, the second variant employs a stochastic learning technique that adapts to network dynamics (i.e., mobility, interference conditions, and traffic conditions). The primary advantage of our assignment scheme is that it is based on passive learning. This is because in DDMAC, CR users always listen to the control channel in order to overhear control-packet exchanges, including those not destined to them. CR users use the control information to identify the preferable channels. DDMAC has the following attractive features: • It does not make any assumptions about the activity patterns of the underlying networks or about user distribution. • It is easy to implement in practical settings and its processing overhead is small. • It is transparent to PR users, i.e., does not require coordination with them. • It inherently improves the fairness among CR users, compared to typical multi-channel CSMA-based protocols. • Under low load and several available channels, DDMAC gracefully degrades to the BMC approach. To evaluate the performance of DDMAC, we conduct simulations over a dynamic CRN with mobile users. Our simulation results show that by being distance- and traffic-aware, DDMAC significantly improves network throughput while preserving fairness. The results also indicate that compared with typical multi-channel CSMA-based protocols, DDMAC decreases the connection blocking rate in a CRN by up to 30%. By injecting artificial errors into the estimated distances, our evaluation reveals that DDMAC is robust against estimation errors. It should be noted that selecting a preferable channel list was also proposed in the MMAC protocol [11]. However, MMAC does not support multiple-channel assignment (it is limited to one channel per user). Specifically, the channel selection criterion in MMAC is to use a channel with the lowest count of source-destination pairs that have selected the channel. In DDMAC, the preferable channel list per node is constructed by accounting for the challenges associated with CRs (i.e., low transmit power, presence of PR users, widely-separated noncontiguous available bands). Unlike DDMAC, the objective in MMAC was not to address spectrum sharing while improving the overall throughput, but rather to handle multi-channel hidden terminals using a single transceiver and to balance the channel usage over all available channels. In addition, MMAC requires global network synchronization, which is not a requirement in DDMAC. C. Organization The rest of this paper is organized as follows. Section II gives an overview of related work. In Section III-A, we introduce our system model and state the main assumptions. The SINR analysis is presented in Section III-B. Section III-C illustrates the effect of the carrier frequency and transmission distance on the path loss. In Section IV, we formulate the optimal channel assignment problem. Section V introduces our proposed distance- and traffic-aware channel assignment algorithm. Section VI describes the proposed DDMAC protocol and outlines its benefits and associated overhead. We evaluate DDMAC in Section VII. Finally, Section VIII gives concluding remarks. II. R ELATED W ORK Recently, several attempts were made to develop MAC protocols for CRNs (e.g., [6], [12]–[16]). In [6], the authors developed a CRN MAC protocol with a common control channel. This protocol jointly optimizes the channel/power/rate assignment, assuming a given power mask on CR transmissions. DC-MAC [12] is a cross-layer distributed scheme for spectrum allocation/sensing. It provides an optimization framework based on partially observable Markov decision processes, assuming that PR and CR users share the same slotted transmission structure. In [13], the authors investigated continuous-time Markov models for dynamic spectrum access in open spectrum wireless networks. Using such models, a distributed random access protocol is proposed to achieve airtime fairness between dissimilar unlicensed users. The FCC defined the interference temperature model [17], which provides a metric for measuring the interference experienced by licensed receivers. In [14], the authors studied the issue of spectrum sharing among a group of spreadspectrum users subject to constrains on the SINR and on the interference temperature. In [18], the interference temperature model was used for optimal selection of spectrum and transmission powers for CR users. In [16], the authors proposed a decentralized channel-sharing mechanism for CRNs based on a game-theoretic approach for both cooperative and noncooperative scenarios. In [19], the concept of a time-spectrum block is introduced to model spectrum reservation in a CRN. Based on this concept, the authors presented centralized and distributed CRN protocols with a common control channel for spectrum allocation. The above protocols were designed without exploiting the dependence of the number of allowable CR transmissions on the carrier frequency and the transmission distance. They are limited to the analytical aspects of MAC design, with no complete operational details. To the best of our knowledge, DDMAC is the first CRN MAC protocol that aims at improving the CRN throughput by exploiting the dependence on the RF signal’s attenuation model and the transmission distance while considering the prevailing traffic and interference conditions. III. P RELIMINARIES Users in PRN1 CR CR Fig. 2. Users in PRN2 CR Example of an opportunistic CRN that coexists with two PRNs. A. Network Model We consider a CRN with decentralized control (i.e., an ad hoc network). This CRN coexists geographically with M different PRNs. PR users are legacy radios that cannot be controlled by the CRN. Figure 2 shows a conceptual view of the scenario under consideration with M = 2. The PRNs are licensed to 3 Fig. 3. Operating spectrum in the hybrid network. operate over non-overlapping frequency bands. We assume that all the PRN bands have the same bandwidth (BW ). In reality, a PRN may occupy multiple, non-contiguous, frequency bands. Such a PRN can be easily represented in our setup by using multiple equal-bandwidth virtual PRNs, each operating over its own carrier frequency. For the ith PRN, we denote its carrier frequency by fi . As shown in Figure 3, the available bandwidth (BW ) of a PRN is divided into L adjacent but non-overlapping frequency channels each of Fourier bandwidth W (in Hz). Such L channels are collectively referred to as a band. Let N denote the total number of channels in all bands; N = LM . Without loss of generality, we assume that BW is sufficient to support at least one CR transmission. This is an acceptable assumption in many wireless systems that are built to operate in the unlicensed bands, including IEEE 802.11/a/b/g-compliant devices. Each CR user is equipped with nt radio transceivers, 1 ≤ nt ≤ L, that can be used simultaneously. In theory, a CR user can transmit over an arbitrary segment of the available bandwidth by using tunable filters. In practice, however, a CR typically implements a bank of fixed filters, each tuned to a given carrier frequency with fixed bandwidth, allowing the CR user to choose from a fixed number of channels. In our setup, we assume the latter (more practical) capability, which can be used to approximate the tunable filter scenario. To avoid corrupting the transmissions of licensed users, a mask is enforced on the transmission power of a CR user (i) (i) over each band, i.e, Pt ≤ Pmask , i = 1, 2, . . . , M . The determination of an appropriate power mask is an important topic, which has been investigated under certain simplifying assumptions (e.g., [18], [20]). The spectrum sharing protocols in [18] and [20] were designed such that the maximum transmission powers of CR users over various bands are dynamically computed based on the PR’s interference margins (set by the FCC) and local traffic conditions. In [20], the authors provided a neighborhood-dependent adaptive power mask on CR transmissions that ensures a statistical (soft) guarantee of the outage probability of PRNs (the probability that the total interference power at a PR receiver exceeds the maximum tolerable interference). The authors provided closed-form expressions for the resulting power mask. For our purposes, we assume that a similar mechanism for determining the power mask is in place. A CR user transmits data to other CR users using the maximum allowable power vector Pmask . When not transmitting, a CR user is capable of measuring the total noise-plus-interference I (i) over all bands i = 1, 2, . . . , M 1 . This requires a wideband sensing capability with a narrowband resolution. The technology to support such 1 The quantity I (i) includes the PR-to-CR interference as well as the thermal noise. M L N, N BW , W nt Pt (i) (i) Pmask I (i) (i) SINRj PL (fi ) Pr (fi ) d J (i) cj Cj µ∗ i mj Mj R Rc , rc Ri ri , ri−1 Di di , di−1 m Twin pi (t) p ei (t) α Ωi (A), K CCL(A, B) Φ(A, B) Number of PRNs Number of channels in a PRN N is the total number of channels, N = {1, . . . , N } Bandwidth of a band and a channel, respectively Number of transceivers per CR user CR transmit power Interference power mask on channel i Noise-plus-interference over channel i Measured SINR over band i at receiver j Path loss associated with band i Received power at a CR receiver over band i Transmitter-receiver distance Set of all CR transmission requests in a locality ith selected channel’s data rate for transmission j Rate demand of the jth CR transmission SINR threshold over channel i Number of selected channels for the jth transmission Mj is the set of mj selected channels for the jth transmission Random variable represents the distance to the intended receiver RC is the transmission region, Rc = πrc2 ith ring around a CR in static channel assignment Radii that define Ri , i = 1, . . . M ith ring around a CR in dynamic channel assignment Radii that define Di Number of non-overlapping Di rings Observation window time Probability of Di at time t Weighted average of pi (t) Forgetting factor CR A’s preferable channel list for region i, i = 1, ..., K Common channel list available for A → B transmission Preferable available channels for A → B transmission TABLE I S UMMARY O F N OTATIONS U SED I N T HE PAPER . capability is readily available through a wideband antenna, a power amplifier, adaptive filters, and a DSP technique called cyclostationary feature detection [21], [22]. Thus, a CR user can simultaneously sense several GHz-wide bands and estimate the instantaneous interference over each band [22]. Alternatively, a sequential partial sensing approach can be employed at the cost of negligible switching/sensing overhead [21], [23]. It is worth mentioning that off-the-shelf wireless cards can readily serve as a fully functional wideband multi-channel CR interface. Such an interface enables a CR user to perform analysis of the RF spectrum (i.e., sensing) in real time. C. Carrier Frequency and Distance Effects on Path Loss In this section, we discuss the effect of the carrier frequency and transmission distance on the path loss. For a given carrier frequency f , let do (f ) be the close-in distance, i.e., the distance from the transmitter after which the RF channel can be approximated by the free-space model; do (f ) can be determined from measurements or can be estimated by [8]:  2  2Da f c do (f ) = max (2) , Da , c f where Da is the antenna length of the transmitter and c is the speed of light. Let Po (f ) and Pt (f ) respectively denote the received power at the close-in distance and the CR transmit power. Then, Po (f ) can be estimated as follows [8]: Po (f ) = c2 Gt (f )Gr (f ) Pt (f ) (4πdo (f ))2 f 2 where n is the path loss exponent (typically, 2 ≤ n ≤ 6). Note that, in practice, do (f ) is of the same order of magnitude as the node’s dimensions. For example, for a mobile phone operating in the 900 MHz band with Da = 5 cm, do (f ) = 33 cm. For an 802.11 WLAN card operating in the 2.4 GHz band and the same antenna size, do (f ) = 12 cm. Accordingly, it is reasonable to assume that the probability that d is less than do (f ) is very small (i.e., Pr(d < do (f )) ≈ 0). Using (2), (3), and (4), the path loss PL (f ) can be expressed as: Pt (f ) PL (f ) = 10 log = −10 × Pr (f )  o n 2 n−2 c2 γDa c 2Da f  , , ∀ f s.t. D ≥ max log  2 n a f d c o  n f  n 2D2 f ∀ f s.t. fc ≥ max Da , ca log fcn dγn , (5)  o n n−2  4−n 2 2  c γ(2Da ) 2D f  log , ∀ f s.t. ca ≥ max Da , fc f 4−n dn def γ= Based on the aforementioned characteristics of the CRN, the average measured SINR (SINR) at a CR receiver at a given time over band i is mainly determined by: (1) the path loss associated with that band (PL (fi )); (2) the average interference over that (i) band (I ), which can be estimated based on the sensing history and the spectrum occupancy statistics (e.g., using the techniques (i) in [12], [24]; and (3) the enforced power mask Pmask . Formally, (i) SINR (dB) is given by: (i) (i) SINR (dB) = Pmask (dB) − PL (fi )(dB) − I (i) (dB). (1) Note that in [25] and [26], it was shown that for a given CRN and due to PRN’s activity, CR users that are far away from each (i) other can experience different average interference I , which may vary with time. On the other hand, CR users in close proximity typically share the same view of the surrounding RF environment. Table I summarizes the main notation used in the paper. (3) where Gt (f ) and Gr (f ) are the transmit and receive antenna gains, respectively. Let Pr (f ) denote the received power at distance d from the transmitter, d ≥ do (f ). Then, n  do (f ) (4) Pr (f ) = Po (f ) d where B. Analysis of the Average SINR 4 Gt (f )Gr (f ) . (4π)2 (6) Note that the dependence of PL (f ) on d (i.e., d1n ) is the same for any given carrier frequency. Figure 4 depicts the path loss for a wide range of carrier frequencies and two values of n at a distance d = 1 meter. This figure and equation (5) reveal that the signal attenuation increases as the distance between the two communicating users increases, and as the frequency used for communication increases. These observations provide the motivation for our distance-dependant channel assignment, discussed in Section V. IV. O PTIMAL C HANNEL A SSIGNMENT P ROBLEM Our objective is to maximize the number of simultaneous CR transmissions, and consequently the overall network throughput. Toward this end, we define the term local spectrum utilization as the total number of simultaneous CR transmissions that can be supported in a given (one-hop) locality while meeting a predefined power mask. Before formulating the problem, we discuss the requirements for a successful CR transmission. 5 demands are available, the above ILP problem belongs to the class of NP-hard problems [9]. In this paper, we develop a heuristic channel assignment scheme that provides a suboptimal solution with low complexity and good spectrum utilization. Our heuristic exploits distance and traffic awareness. The key idea behind it is to assign channels with low SINR to shortdistance transmissions. Also, local traffic information is used to assign more channels to more likely transmission distances. 70 Path Loss (dB) 60 50 40 30 n=4 n=2 20 0 1 2 3 f (Hz) 4 5 6 9 x 10 Fig. 4. Path loss vs. carrier frequency for two path loss exponents (Da = 5 cm, Gt (f ) = Gr (f ) = 1). A. CRN Transmission Requirements Within a given neighborhood, multiple CR users may contend for access to one or more of the available channels. Let N and J denote the set of all N channels and the set of all CR transmission requests in the local neighborhood at a given time, respectively. We assume that the jth CR transmission (j ∈ J ) is successful if both of the following two conditions are met: • It is possible to find mj available channels from the set Pmj (i) (i) cj ≥ Cj , where cj is the data rate of N such that i=1 the ith selected channel and Cj is the total rate demand for the jth CR transmission. • Let Mj be the set of mj selected channels. Then, the (i) received SINR of every i ∈ Mj (SINRj ) must be greater ∗ than the SINR threshold (µi ) that is required at the CR receiver to achieve a target bit error rate over channel i. B. Maximizing the Utilization of Local Spectrum (i) Let δj be a binary variable denoting whether or not channel i is assigned for transmission j. Formally,  1, if channel i is assigned for transmission j (i) δj = (7) 0, otherwise. Similar to [10], [27], the problem of maximizing the total number of simultaneous CR transmissions in a given neighborhood can be formally stated as follows: hP i P (i) (i) maxδ(i) ∈{0,1} j∈J 1 (8) ≥ C c δ j j i∈N j j P (i) (9) j∈J δj ≤ 1, ∀i ∈ N P (i) (10) i∈N δj ≤ nt , ∀j ∈ J (i) (i) SINRj ≥ µ∗i , ∀j ∈ J , s.t. δj = 1 (11) where 1[.] is the indicator function. The constraint in (9) ensures that a channel cannot be assigned to more than one CR transmission in the same vicinity. The constraint in (10) ensures that at most nt channels can be assigned to a CR transmission. For an ad hoc CRN, the above optimization problem must run in a distributed manner at each CR user in the network. This implies that each CR user must exchange instantaneous SINR and rate demand information with neighboring CR users before selecting channels, which incurs high control overhead and delay (i.e., information may not be up-to-date). Even if perfect knowledge of the SINR of each link and the rate V. D ISTANCE -D EPENDENT C HANNEL A SSIGNMENT A LGORITHM In this section, we describe our proposed channel assignment mechanism. The assignment process identifies a “preferable” channel list for each CR user. Such a list indicates which channels are preferable to use depending on the estimated distance between the transmitter and the receiver. It is worth mentioning that many techniques for estimating the transmitterreceiver distance in wireless networks have been proposed in the literature, including the Received Signal Strength Indicator (RSSI), the Time of Arrival (ToA), and the Time Difference of Arrival (TDoA) [28]. For our purposes, any of these schemes can be used. In Section VII, we investigate the robustness of our scheme under inaccurate distance estimation, which is mainly caused by mobility, multi-path propagation, reflection, and fading effects. Two variants of the channel assignment mechanism are proposed. The first variant is suitable for offline planning of spectrum sharing in networks with known traffic patterns, whereas the second variant is for online spectrum allocation in dynamic (mobile) networks with unknown traffic patterns. A. Spectrum Assignment for Known Traffic Profiles Given a CR user with a packet to transmit, let r be the estimated distance to the intended receiver; r ≤ rc , where rc is the maximum transmission range. rc represents the largest distance from a CR transmitter over which the transmission at maximum power can be correctly decoded over all selected channels in the absence of interference from other terminals def (CR or PR users). Let FR (r) = Pr{R ≤ r}. The functional form of FR depends on both node distribution as well as the distance traffic profile, which for now we assume to be given. Given FR , the channel assignment process is conducted as follows: • The available bands are divided according to their measured SINR (given in (1))2 into M sets S1 , S2 , . . . , SM , where each band consists of multiple channels. The set S1 contains the frequency channels of the band that has the highest SINR, S2 contains the next highest SINR, and so on. • A CR user, say A, divides its maximum transmission def region Rc = πrc2 into M non-overlapping “rings” R1 , . . . , RM . The ith ring contains the CR users whose distances to A fall in (ri−1 , ri ], where i = 1, . . . , M and 0 = r0 ≤ r1 ≤ r2 ≤ . . . ≤ rM = rc . The rings are divided such that the probability of communicating with a CR receiver that falls within any of the M rings is the same, i.e., 1 FR (ri ) − FR (ri−1 ) = , i = 1, . . . , M. (12) M 2 Note that P ’s dependence on d is the same for all bands. Thus, for the L purpose of SINR comparison, we set d = 1 meter. User A computes the radii ri , i = 1, . . . M , by substituting for FR (ri ) in (12) and solving for ri . • Finally, A constructs a preferable channel list for each ring by assigning channels with lower SINR to shorter transmission distances and channels with higher SINR to longer transmission distances, i.e., assign SM to R1 , SM−1 to R2 , . . ., and S1 to RM . To illustrate the idea, we consider a uniformly distributed CRN and assume that a CR transmitter randomly chooses a destination for its data from within Rc . Therefore, FR (r) is given by: ( 2 r rc2 , r ≤ rc . FR (r) = (13) 1, r ≥ rc Using (12) and (13), and noting that r0 = 0, we arrive at the following expression for ri : s r  2 ri−1 i 1 ri = rc = + 2 rc . (14) M rc M Figure 5 illustrates the non-overlapping rings around a CR transmitter when M = 4. Within these rings, other CR and PR users may exist. Assume rc = 100 meters. Then, r1 , . . . , r4 are given by 50, 70.71, 86.6, 100 meters, respectively. 6 distances fall in the range (di−1 , di ] (how to convey transmitterreceiver distance information will be discussed later). Note that the proper setting of Twin depends on the dynamics of the network. The effect of Twin is studied in Section VII. Fig. 6. Time diagram of pmf’s updating process. To initialize the assignment algorithm, all CR users employ the BMC scheme discussed in Section I. At any time t, CR user A constructs its transmission distance table based on control packets it overheard during the observation window [t−Twin , t]. Using the transmission distance table, A estimates the current probability mass function pi (t) of the distance r at time t (see Figure 6). It then computes an exponentially weighted average of pi (t) : pei (t) = αpi (t) + (1 − α)pei (t − Twin ), Fig. 5. Four regions around a CR transmitter for assigning channels. B. Spectrum Assignment for Unknown Traffic Profiles For offline spectrum planning, we assumed in the previous section a fixed network and prior knowledge of the distancetraffic pattern (i.e., the form of FR ). During network operation, however, the distance-traffic pattern may change with time, depending on network dynamics and user mobility. Because users only possess local knowledge of their neighborhoods, it is difficult to maintain the optimal network performance. Nevertheless, we can develop a stochastic learning algorithm that performs well and uses only localized information. Stochastic learning techniques have been widely used in wireless networks for online traffic prediction, tracking, and power control [29], [30]. Our proposed learning approach is a distributed algorithm that runs at each CR user in the network. A CR user, say A, evenly divides its maximum transmission region Rc into m nonoverlapping regions, where m ≫ M . The ith region, Di , forms a ring, defined by the area {(x, y) : d2i−1 < x2 + y 2 ≤ d2i }, where di = i rmc , and di−1 < di i = 1, . . . , m. CR user A maintains an m-entry transmission distance table. The ith entry in that table corresponds to the region Di , and contains the number of overheard CR packet requests during the recent observation window Twin for which the transmitter-receiver (15) where α is a forgetting factor, 0 < α ≤ 1. Once pei (t) is computed, A computes the preferable channel list for each ring. Let Ωi (A) denote the preferable channel list for ring Di at CR user A (how to construct Ωi (A) will be given later). The new preferable channel lists will be used during the next observation window time. The proposed channel assignment process merges the Di ’s into K regions according to pi (t), where K ≤ M . It then assigns preferable channels for each region. The process is now described in detail: Pk−1 1) User Pm A determines the integer k such that | i=0 pei (t)− ei (t)| is minimized, i.e., it divides the regions i=k p into two groups; short-distance and long-distance groups. The probabilities of the short-distance and long-distance groups are given by: Pshort = k−1 X pei (t) (16) pei (t). (17) i=0 and Plong = m X i=k 2) User A divides the M bands into two frequency sets: low SINR frequency set and high SINR frequency set. It assigns the low SINR frequency set to the short-distance group and the high SINR frequency set to the longdistance group. The numbers of bands in the high (nH ) and low (nL ) frequency sets depend on Pshort and Plong , as follows:   Pshort nH = M Pshort + Plong nL = M − nH (18) where ⌈.⌉ is the ceiling function. 3) Step 1 and 2 are repeated for every group until either only one band is assigned to that group or the group contains only one region. Note that when repeating the above process for a group, m in (17) and M in (18) are replaced by the number of regions in that group and the number of channels assigned to that group, respectively. Using this recursive procedure, the preferable channel list Ωi (A), for all i, is computed for one observation window. C. Complexity Claim 1: The worst-case complexity for selecting the preferable channel list Ωi (A), for all i, may be obtained using the above recursive procedure in O(mK) time, where K ≈ min[N, m]. Proof: In the worst case, our proposed algorithm requires O(m) comparisons to perform one iteration (steps 1 and 2). In addition, it requires at most K = min[N −1, m−1] iterations to obtain Ωi (A), for all i. Hence, Ωi (A), for all i, may be obtained using the proposed algorithm with a complexity of O(mK), where K ≈ (m min[N, m]). For N ≥ m, K ∼ O(m). On the other hand, for N < m, K ∼ O(N ). D. Illustrative Examples We illustrate the previously discussed channel assignment process using the following examples. 1) Example 1: Consider four PRNs and one CRN. Each PRN occupies two adjacent non-overlapping channels. The PRNs are labeled such that f1 < f2 < f3 < f4 . Consider (1) (2) (3) a CR user A with SINR > SINR > SINR > (4) SINR . Suppose that A divides its transmission region Rc into 8 rings, D1 , D2 , . . . , D8 . At a given time t, assume that the weighted average pmf {pei (t) : i = 1, . . . , 8} is given by {0.25, 0.1, 0.15, 0.05, 0.05, 0.15, 0.05, 0.2}. Figure 7 shows how the proposed channel assignment process is conducted. The outcome of this process is as follows: • Band 4, which includes two channels, is assigned to all CR transmissions whose distances are in D1 (i.e., Ω1 (A) = {4}). • Band 3, which includes two channels, is assigned to all CR transmissions whose distances are in D2 and D3 (i.e., Ω2 (A) = Ω3 (A) = {3}). • Band 2, which includes two channels, is assigned to all CR transmissions whose distances are in D4 , D5 , and D6 (i.e., Ω4 (A) = Ω5 (A) = Ω6 (A) = {2}). • Band 1, which includes two channels, is assigned to all CR transmissions whose distances are in D7 and D8 (i.e., Ω7 (A) = Ω8 (A) = {1}). Di 1 2 3 4 5 6 7 8 p~ i 0.25 0.1 0.15 0.05 0.05 0.15 0.05 0.2 {f4, f3} 0.5 {f4} {f3} {f2} {f2, f1} 0.5 {f1} 0.25 0.25 0.25 0.25 Fig. 7. Example that illustrates the channel assignment process in a dynamic CRN. 7 2) Example 2: Consider 8 PRNs and one CRN. The PRNs are labeled such that f1 < f2 < . . . < f8 . Suppose that A divides its transmission region into 2 rings. At a given time t, assume that the weighted average pmf {pei (t) : i = 1, 2} is given by {0.25, 0.75}. Then, the outcome of our preferable channel assignment is as follows: • Channels 1 and 2 (total of 2 channels are assigned to all CR transmissions whose distances are in D1 ). • Channels 3, . . . , 8 (total of 6 channels are assigned to all CR transmissions whose distances are in D2 ). The above example reveals that our algorithm assigns more preferable channels (total of 6 channels) to the more frequently used transmission distances (D2 , pe2 (t) = 0.75). VI. DDMAC P ROTOCOL Based on the channel assignment process presented in Section V, we now propose a distributed, asynchronous MAC protocol for CRNs. The proposed DDMAC is a CSMA/CAbased scheme that uses contention-based handshaking for exchanging control information. It is worth mentioning that the most common configuration for upcoming CRNs is to use CSMA/CA-like MAC access [6], [19], [20], [23], [25], [26], [31]. Thus, in designing the channel access in DDMAC, we focus on extending the CSMA/CA scheme due to its maturity and wide deployment in many wireless packet networks. Note that the handshaking procedure is essential in multi-channel systems. Besides mitigating the hidden-terminal problems, there are two other main objectives for the use of RTS/CTS: (1) conducting and announcing the channel assignment, and (2) prompting both the transmitter and the receiver to tune to the agreed on channels before transmission commences. Before describing our protocol in detail, we first state our main assumptions. A. Assumptions In designing DDMAC, we make the following assumptions: • For each frequency channel, the channel gain is stationary for the duration of three control packets and one data and ACK packet transmission periods. As explained in [32], this assumption holds for typical mobility patterns and transmission rates. • Channel gains between two CR users are symmetric. This is a typical assumption in any RTS/CTS-based protocol, including the IEEE 802.11 scheme. • CR transmissions use the maximum allowable power vector (Pmask ). The key idea behind this choice is as follows. It is well-known that using as many channels as possible for a transmission reduces the CR-to-PR interference [6] due to the reduction in transmission power. However, because DDMAC enforces an exclusive channel occupancy, which prevents two neighboring CR users from using common channels3, such a channel assignment policy may lead to channel over-assignment, which reduces the opportunity for finding available channels by other neighboring CR transmitters (thus reducing the CRN’s throughput). Therefore, in DDMAC, we tackled the CR-toPR interference problem by assuming a given power mask to protect PR users while trying to use the least possible number of selected channels per transmission. This can be 3 The exclusive channel occupancy excludes CR-to-CR interference although it still allows for the typical co-channel PR-to-CR interference, thus largely simplifying the CR-to-PR interference management process. • • • done by transmitting at the highest possible transmission power over each selected channel, which results in less number of assigned channels per CR transmission. This increases the opportunity for finding available channels by other neighboring CR transmitters. The total rate demand of a CR user A (denoted by CA ) is met by aggregating the transmission rates of several selected channels. Note that CA can vary from one packet to another. A prespecified control channel with Fourier bandwidth Bc is available, where Bc ≪ B. This channel does not need to be reserved for the CRN. It can, for example, be one of the subchannels in an ISM band. Contending CR users follow similar interframe spacings and collision avoidance strategies of the 802.11 protocol (over the control channel) by using physical-carrier sensing and backoff before initiating control packet exchanges. We also assume that data packet sizes are significantly larger than control packets, and therefore, the use of the RTS/CTS handshake is justified. B. Channel Access in DDMAC The channel access mechanism allows the CR transmitter and receiver to agree on the set of channels to use for communication and to allocate their rates. Rate is allocated in a manner that ensures that the power mask and the rate demands are met. A CR user A views its transmission region as K non-overlapping regions, where each region is associated with a preferable channel list Ωi (A), i = 1, . . . , K, determined according to Section V. This user maintains an N -entry channel list and an m-entry transmission distance table (as described in Section V). The jth entry of the channel list indicates the status of the jth channel; 1 if the channel is available and 0 if the channel is occupied or reserved by any of A’s CR neighbors. Recall that each CR user is equipped with nt transceivers. One of these transceivers is tuned to the control channel, while the other nt − 1 transceivers can be tuned to any data channels. As a result, CR users can always hear control messages over the common control channel even when they are transmitting/receiving data over other data channels. Thus, every CR user listens to the control channel, and accordingly updates its channel list and transmission distance table. Suppose that CR user A has data to transmit to another CR user B at an aggregate rate demand CA . Then, A reacts as follows: • If user A does not sense a carrier over the control channel for a random duration of time, it sends an RTS message at the maximum (known) power Pmax . This Pmax is constrained by the power mask imposed on the prespecified control channel. The RTS includes CA , the packet size (in bytes), and the list of all available channels at A (see Figure 8). • The neighbors of A (other than B) that can correctly decode the RTS refrain from accessing the control channel until they receive one of two possible control packets, denoted by EPCA and ENCA (explained below). • Upon receiving the RTS packet, B estimates the distance between A and B (dAB ) (using one of the techniques described in Section V). It identifies the preferable channel list Ωi (B) that corresponds to dAB . Based on the available channels at A and B, and the instantaneous interference level over these channels as measured at B, user B removes any channel that has a received SINR less than its threshold SINR and determines the common channel 8 RTS Transmitter Receiver ID ID Packet size Rate demand (CA) Available channel list PCA/ EPCA Transmitter Receiver ID ID TX duration Distance (d) Assigned channel list NCA/ ENCA Transmitter Receiver ID ID Distance (d) Fig. 8. • • • Formats of DDMAC control packets. list that is potentially available for A → B transmission, denoted by CCL(A, B). User B then computes the intersection between Ωi (B) and CCL(A, B) to identify a preferable set of channels for A → B (Φ(A, B)). To achieve good throughput, B sorts the channels in Φ(A, B) in a descending order of their maximum possible data rate (calculated according to Shannon’s formula4). Then, user B appends the rest ofthe common available channels  that T are not in Φ(A, B) i.e., CCL(A, B) Φ(A, B) , also listed in a descending order of their maximum possible data rate, to the bottom of the sorted preferable channels. User B cumulatively adds channels from the top of the new sorted list until either the aggregate rate CA is satisfied or the list is exhausted, i.e., no feasible channel assignment is found. If there is no feasible channel assignment, then B responds by sending a Negative-Channel-Assignment (NCA) message that includes the distance dAB (see Figure 8). The purpose of this packet is to help B’s neighbors estimate the network distance-traffic pattern and prompt A to back off and retransmit later. If B can find a set of available channels that can support a total demand CA , it sends a Positive-Channel-Assignment (PCA) message to A, which contains the assigned channels for the transmission A → B, the distance dAB , and the duration needed to hold the assigned channels for the ensuing data transmission and corresponding ACK packet. The PCA packet implicitly instructs B’s CR neighbors to mark the set of assigned channels as unavailable for the indicated transmission duration. It also helps these neighbors estimate the network distance-traffic pattern. Depending on which control message is received, user A reacts as follows: – If A receives an NCA message, it responds by sending an Echo-NCA (ENCA) message, which includes the distance dAB . The purpose of this packet is to help A’s neighbors estimate the network distance-traffic pattern. – If A receives a PCA message, it replies back with an Echo-PCA (EPCA) message, informing its neighbors of the selected channel list, the distance dAB , and the transmission duration. This EPCA also announces the success of the control packet exchange between A and B to A’s neighbors, which may not have heard B’s PCA. Once the RTS-PCA-EPCA exchange is completed, the data transmission A → B proceeds. Once completed, B 4 Other rate-vs-SINR relationships, such as a staircase function, can be used for calculating the achievable data rates. 9 sends back an ACK packet to A over the best assigned channel, i.e., the channel that has the highest rate. A time diagram of the RTS-PCA-EPCA-DATA-ACK exchange is depicted in Figure 9. * * CR user * * * * Users in PRN 1 Users in PRN 2 C A BA R P T C S A CD C E P C A R P T C S A E P C A * * * * ……….. * * C + * * * D A ACK Data TX 1 D * A DATA A->B B * B * * * ACK DATA C->D Data TX 2 * * . . . Fig. 9. RTS-PCA-EPCA-DATA-ACK packet exchange. It is worth mentioning that there is no interference between data and control packet transmissions because the two are separated in frequency. Therefore, a CR user that hears the RTS packet from A defers its attempt to access the control channel until it receives an EPCA or an ENCA packet from A. In addition, a CR user that receives only a PCA or an NCA should defer its attempt to access the control channel for the expected time of the EPCA/ENCA packet (to avoid a collision between control packets). This allows for more parallel transmissions to take place in the same neighborhood (see Figure 9). Remark: DDMAC’s channel assignment is performed on a per-packet basis, with the channels assigned to different interfaces dynamically changing. This type of channel assignment requires channel switching to occur at a very small time scale, which is in the range of micro-seconds5. (a) Allowed channel reuse Fig. 10. Scenarios in which a CR transmitter C can/cannot reuse the channels assigned to A. Solid circles indicate data-transmission ranges, whereas dashed circles indicate control-transmission ranges. • radio technology allows channel switching to be done in a few microseconds (i.e., < 10 µs [23], [33]) = φ : i = 1, . . . , m − 1. = N : i = m. Recall that N denotes the set of available channels. In other words, no channels will be assigned to ring, i, i = 1, 2, . . . , m − 1, and all channels will be assigned to the mth ring. rc A d1 rc d m-1 A d1 d m-1 (a) Scenario I Fig. 11. • 5 Current Scenario I: At a given time t, assume that the weighted average pmf {pei (t) : i = 1, . . . , m} has a value of 1 at i = m and 0 otherwise (i.e., most likely, transmission distances are within Dm ). This scenario represents the case when all of A’s neighbors are located near the border of A’s transmission range (Figure 11(a)). According to the channel assignment algorithm, the preferable channel list is identified as follows: Ωi (A) Ωi (A) C. Spatial Reuse and DDMAC We consider a CSMA/CA-based multi-hop CRN environment, which consists of multiple contention regions (neighborhoods) that permit spatial reuse. Specifically, non-neighboring CR users may access the same channel on different contention domains. To illustrate the idea of spatial reuse, Figure 10 depicts two scenarios for the operation of DDMAC. In the first scenario (Figure 10(a)), the two transmitters A and C cannot hear each other’s control packets. So, according to CSMA/CA, the transmissions A → B and C → D can overlap in their data channels, i.e., the assigned channels for A → B transmission are reserved only within the area of A’s and B’s control range (spatial reuse case). In Figure 10(b), node C falls in the control region of node A (and vice versa). The exclusive channel occupancy policy prevents A and C from using common channels. However, the two transmissions can proceed simultaneously if A and C can find two nonintersecting sets of channels to support their rates. D. Worst-Case Scenarios for DDMAC We illustrate two extreme scenarios under which the DDMAC protocol gracefully degrades into the BMC scheme. Recall that a CR receiver A divides its transmission range into m regions. (b) Unallowed channel reuse (b) Scenario II Illustration of two worst-case scenarios in DDMAC. Scenario II: At a given time t, assume that the weighted average pmf {pei (t) : i = 1, . . . , m} has a value of 1 at i = 1 and 0 otherwise (i.e., most likely, transmission distances are within D1 ). This scenario represents the case where all A’s neighbors are located close to A (Figure 11(b)). According to the proposed channel assignment algorithm, the preferable channel list is identified as follows: Ωi (A) Ωi (A) = = N : i = 1. φ : i = 2, . . . , m. According to DDMAC, the sorted channel list from which a CR user assigns channels to its transmission is constructed by appending the common sorted available channels that are not in the sorted preferable channels to the bottom of the sorted preferable channels list. Thus, for the above two scenarios and depending on the transmitter-receiver distance, the sorted channel list of DDMAC is as follows: • • If the distance falls in the range (dm−1 , dm = rc ] or (0, d1 ], the preferable channel list is the set of all available channels. Therefore, the sorted channel list of DDMAC is the same as that of the BMC scheme. Consequently, DDMAC gracefully degrades into the BMC scheme. If the distance falls within the transmission range Rc but not in the range (dm−1 , dm ] or (0, d1 ], the preferable channel list is empty whereas the available channel list contains all the available common channels. Therefore, the sorted channel list of DDMAC is the same as that of BMC. Consequently, DDMAC gracefully degrades into the BMC scheme. Protocol Overhead Claim 2: DDMAC and BMC have comparable overheads. Proof: Both DDMAC and BMC use a three-way handshake to send one data packet. Thus, DDMAC does not introduce any additional control message overhead. VII. P ROTOCOL E VALUATION We now evaluate the performance of the DDMAC via simulations and compare it with CSMA/CA variants. Our results are based on simulation experiments conducted using CSIM (a C-based, process-oriented, discrete-event simulation package [34]). Each CR user generates packets according to a Poisson process with rate λ (in packet/time slot), which is the same for all users. For simplicity, data packets are assumed to be of a fixed size (2 Kbytes). Each CR user requires an aggregate transmission rate of 5 Mbps. We divide time into slots, each of length 3.3 ms. A time slot corresponds to the transmission of one CR packet at a rate of 5 Mbps. We set the CRN SINR threshold to 5 dB and the thermal noise to (i) Pth = 10−21 Watt/Hz for all channels. Because DDMAC and the compared with CSMA/CA-based protocols have the same maximum transmission ranges and use the same channel access mechanism, it is reasonable to assume that all protocols achieve the same forward progress per hop. Consequently, our performance metrics are: (1) one-hop throughput, i.e., the destination of a packet is restricted to one hop from the source, (2) connection blocking rate, and (3) the fairness index [35]. The connection blocking rate is defined as the percentage of CR packet requests that are blocked due to the unavailability of a feasible channel assignment. We use Jain’s fairness index [35] to quantify the throughput fairness of a scheme6 . Fairness index values closer to 1 indicate better fairness. The signal propagation model in (4) is used with n = 4, the antenna length (D) is 5 cm, and Gt (f ) = Gr (f ) = 1 for every carrier frequency f . A. Single-hop Scenarios 10 1) Simulation Setup: We first simulate a small-scale network for the purpose of highlighting the advantages and operational details of DDMAC. DDMAC is compared with three multichannel CSMA-based protocols that use different channel selection schemes: an optimal scheme (which uses exhaustive search), the BMC scheme [3] (which is based on a greedy strategy that selects the best available channels for a given transmission), and a naive scheme (which always tries to select high-frequency channels if available for a given transmission, while leaving low-frequency channels for other users). Specifically, we consider a single-hop CRN, where all users can hear each other. This CRN coexists with two PRNs in a 100 meter × 100 meter field. The PRNs operate in the 600 MHz and 2.4 GHz bands. Each PRN band consists of one channel of bandwidth 1.5-MHz. The number of PR users in each PRN is 50. Each user in the ith PRN acts as an ON/OFF source, where it is ON while transmitting and OFF otherwise. We define the “activity factor” αi as the fraction of time in which the ith type PR user is ON (i.e., the probability that the source is in the ON state). The source is further characterized by the distribution of its ON and OFF periods, which are both taken to be exponential. We set the average ON period to be the duration of one time slot. In other words, traffic correlations are captured using a two-state Markov model. The appropriateness of the 2-state ON/OFF model has been demonstrated in several previous works, e.g., [6], [13], [20], [36], [37]. In essence, the ON/OFF behavior is attributed to the bursty nature of many types of network traffic, including voice traffic and VBR video streaming. Note that one potential PRN is a cellular network that transports voice traffic. We set the αi probabilities for the two PRNs to 0.5 and 0.3, respectively. The transmission power for each PR user is 0.5 Watt. For the CRN, we consider 40 mobile users. The random waypoint model is used for mobility, with the speed of a CR user uniformly distributed between 0 and 2 meters/sec. This results in dynamic, time-varying topologies. We assume that a CR user can use up to two data channels simultaneously. We (1) (2) set the interference mask to Pmask = Pmask = 50 mW. We also set the forgetting factor α to 0.6, the observation window Twin = 0.5 second, and the number of rings around a CR user m = 12. For a fair comparison, we let all schemes use the maximum allowable power vector Pmask . 2) Results: Under the above setup, Figure 12 shows that DDMAC improves the one-hop throughput by up to 25% (compared to BMC) and 34% (compared to the naive approach). More importantly, its throughput is within 7% of the optimal throughput, obtained via exhaustive search. Note that the exhaustive search implies that the instantaneous SINR values, location information, and rate demands are known to the decision-making entity that assigns channels to CR users (i.e., such search requires global information). Even if perfect knowledge of the SINR of each link and the rate demands are available, for large-scale networks, finding the optimal solution requires exhaustive search over a large state space, which grows exponentially with the number of CR users and the available channels. B. Multi-hop Scenarios 6 For our simulation setup, CR user demands are uniform. The destination CR user is uniformly selected from the one-hop neighbors and the packet generation rate are the same for all CR users. Thus Jain’s fairness index provides a meaningful metric for comparing the fairness of DDMAC and BMC. 1) Simulation Setup: We now evaluate the performance of DDMAC in more realistic (large-scale) network scenarios and contrast it with a typical multi-channel CSMA-based protocol that uses BMC for channel selection [3]. We consider four 11 0.45 1.25 1 0.75 Blocking rate (%) Throughput (Packet/time slot) 1.5 Optimal DDMAC BMC Naive approach 0.5 0.25 0 0 10 20 30 λ (Packet/sec) 40 BMC DDMAC 0.4 0.35 0.3 0 0.1 Fig. 12. Throughput vs. λ for a small-scale network (comparison with the optimal scheme). 7 Random-grid is a realistic topology that models constrained scenarios. For example, a building could have various offices, where each office may contain several wireless devices. 0.5 (a) Blocking rate vs. λ. Single hop througput (Packet/time slot) PRNs and one CRN. Users in each PRN are uniformly distributed over a 500 meter × 500 meter area. The PRNs operate in the 600 MHz, 900 MHz, 2.4 GHz, and 5.7 GHz bands, respectively. Each PRN band consists of three non-overlapping 1-MHz channels. The number of PR users in each PRN is 300. The αi probabilities for the four PRNs are 0.5, 0.3, 0.3, 0.1, respectively. The transmission power for each PR user is 0.5 Watt. For the CRN, we consider a random-grid topology7, where 225 mobile CR users are placed within the 500 meter × 500 meter field. The field is split into 225 smaller squares, one for each CR user. The location of a mobile user within the small square is selected randomly. For each generated packet, the destination is selected randomly from the one-hop neighbors. Within each small square, the random waypoint model is used for CR mobility, with the speed of a CR user uniformly distributed between 0 and 2 meters/sec. We assume that a CR user can use up to three data channels simultaneously. We set (1) (2) (12) the interference mask to Pmask = Pmask = . . . = Pmask = 50 mW. The reported results are the average of 100 experiments. In our design, we assume an exclusive channel occupancy policy on CR transmissions (i.e., no CR-CR interference). However, hidden-terminal problem can still occur in this scenario due to imperfect control. Our simulations relax these assumptions and account for all sources of interference, including those that are far away from a receiver and use common channels. Remark: Our simulations only address the MAC layer aspects and assume that route computations have already been carried out. Taking the destination from a node’s one-hop neighbors is intended only to convey the need for channel access. A “destination” in this context could be the next hop where a packet is to be forwarded to or the final packet destination. Randomly selecting a neighbor as a destination is realistic in terms of packet forwarding, especially when multiple flows (like file transfers, messaging, or VoIP) pass through a node. 2) Results: We first compare the performance of DDMAC to that of the BMC scheme. For a fair comparison, we let both schemes use the maximum allowable power vector Pmask . We set α to 0.6, Twin to 0.5 second, and m to 12. Figures 13(a) and (b) show that under moderate and high traffic loads, DDMAC significantly reduces the connection blocking rate and improves the overall one-hop throughput by up to 30%. This 0.2 0.3 0.4 Packet genaration rate 40 35 30 25 20 BMC DDMAC BMC−Raleigh DDMAC−Raleigh 15 10 5 0 0 0.1 0.2 0.3 0.4 Packet generation rate (Packet/sec) 0.5 (b) Throughput vs. λ (with and without a Raleigh fading component). Fig. 13. Performance of a CRN. improvement is attributed to the increase in the number of simultaneous transmissions in DDMAC. Note that under low traffic load, the throughput of DDMAC gracefully degrades to that of BMC due to the availability of a sufficient number of channels. The system performance under Raleigh channel model (i.e., varying channel conditions) is also investigated in Figure 13(b). We consider normalized random variables to capture the fading processes [8]. The results show that, under such varying channel conditions, the same trends that were noticed for the AWGN channel are observed here. Figure 13(b) shows that the impact on throughput is almost the same under AWGN and Raleigh channel models. Similar observations have also been reported in an experimental study in [38] for IEEE 802.11b wireless LAN. In [38], the authors showed that the impact on throughput and packet error rate were virtually identical under AWGN and Raleigh channel models. In Figure 14(a), we focus on the per-user throughput performance under DDMAC8 . As shown in this figure, although DDMAC requires a pair of CR users to communicate over a set of channels that may not be optimal from one user’s perspective, the per-user throughput of DDMAC under moderate and high traffic loads is still greater than that of the BMC scheme. This is because DDMAC attempts to serve a given CR transmission first using the preferable channel list and preserves 8 This figure shows the average worst-case throughput performance among all CR users. 0.1 0.08 0.06 0.04 BMC DDMAC 0.02 0 0 0.1 0.2 0.3 Packet generation rate 0.4 0.5 (a) Per-user single-hop throughput. Single−hop throuhput (Packet/time slot) Single−hop throughput (packet/ time slot) 12 0.12 40 λ = 0.2 λ = 0.25 λ = 0.3 35 30 25 20 5 10 m 15 (a) Throughput vs. number of rings (m) around a CR user. Fairness Index 0.9 0.8 BMC DDMAC 0.7 0.6 0.5 0 0.1 0.2 0.3 0.4 Packet generation rate 0.5 Single−hop throuput (Packet/time slot) 1 38 T 36 Twin = 0.3 s 34 Twin = 1 s 32 win = 0.03 s λ = 0.3 Twin = 4 s 30 28 26 24 22 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Forgitting factor α 0.8 0.9 1 (b) Fairness index vs. λ. Fig. 14. Per-user throughput and fairness Performance. (b) Throughput vs. α for different Twin values. Fig. 15. the “better” channels for other transmissions. However, if the aggregate rate of this transmission cannot be satisfied using the preferable list, DDMAC attempts to serve this transmission using the remaining available channels. Next, we compare the fairness index of DDMAC to that of BMC. Compared to BMC, Figure 14(b) shows that DDMAC slightly improves the network fairness and preserves longterm fairness properties. This improvement occurs because DDMAC motivates cooperation among neighbors to maximize their network-wide benefit. The effect of dividing the transmission range of a CR user is depicted in Figure 15(a) for different values of λ. As m increases, the throughput increases up to a certain point. For m ≥ 12, no significant improvement is observed in the network throughput. This is because the preferable-channel assignment mechanism merges the m regions into K ≤ m regions, i.e., over-splitting Rc is not useful. In Figure 15(b), we study the impact of α and Twin on the performance of DDMAC. We set λ = 0.3 packet/slot. The network throughput versus α for different values of Twin is shown in the figure. It is clear that the throughput depends on the choice of α and Twin . As Twin increases, α should increase to give much more importance to recent observations without entirely discarding older observations. Table II shows the best throughput performance and the associated optimal value of α (α∗ ), obtained from simulation, for different values of Twin . It is clear that if Twin is too small or too large, the throughput reduces significantly. Scheme BMC DDMAC(Twin DDMAC(Twin DDMAC(Twin DDMAC(Twin DDMAC(Twin Performance of DDMAC. = 0.03 s) = 0.3 s) = 0.4 s) = 1 s) = 4 s) α∗ 0.1 0.6 0.6 0.8 1.0 Best throughput (packet/slot) 25 26 33.6 33.85 33.89 28 TABLE II P ERFORMANCE OF DDMAC AT THE OPTIMAL α AS A FUNCTION OF Twin . We also investigate the robustness of DDMAC under inaccurate distance estimation, which is mainly caused by mobility, multi-path propagation, reflection, and fading effects. The estimated distance de is given by (1 + ξ) d, where ξ is a uniform estimation error (ξ ∼ Uniform[−ǫ, ǫ]). Figure 16(a) shows the effect of inaccurate distance estimation on throughput as a function of ǫ under different traffic loads. It can be observed that there are no significant changes in the throughput for different values of ǫ. Figure 16(b) gives the percentage reduction in throughput due to inaccurate d as a function of λ for different values of ǫ. This figure shows that the maximum percentage of reduction in throughput due to inaccurate estimation of d is less than 6%. The results in Figure 16 indicate that channel assignment 40 35 8 6 BMC DDMAC 4 2 30 0 0 25 20 Fig. 17. 0.1 0.2 0.3 0.4 Packet generation rate 0.5 End-to-end throughput vs. λ 15 Low load 10 Moderate load 5 0 0.05 High load 0.1 0.15 ε 0.2 0.25 Percentage of reduction in throughput (a) Throughput vs. ǫ. 6 ε = 5% 5 ε = 15% ε = 25% 4 3 2 1 0 0 0.1 0.2 0.3 0.4 Packet generation rate channel assignment process in the design of DDMAC. We compared the performance of DDMAC with that of a reference multi-channel MAC protocol that is designed for typical multichannel systems (BMC). We showed that, under moderate and high traffic loads, DDMAC achieves about 30% increase in throughput over the BMC scheme, with manageable processing overhead. Although DDMAC requires a pair of CR users to communicate on a channel that may not be optimal from a user’s perspective, we showed that the average per-user throughput of DDMAC under moderate and high traffic loads is greater than that of the BMC scheme. Furthermore, DDMAC preserves (even slightly improves) throughput fairness relative to BMC. In summary, DDMAC provides better spectrum utilization by reducing the connection blocking probability and increasing the system throughput. To the best of our knowledge, DDMAC is the first CRN MAC protocol that utilizes the radio propagation characteristics to improve the overall network throughput. 0.5 (b) Percentage of reduction in throughput vs. λ. Fig. 16. 13 10 End−2−end Throughput (Packet/time slot) Single−hop throughput (Packet/time slot) in DDMAC is quite robust to distance estimation errors. This is because DDMAC requires only rough estimates of user distribution, distances among users, and local traffic conditions in order to dynamically adapt channel assignments to current network traffic. Impact of inaccurate distance estimation in DDMAC. Finally, we study the end-to-end throughput for both BMC and DDMAC. Specifically, for each generated packet, the destination node is randomly selected to be any node in the network. We use a min-hop routing policy, but we ignore the routing overhead. For both schemes, the next-hop candidates are nodes that are within the transmission range of the transmitter. Figure 17 shows that under moderate and high traffic loads, DDMAC significantly improves the overall network throughput (inline with the results in Figure 13(b)). VIII. C ONCLUSIONS In this paper, we proposed an opportunistic distancedependent MAC protocol for CRNs (DDMAC). DDMAC improves the CRN throughput through cooperative channel assignment, taking into consideration the non-adjacency of frequency channels and the imposed power masks. We presented a heuristic stochastic channel assignment scheme that dynamically exploits the dependence between the signal attenuation model and the transmission distance. Our scheme accounts for traffic dynamics. It assigns channels with lower average SINR to shorter transmission distances to increase the number of simultaneous transmissions. We integrated the R EFERENCES [1] “FCC, spectrum policy task force report, ET docket no. 02-155,” Nov. 2002. [2] H. Bany Salameh, T. Shu, and M. Krunz, “Adaptive cross-layer MAC design for improved energy efficiency in multi-channel wireless sensor networks,” Ad Hoc Networks, vol. 5, no. 6, pp. 844–854, 2007. [3] N. Jain, S. Das, and A. Nasipuri, “A multichannel CSMA MAC protocol with receiver-based channel selection for multihop wireless networks,” in Proceedings of the 9th Int. Conf. on Computer Communications and Networks (IC3N), Oct. 2001, pp. 432–439. [4] H. Bany Salameh and M. Krunz, “Channel Access Protocols for Multihop Opportunistic Networks: Challenges and Recent Developments,” IEEE Network-Special Issue on Networking over Multi-hop Cognitive Networks, July 2009. [5] M. Riback, J. Medbo, J. Berg, F. Harrysson, and H. Asplund, “Carrier frequency effects on path loss,” in Proceedings of Vehicular Technology Conference (VTC), May 2006, pp. 2717–2721. [6] T. Shu, S. Cui, and M. Krunz, “Medium access control for multi-channel parallel transmission in cognitive radio networks,” in Proceedings of the IEEE GLOBECOM Conference, Nov. 2006. [7] A. Nasipuri and S. Das, “Performance of multi-channel ad hoc networks,” International Journal of Wireless and Mobile Computing, vol. 1, no. 3/4, pp. 191–203, 2006. [8] T. S. Rappaport, Wireless Communications-Principles and Practice, Prentice-Hall Press, 2001, 2nd edition. [9] E. Arikan, “Some complexity results about packet radio networks,” IEEE Transactions on Information Theory, vol. 30, no. 4, pp. 681–685, 1984. [10] A. Behzad and I. Rubin, “Multiple access protocol for power-controlled wireless access nets,” IEEE Transactions on Mobile Computing, vol. 3, no. 4, pp. 307–316, 2004. [11] J. So and N. Vaidya, “Multi-channel MAC for ad hoc networks: Handling multi-channel hidden terminals using a single transceiver,” in Proceedings of the ACM International Symposium on Mobile and Ad-Hoc Networking and Computing (MobiHoc), Tokyo, Japan, May 2004, pp. 222 – 233. [12] Q. Zhao, L. Tong, and A. Swami, “Decentralized cognitive MAC for dynamic spectrum access,” in Proceedings of the IEEE DySPAN Conference, Nov. 2005, pp. 224–232. 14 [13] Y. Xing, R. Chandramouli, S. Mangold, and S. Shankar, “Dynamic spectrum access in open spectrum wireless networks,” IEEE Journal on Selected Areas in Communications, vol. 24, no. 3, pp. 626–637, 2006. [14] Y. Xing, C. Mathur, M. Haleem, R. Chandramouli, and K. Subbalakshmi, “Dynamic spectrum access with QoS and interference temperature constraints,” IEEE Transactions on Mobile Computing, vol. 6, no. 4, pp. 423–433, 2007. [15] R. Menon, R. Buehrer, and J. Reed, “Outage probability based comparison of underlay and overlay spectrum sharing techniques,” in Proceedings of the IEEE DySPAN Conference, Nov. 2005, pp. 101–109. [16] N. Nie and C. Comaniciu, “Adaptive channel allocation spectrum etiquette for cognitive radio networks,” in Proceedings of the IEEE DySPAN Conference, Nov. 2005, pp. 269–278. [17] S. Haykin, “Cognitive radio: Brain-empowered wireless communications,” IEEE Journal on Selected Areas in Communications, vol. 23, no. 2, pp. 201–220, 2005. [18] T. C. Clancy, “Achievable capacity under the interference temperature model,” in Proceedings of the IEEE INFOCOM Conference, May 2007, pp. 794–802. [19] Y. Yuan, P. Bahl, R. Chandra, T. Moscibroda, and Y. Wu., “Allocating dynamic time-spectrum blocks in cognitive radio networks,” in Proceedings of the ACM International Symposium on Mobile and Ad-Hoc Networking and Computing (MobiHoc), Sept. 2007. [20] H. Bany Salameh, M. Krunz, and O. Younis, “MAC protocol for opportunistic cognitive radio networks with soft guarantees,” IEEE Transactions on Mobile Computing, vol. 8, no. 10, pp. 1339–1352, Oct. 2009. [21] D. Cabric, S. Mishra, and R. Brodersen, “Implementation issues in spectrum sensing for cognitive radios,” in Proceedings of the 38th Asilomar Conference on Signals, Systems and Computers, Nov. 2004, pp. 772–776. [22] D. Cabric and R. Brodersen, “Physical layer design issues unique to cognitive radio systems,” in Proceedings of the IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), Sept. 2005, pp. 759–763. [23] A. Sabharwal, A. Khoshnevis, and E. Knightly, “Opportunistic spectral usage: Bounds and a multi-band CSMA/CA protocol,” IEEE/ACM Transactions on Networking, vol. 15, no. 3, pp. 533–545, 2007. [24] S. Jones, N. Merheb, and I. Wang, “A cognitive MAC protocol using statistical channel allocation for wireless ad-hoc networks,” in Proceedings of Wireless Communications and Networking Conference (WCNC), March 2007, pp. 105–110. [25] J. Zhao, H. Zheng, and G.-H. Yang, “Distributed coordination in dynamic spectrum allocation networks,” in Proceedings of the IEEE DySPAN Conference, Nov. 2005, pp. 259–268. [26] I. Akyildiz, W.-Y. Lee, M.C. Vuran, and S. Mohanty, “Next generation dynamic spectrum access cognitive radio wireless networks: A survey,” Computer Networks, vol. 50, no. 13, pp. 2127–2159, 2006. [27] A. Hoang and Y. Liang, “Maximizing spectrum utilization of cognitive radio networks using channel allocation and power control,” in Proceedings of Vehicular Technology Conference (VTC), Sept. 2006, pp. 1–5. [28] M. Youssef and A. Agrawala, “The Horus WLAN location determination system,” in MobiSys’05: Proceedings of the 3rd International Conference on Mobile Systems, Applications, and Services, 2005, pp. 205–218. [29] S. Kiran and R. Chandramouli, “An adaptive energy-efficient link layer protocol using stochastic learning control,” in Proceedings of the IEEE ICC Conference, May 2003, pp. 1114–1118. [30] Y. Chang, T. Ho, and L. Kaelbling, “Mobilized ad-hoc networks: A reinforcement learning approach,” in Proceedings of the International Conference on Autonomic Computing, May 2004, pp. 240–247. [31] H. Khalife, S. Ahuja, N. Malouch, and M. Krunz, “Probabilistic path selection in opportunistic cognitive radio networks,” in Proceedings of the IEEE GLOBECOM Conference, Nov. 2008. [32] A. Muqattash and M. Krunz, “POWMAC: A single-channel power control protocol for throughput enhancement in wireless ad hoc networks,” IEEE Journal on Selected Areas in Communications, vol. 23, no. 5, pp. 1067– 1084, May 2005. [33] R. Garces and J. Garcia-Luna-Aceves, “Collision avoidance and resolution multiple access for multichannel wireless networks,” in Proceedings of the IEEE INFOCOM Conference, March 2000, pp. 595–602. [34] “Mesquite Software Incorporation,” www.mesquite.com. [35] R. Jain, The Art of Computer System Performance Analysis, New York: John Wiley & Sons, 1991. [36] E. Sousa, “Performance of a spread spectrum packet radio network link in a poisson field of interferers,” IEEE Transactions on Information Theory, vol. 38, pp. 1743–1754, Nov. 1992. [37] E. Sousa and J. Silvester, “Optimum transmission ranges in a directsequence spread-spectrum multihop packet radio network,” IEEE Journal on Selected Areas in Communications, vol. 8, no. 5, pp. 762–771, June 1990. [38] C. Steger, P. Radosavljevic, and J. Patrick Frantz, “Performance of IEEE 802.11b wireless LAN in an emulated mobile channel,” in Proceedings of IEEE Fall Vehicular Technology Conference (VTC), April 2003. Haythem A. Bany Salameh received the Ph.D. degree in electrical and computer engineering from the University of Arizona, Tucson, in 2009. He is currently an Assistant Professor of electrical engineering with the Hijawi Faculty for Engineering Technology, Yarmouk University (YU), Irbid, Jordan. He joined YU in August 2009, after a brief postdoctoral position with the University of Arizona. His current research interests are in system architecture and communication protocol designs for cognitive radio networks with emphasis on spectrum access and channel/power assignment. In Summer 2008, he was a member of the R&D LTE (Long Term Evolution) Development Group, QUALCOMM, Inc., San Diego. He serves as a reviewer for many IEEE conferences and journals. Marwan Krunz is a professor of electrical and computer engineering at the University of Arizona. He directs the wireless and networking group and is also the UA site director for Connection One, a joint NSF/state/industry IUCRC cooperative center that focuses on RF and wireless communication systems and networks. Dr. Krunz received his Ph.D. degree in electrical engineering from Michigan State University in 1995. He joined the University of Arizona in January 1997, after a brief postdoctoral stint at the University of Maryland, College Park. He previously held visiting research positions at INRIA, HP Labs, University of Paris VI, and US West (now Qwest) Advanced Technologies. His research interests lie in the fields of computer networking and wireless communications. His current research is focused on cognitive radios and SDRs; distributed radio resource management in wireless networks; channel access and protocol design; MIMO and smart-antenna systems; UWB-based personal area networks; energy management and clustering in sensor networks; media streaming; QoS routing; and fault monitoring/detection in optical networks. He has published more than 150 journal articles and refereed conference papers, and is a co-inventor on three US patents. M. Krunz is a recipient of the National Science Foundation CAREER Award (1998). He currently serves on the editorial boards for the IEEE Transactions on Mobile Computing and the Computer Communications Journal. He previously served on the editorial board for the IEEE/ACM Transactions on Networking (2001-2008). He was a guest co-editor for special issues in IEEE Micro and IEEE Communications magazines. He served as a technical program chair for various international conferences, including the IEEE WoWMoM 2006, the IEEE SECON 2005, the IEEE INFOCOM 2004, and the 9th Hot Interconnects Symposium (2001). He has served and continues to serve on the executive and technical program committees of many international conferences and on the panels of several NSF directorates. He gave keynotes and tutorials, and participated in various panels at premier wireless networking conferences. He is a consultant for a number of companies in the telecommunications sector. Ossama Younis is a senior research scientist in Applied Research at Telcordia Technologies, Inc. He received his B.S. and M.S. degrees in computer science from Alexandria University, Egypt. Then, he received his Ph.D. degree in computer science from Purdue University, USA in August 2005. He has served as the web chair of IEEE ICNP 2008 and on the Technical Program Committee of several international conferences. His research interests are in the architecture and experimentation of network protocols and applications, especially wireless sensor networks, cognitive-radio networks, and Internet tomography. distributed systems, secure protocol design, and cross-layer optimization. He is a member of the ACM and the IEEE.