Model-Based Opportunistic Channel Access in Cognitive Radio Enabled Dynamic Spectrum Access Networks

Transcription

1 Model-Based Opportunistic Channel Access in Cognitive Radio Enabled Dynamic Spectrum Access Networks Manuj Sharma, Anirudha Sahoo and K. D. Nayak Indian Institute of Technology Bombay Mumbai, India Advanced Numerical Research and Analysis Group Defence Reseacrh and Development Organization Hyderabad, India Abstract We propose a model-based channel access mechanism for cognitive radio-enabled secondary network, which opportunistically uses the channel of an unslotted primary network when the channel is sensed idle. We refer to primary network as the network which carry the main traffic in a designated spectrum band. We have considered IEEE WLAN as a de facto primary network operating in ISM band. Our study focuses on a single WLAN channel that is used by WLAN clients and a WLAN server for a mix of , FTP, and HTTP-based web browsing applications. We model the occupancy of the channel by primary WLAN nodes as an alternating renewal process. When the secondary sender node has one or more frames to send, this model is used by the the sender and receiver pair to estimate residual idle time duration after the channel is sensed as idle. The secondary sender then opportunistically transmits frames in that duration without significantly degrading performance of the primary WLAN applications. Our simulation results show that the performance of secondary network is sensitive to the channel sensing duration and that high secondary throughput can be achieved without affecting the primary network significantly by choosing appropriate value of channel sensing duration. I. INTRODUCTION Opportunistic Spectrum Access (OSA) has emerged as a promising approach to efficiently utilize the electromagnetic spectrum. In this approach, the secondary users use those parts of the spectrum band that are not currently utilized in space or time by any primary user. In doing so, the secondary devices must ensure that their transmissions do not cause unacceptable interference at any primary receiver within their interference range. In this report, we propose a model-based approach for opportunistic access of unslotted primary network channel by secondary network nodes. Each secondary node collects the statistical data about the channel occupancy by primary nodes. It then fits an appropriate distribution to the statistical data to construct a channel occupancy model and proceeds with the algorithms presented in this report to opportunistically use the channel. We refer to primary network as the network which carry the main traffic in a designated spectrum band. For our work, we have considered based WLAN as a de facto primary network operating in ISM band. WLAN devices constitute the primary devices in this band. Other networks, such as Bluetooth network or sensor networks, which operate in these bands, can be treated as secondary networks. Coexistence of WLAN as a primary network and Bluetooth as a secondary network has been investigated in [1]. As mentioned above, each secondary node collects the statistical data about the channel occupancy by primary nodes and fits an appropriate distribution to construct the channel occupancy model. In this study, we have fitted the hyperexponential distribution (HED) to idle and busy time data values. For WLAN based primary network, HED is a preferred choice because it provides better fit for highly variable channel idle time statistics. Channel busy time statistical data is not highly variable, so either HED or exponential data can provide a good fit for busy time distribution. Statistical parameters for idle and busy time data values and the fitted distributions is given in Section IV. When secondary sender has one or more frames to send, it senses the primary channel. If the channel is busy, it backs off randomly. Otherwise, it uses the model to estimate the residual idle time on the channel, and transmit as many frames as it can during the estimated remaining idle time. The proposed protocol enables better sharing of WLAN channel between the primary WLAN and the secondary network, leading to better channel utilization without significantly degrading the performance of the primary network. The secondary network considered in our work is an unslotted network. We have constructed and applied the model for a single WLAN primary channel, but the approach can be used for multichannel primary networks as well. We believe that the proposed protocol for secondary networks is equally applicable and useful for other primary network such as WiMAX, GSM or TV network. II. ASSUMPTIONS In this section, we specify the assumptions made in our work.

2 1) There is one WLAN channel in the primary network, which is used for data communication. Control frames are transmitted on a separate control channel. 2) The primary network is an unslotted network, which uses CSMA/CA-based protocol. We consider IEEE as the de facto primary network in ISM band. 3) The secondary nodes perform perfect sensing. That is, the probability of false alarm and missed detection is zero. 4) We assume that each secondary node has a separate sensing module for the data channel to obtain channel occupancy statistics. For a single channel network, the receiver module of the secondary node could as well sense the channel when the node is not transmitting. But when the protocol is extended for a multi-channel network, using a separate sensing module is more desirable so that when the transmitter transmits on one channel, the sensing module can sense the other channels. The assumption of a separate sensing module is made in our work keeping this extension in view. III. NOTATIONS In this section, we introduce the main notations used in rest of the paper. I, B: Random variables denoting channel idle and busy periods respectively. C: Random variable denoting cycle time for the channel. (C I + B) ξ(t): Ending time of the last (previous) renewal cycle at sensing time t. ph I : Number of phases in Hyperexponential distribution (HED) for idle time random variable I. ph B : Number of phases in Hyperexponential distribution (HED) for busy time random variable B. λ i : Rate parameter of the i th phase (exponential) of the HED for random variable I. (i 1... ph I ) α i : Probability of the i th phase (exponential) of HED for random variable I. ( ph I i1 α i 1) µ i : Rate parameter of the i th phase (exponential) of HED for random variable B. (i 1... ph B ) β i : Probability of the i th phase (exponential) of HED for random variable B. ( ph B i1 β i 1) HED(ph I, λ i, α i ): Denotes Hyperexponential distribution (for random variable I) with parameters (ph I, λ i, α i ), where i 1... ph I. HED(ph B, µ i, β i ): Denotes Hyperexponential distribution (for random variable B) with parameters (ph B, µ i, β i ), where i 1... ph B. T obs : Denotes the duration for which secondary nodes passively sense the channel to gather occupancy statistics (idle and busy time durations) of the channel by primary nodes. In our work, this duration is first 5 hours of the simulation time (out of total simulation time of 8 hours). T comm : Denotes the duration for which secondary nodes opportunistically communicate on the primary WLAN Fig. 1: Alternating renewal process channel. In our work, this duration is last 3 hours of the simulation time (after T obs ). S(t): A boolean variable that denotes the outcome of sensing the channel at time instant t by a secondary node. S(t) can either take value BUSY or IDLE. T RI (t): A random variable that denotes residual idle time for the channel at sensing instant t. T h RI : A predefined threshold for residual idle time; secondary node uses the channel opportunistically only if the estimated residual idle time is more than this threshold. (T h RI (one secondary frame transmission time + cushion time) 3 microseconds). T ERI (t): Effective residual idle time on the channel. This is computed as: T ERI (t) min{(t RI (t)) sender, (T RI (t)) rcv }, where (T RI (t)) sender is the residual channel idle time estimated by the sender at time t and (T RI (t)) rcv is the residual channel idle time at time instant t as estimated by the receiver using channel occupancy model. SRT S: Secondary RTS frame SCT S: Secondary CTS frame CON F : Confirmation frame IV. PRIMARY NETWORK CHANNEL OCCUPANCY MODEL We treat a channel as a 2-state system. We consider a channel to be idle (or available) (from SU s perspective) when it is not used by any primary user, and busy (or occupied) (from SU s perspective) when it is used by any primary user. Channel occupancy modeling is performed by secondary user. In order to construct the channel occupancy model, each secondary node passively gathers the primary user s channel occupancy statistics (idle and busy durations) for T obs duration. Channel occupancy is modeled as an alternating renewal process in which a cycle, consisting of idle duration followed by busy duration, repeats (renews) in time. Renewal of a cycle is said to occur when the channel becomes idle (i.e. the primary user stops transmitting on the channel). On time line, such alternating renewal process is shown in Figure 1. In the figure, a vertical arrow denotes a renewal event; time between two vertical arrows denote a renewal cycle; solid segment within one cycle denote idle time; and dashed segment within a cycle denote busy time. Random variables T I, T B, and T C denote the idle time, busy time, and total cycle time respectively. Since T C is sum of T I and T B, its density can be computed as convolution of the densities of T I and T B. Two related concepts of interest to our work are residual idle time and residual busy time. We explain these two measures with the help of Figure 2 and Figure 3. In both the figures, T I (solid segment) and T B (dashed segment) denotes the idle time and busy time respectively within a cycle. t denotes the instant at

3 ph F B (t) β i (1 e µit ), (t ) (4) i1 The density function of renewal cycle random variable C can be computed as convolution of the density functions of random variables I and B: Fig. 2: Residual idle time at sensing instant t Fig. 3: Residual busy time at sensing instant t which we incident in the cycle (in our work, t is typically the time instant when a secondary node senses the channel). T R (t) denote the residual time of the complete cycle at instant t. If the channel is idle at instant t (see Figure 2), then the residual idle time is shown as T RI (t). Similarly, if the channel is busy at instant t (see Figure 3), then the residual busy time is shown as T RB (t) and is equal to the total cycle s residual time. Having gathered the primary network channel occupancy data during T obs duration, each secondary node fits an appropriate distribution to idle and busy periods. Phase type distributions, such as Hyperexponential distribution (HED) [2], have been shown to provide good fit to the values of non-negative continuous random variables [3], especially for highly variable data (with coefficient of variation > 1), but are non-trivial to track analytically. Exponential distributions are analytically tractable but may not provide satisfactory fit for highly variable data. In our work, we fit Hyperexponential distribution to both idle time values and busy time values using Expectation Maximization (EM) algorithm proposed in [4]. The Hyperexponential density and distribution functions (see [2] for description of these functions) of idle time random variable I (with estimated parameters (ph I, λ i, α i )), and busy time random variable B (with estimated parameters (ph B, µ i, β i )) are given as follows: 1) Hyperexponential density and distribution function of I: ph I f I (t) α i λ i e λit, (t ) (1) i1 ph I F I (t) α i (1 e λit ), (t ) (2) i1 2) Hyperexponential density and distribution function of B: ph f B (t) β i µ i e µit, (t ) (3) i1 f C (t) f I (t y)f B (y)dy ( ph I α i λ i e λi(t y))( ph B β j µ j e µjy) dy i1 ( ph I ph B j1 α i λ i β j µ j e λi(t y) e µjy) dy α i λ i β j µ j e λit e (λi µj)y dy α i λ i β j µ j e λit e (λi µj)y dy f C (t) α i λ i β j µ j e λit (e (λi µj)t 1) (λ i µ j ) α i λ i β j µ j (λ i µ j ) (e µjt e λit ) (5) The distribution function of random variable C can be computed using (5) to obtain: F C f C (u)du { ph I ph B α i λ i β j µ j (λ i µ j ) α i λ i β j µ } j (λ i µ j ) (e µju e λiu ) du (e µju e λiu )du (6) Let us denote the integral in the above equation as M and solve it. M (e µju e λiu )du e µju du e λju du (1 e µjt ) µ j + (e λit 1) λ i λ i(1 e µjt ) + µ j (e λit 1) λ i µ j M (µ je λit λ i e µjt ) + (λ i µ j ) λ i µ j (7)

4 TABLE I: Statistics of Channel Idle Time Samples and Fitted Distributions Mean Variance CoV Collected Data Sample EXPO HED HED HED HED TABLE II: Statistics of Channel Busy Time Samples and Fitted Distributions Mean Variance CoV For Sample EXPO HED HED HED HED Substituting the value of M from eq. (7) into eq. (6), we get, F C α i λ i β j µ { j (µj e λit λ i e µjt ) + (λ i µ j ) } (λ i µ j ) λ i µ j { αi β j (µ j e λit λ i e µjt ) } + α i β j (λ i µ j ) Equation (8) gives the distribution function of C. Table I shows the mean, variance and coefficient of variation of idle time data samples, as well as exponential (EXPO) and Hyperexponential distributions (HED-n) fitted to these samples (n denotes the number of phases in the fitted HED). Similarly, Table II shows the mean, variance and coefficient of variation of busy time data samples, as well as the exponential (EXPO) and Hyperexponential distributions (HED-n) fitted to these samples. Typically the range of the collected idle time sample values is very large as compared to the busy time sample values, indicating that the channel idle time sample exhibit large variability in its data values as against busy time sample. The high value of idle sample variance and idle sample coefficient of variation (CoV) in Table I is indicative of this fact. The range of idle time sample values collected in this study (through simulations) is seconds (minimum:.1 seconds; maximum: seconds). Similarly, the range of busy time sample values collected in this study (through simulations) is.6466 seconds (minimum:.241 seconds; maximum:.677 seconds). Since the coefficient of variation of idle time samples is significantly greater than 1, Hyperexponential distribution is a better fit for channel idle time. The coefficient of variation (CoV) of busy time samples (in Table II) is less than 1, therefore, both (8) TABLE III: Hyper Exponential Distribution (HED-5) Parameters for Idle Time values phase-1 phase-2 phase-3 phase-4 phase-5 α λ TABLE IV: Hyper Exponential Distribution (HED-5) Parameters for Busy Time values phase-1 phase-2 phase-3 phase-4 phase-5 β µ Fig. 4: Residual idle time computation HED and exponential distributions can provide a satisfactory fit. Readers should note from Table I that the variance and coefficient of variation of Exponential distribution fitted to idle time samples are extremely off-the-mark as compared to variance and coefficient of variation of collected data samples, indicating that exponential distribution is unlikely to be a good fit for channel idle time. Similarly, variance and coefficient of variation values of 2-phase idle time HED (denoted as HED- 2 in Table I) are significantly larger than sample variance and coefficient of variation. The idle and busy time HEDs with other number of phases have variance and coefficient of variation values in nearly similar range, and not too offthe-mark from the sample statistics. Increasing the number of phases for idle time HED from 4 to 1 does not change the statistics significantly. We have fitted 5-phase HEDs for both idle and busy time data samples in our simulations. The parameters of the fitted distributions (HED-5) for idle and busy time values are shown in Table III and Table IV respectively. V. MODEL BASED CHANNEL ACCESS PROTOCOL In this section, we describe our proposed model-based channel access protocol. A. Residual Idle Time Computation Based on the Fitted Model Let us assume that during the channel observation period (T obs ), a secondary node obtains the following idle and busy time data values: X 1, Y 1,... X n, Y n, where X i represents the i th idle time data value and Y i represents the i th busy time data value on the channel (see Figure 4). Then the i th observed cycle time can be computed as : c i X i + Y i. If we use ξ(t) to denote, in general, the ending time of

5 Algorithm 1 Algorithm for computing residual channel idle time at secondary node at sensing instant t 1: procedure COMPUTERESIDLETIME(t, ξ) 2: Generate idle random variable value I (I HED(p I, λ i, α i )). 3: Generate busy random variable value B (B HED(p B, µ i, β i )). 4: Compute the cycle time: C I + B. 5: while (ξ + C) < t do 6: ξ ξ + C 7: I HED(p I, λ i, α i ) 8: B HED(p B, µ i, β i ) 9: C I + B 1: end while 11: while ξ + I < t do 12: I HED(p I, λ i, α i ) 13: end while 14: T RI (t) (ξ + I) t 15: Return T RI (t) 16: end procedure the last renewal cycle prior to time t, then we can write ξ(t obs ) n j1 (X j + Y j ) n j1 c j. Note that c n is the last renewal cycle prior to T obs. If the first channel sensing by a secondary node after time T obs occurs at time t 1 and the channel is sensed idle, then the residual time at t 1 is computed as follows: Starting from the last renewal cycle prior to the previous observation/sensing time, (in this case, cycle c n prior to T obs, secondary node alternately generates idle and busy time random variable values using their respective distributions (I HED(ph I, λ i, α i ) and B HED(ph B, µ i, β i )), till it generates m th idle time random variable value (I m ), such that the current sensing time t 1 falls within the m th idle time (as shown in Figure 4). If we denote the i th generated cycle by C i (i.e. C i I i + B i ), then C m 1 is the last cycle prior to the current sensing tim t 1, and ξ(t 1 ) (the ending time of the last cycle (C m 1 ) prior to t 1 ) is calculated as ξ(t 1 ) ξ(t obs ) + m 1 j1 (I j + B j ) ξ(t obs ) + m 1 j1 C j. From Figure 4, we see that the residual idle time at t 1 can be obtained as: T RI (t 1 ) (ξ(t 1 ) + I m ) t 1. Extending the same explanation further, if the next sensing (after t 1 ) occurs at time t 2 and the channel is sensed idle, then the residual idle time at t 2 is computed as follows: Starting from the last renewal cycle prior to the previous observation/sensing time, (in this case, cycle C m 1 prior to t 1 ), secondary node alternately generates idle and busy time random variable values using their respective distributions, till it generates q th idle time random variable value (I q ), such that the current sensing time t 2 falls within the q th idle time (as shown in Figure 4). Then ξ(t 2 ) (the ending time of the last cycle (C q 1 ) prior to t 2 ) is calculated as ξ(t 2 ) ξ(t 1 ) + q 1 jm (I j + B j ) ξ(t 1 ) + q 1 jm C j. From Figure 4, we see that the residual idle time at t 2 can be obtained as: T RI (t 2 ) (ξ(t 2 ) + I q ) t 2. Algorithm for computing residual idle time by a secondary node at sensing Algorithm 2 CG-MAC algorithm at secondary sender node 1: procedure CG-MAC(t) 2: Compute S(t). 3: if S(t) IDLE then 4: T RI (t) ComputeIdleResTime(t, ξ) 5: if T RI (t) > T h RI then 6: Construct SRT S(T RI (t)) frame. 7: Send SRT S(T RI (t)) frame to the receiver. 8: else 9: Perform random backoff before sensing the channel again. 1: end if 11: else if S(t) BUSY then 12: Perform random backoff before sensing the channel again. 13: end if 14: On receiving SCT S(T ERI (t)) frame form the receiver, broadcast CONF (T ERI (t)) frame. 15: Compute number of frames (say, M) that can be transmitted in T ERI (t) duration. 16: Compute number of frames to transmit (say, X): X min(m, number of frames available in transmission queue) 17: Transmit X frames back-to-back. 18: end procedure instant t and using the fitted channel idle and busy time models is given in Algorithm 1. B. The Protocol We assume that each secondary node has constructed channel occupancy model, as described in Section IV. Whenever a secondary sender node has one or more frames to transmit at time instant t, its model-based channel access protocol senses the channel (line 2 of Algorithm 2), and if the channel is sensed idle, it estimates the residual idle time (T RI (t)) for the channel using the channel occupancy model (line 4). If the estimated residual idle time is more than a predefined threshold T h RI (line 5), it sends a SRTS frame (containing the estimated residual channel idle time) to the intended secondary receiver node (lines 6-7); otherwise it performs random backoff. The receiver, on receiving the SRTS frame, senses the channel (line 3 of Algorithm 3), and if the channel is sensed idle (line 4), sends the estimated effective residual idle time back to the sender in an SCTS frame (lines 7-9). On receiving the SCTS frame, the secondary sender node broadcasts a CONF frame with the received Effective Residual Idle Time value (line 14 of Algorithm 2), and transmits the minimum of the number of the frames available in its transmission queue and the maximum number of frames that can be transmitted in the Effective Residual Idle Time duration) (lines of Algorithm 2). SRTS, SCTS and CONF frames are transmitted on the control channel.

6 Algorithm 3 CG-MAC algorithm at secondary receiver node (for Single Channel) 1: procedure CG-MAC(t) 2: Receive SRT S(T RI (t)) frame from the sender. 3: Compute S(t). 4: if S(t) IDLE then 5: T RI (t) ComputeIdleResTime(t, ξ) 6: 7: if TRI c (t) > T h RI then T ERI (t) min{(t RI (t)) sender, (T RI (t))} 8: Construct SCT S(T ERI (t)) frame. 9: Send SCT S(T ERI (t)) frame to the Sender. 1: else 11: Do nothing. Sender will retransmit 12: end if 13: else if S(t)) BUSY then 14: Do nothing. Sender will retransmit 15: end if 16: Tune to the data channel and receive the frames sent back-to-back by the secondary sender node. 17: end procedure VI. SIMULATION RESULTS In this section, we discuss our simulation experiments and the results. A. Simulation Model We consider a WLAN with a single data channel as a primary network. The channel is used by two primary WLAN client nodes (WLAN-Client-1 and WLAN-Client-2), and a primary WLAN server node (WLAN-Server). The WLAN- Client-1 node runs parallel sessions of FTP and applications. WLAN-Client-2 node runs a web browsing application. The WLAN-Server acts as a server for all the three types of applications. In addition to the primary WLAN, we consider one pair of secondary devices (a secondary sender node and a corresponding secondary receiver node) that opportunistically uses the channel. All the five nodes are within the transmission range of each other. We use OPNET simulator [5] to simulate the model. We have used High load , High load FTP, and Heavy browsing HTTP application configurations provided by the simulator to run on the WLAN nodes. The send and receive interarrival times for s are exponentially distributed with mean 36 seconds, whereas the size is 2 bytes (constant). Time between two file transfer requests is also exponentially distributed with mean 36 seconds, and the file size is 5 bytes (constant). The percentage of file get commands to the total FTP commands is 5 %. For HTTP application, the time between page requests is exponentially distributed with mean 6 seconds. Each page has 5 medium image objects and 1 constant object of 1 bytes. The simulation runs for 8 hours. For the first 5 hours (T obs duration), a sensor module within each secondary node passively senses the channel to gather occupancy statistics. This observed data consist of TABLE V: Main Simulation Parameters Parameters Value Number of WLAN nodes 3 Number of secondary nodes (SN) 2 Simulation duration (Hours) 8 Channel occupancy data collection duration (by SN) First 5 Hrs WLAN channel number used in simulation 1 Channel data rate Channel bandwidth 11 Mbps 22 MHz Secondary node frame size (Bytes) 256 Clear Channel Assessment duration for SN Idle channel observation duration for SN (COD) ms Varied alternating sequence of channel idle and busy time durations (due to WLAN applications transmissions). The secondary nodes construct primary network s channel occupancy model based on this observed occupancy data. During the model construction phase, the secondary node subsumes very small idle periods into the busy periods as these very small idle periods are too small to be useful for opportunistic transmissions by secondary nodes. In our simulations, we have used a threshold value of 75 microseconds for this purpose. This value is approximatlely equal to the transmission time of a very small WLAN frame, which is dominated mainly by DIFS time (5 microsec) and slot time (2 microsec). Any idle time data value in the observed data file, which is less than this threshold of 75 microseconds, is considered too small for opportunistic transmission and is subsumed in the busy time values. Once the model is constructed, for the next 3 hours (T comm duration), the secondary node uses the channel model to estimate the residual channel idle durations and use the channel opportunistically along with the primary WLAN applications. A packet generator module within each secondary node generates Poisson traffic, with inter-frame arrival rate varying from 5 frames/sec to 5 frames/sec (across different simulation runs). Table V lists the main simulation parameters. We must emphasize that we have used a fixed observation window (of 5 hours) in order focus our investigation on performance of the model-based approach, once the model is constructed. In actual deployment of the algorithm, the model should be appropriately revised based on the data collected using new, possibly sliding, observation windows, as the traffic on the channel will change with time. With the above simulation model, three different scenarios are simulated. The scenarios are explained below. 1) Scn-1 (Only P N W LAN ): In this scenario, the secondary nodes are disabled and they do not transmit any traffic during the complete simulation. Only primary WLAN

7 nodes running , FTP, and web browsing applications are operational. This scenario is the base case which provides us the performance of primary network in absence of any secondary network. Performance of primary network in this scenario can be compared with its performance in the other two scenarios (Scn-2 and Scn- 3) to assess the impact of opportunistic usage of WLAN channel by a secondary network. 2) Scn-2 (P N W LAN + SN W LAN ): In this scenario, the secondary nodes uses WLAN protocol (i.e. secondary nodes are essentially WLAN nodes only) and do not construct any channel occupancy model. Instead, secondary sender node uses conventional WLAN protocol to exchange frames. In order to compare this scenario with other scenarios, the secondary network remain dormant for first 5 hours of simulation and becomes operational during last 3 hours of the simulation. This scenario enable us to compare the performance of secondary network when it shares the same network as a peer to the primary network. 3) Scn-3 (P N W LAN + SN MOD ): In this scenario, the secondary nodes observe the channel for first 5 hours, and construct the channel occupancy model based on the gathered occupancy data. The secondary nodes fit 5-phase Hyperexponential distributions to the busy and idle periods, and construct Alternating Renewal Processbased channel occupancy model. Whenever the secondary sender node has one or more frames to transmit, and it senses the channel idle, it continue to sense the channel for a predefined channel observation duration. If the channel remains idle even for this additional duration, the node uses the channel occupancy model to predict the residual idle time on the channel (as described in Section V-A) and transmits as many frames as possible in the predicted residual idle time. An important parameter in the proposed model-based approach (Scn-3) is channel observation duration (COD). As mentioned earlier in this section, the secondary node observes the channel idle and busy time values during first 5 hours. Some of these idle time values are large. We assume that the idle periods that are greater than or equal to 1. seconds corresponds to durations when no application session is active. In between these large idle periods, there exist a number of alternative sequence of very small idle and busy time values, which we assume corresponds to idle and busy times due to frame transmissions during active application sessions. We refer to this set of very small idle and busy time values that appear between two large idle time values as a block. The observed data file consist of several such blocks. The COD parameter should be set to such a value that enables the secondary node to exploit large idle times (when no application session is active) and avoid using small idle times that occur during frame transmissions of active application sessions, without decreasing secondary network throughput as much as possible. We study the impact of this parameter on the performance of secondary network. We set the value of COD parameter to four different values:, value-1, value- 2 and value-3. value-1 is obtained by taking the average of all the small idle time values (< 1. seconds) that appear in the observed data file; value-2 is obtained by computing the average of small idle times (< 1. seconds) in each block and taking the maximum of average values across all the blocks; value-3 is equal to the largest of the small idle time values (< 1.) that appear in the observed data file. Setting COD implies that once the secondary node senses the channel idle, it does not additionally senses the channel any longer. The performance of model-based approach for these four COD parameter values is given in the Result subsection. B. Performance Metrics We use the following performance metrics to assess the performance of model-based scenario (Scn-3) and compare it with the other two scenarios (Scn-1 and Scn-2). 1) Average throughput of secondary network (frames/sec): This metric represents the average throughput of secondary network, which is achieved as a result of opportunistic transmission by secondary nodes. It is calculated only for scenarios Scn-2, and Scn-3. The metric computation is based on last 3 hours of simulation time during which the secondary network opportunistically uses the channel. In the first scenario (Scn-1), the secondary nodes are not operational throughout the simulation, and therefore, the secondary network throughput is zero. 2) Average medium access delay (in seconds) for Wireless LAN: This metric represents the average delay in accessing the medium by all the WLAN nodes. It shows the impact of opportunistic transmissions by secondary nodes on the medium access delay experienced by primary WLAN nodes. 3) Average download response time for the application running on WLAN-Client-1 and WLAN-Server nodes 4) Average download response time for FTP application running on WLAN-Client-1 and WLAN-Server nodes 5) Average page response time of HTTP-based web browsing application running on WLAN-Client-2 and WLAN- Server nodes The last three metrics show the impact of opportunistic transmissions by secondary nodes on the performance of primary WLAN applications. C. Results In this section, we present the simulation results and compare the performance of HED model-based channel access protocol (used in Scn-3) with other scenarios (Scn-1 and Scn- 2). Main goal of the work reported in this paper is to study the throughput gains achieved by secondary network nodes by opportunistically transmitting on the primary WLAN channel, and to investigate the cost associated with such transmissions (in terms of user perceived impact on performance of the primary WLAN applications). Figure 5, Figure 6 and Figure 7

8 Fig. 5: Download Response Time (on Y axis) in Scn- 3 for Different COD values (Secondary Node Poisson Traffic Rate 5 Frames/sec; X-axis: Simulation time in hours) Fig. 7: HTTP Page Response Time (on Y axis) in Scn-3 for Different COD values (Secondary Node Poisson Traffic Rate 5 Frames/sec; X-axis: Simulation time in hours) Fig. 6: FTP Download Response Time (on Y axis) in Scn-3 for Different COD values (Secondary Node Poisson Traffic Rate 5 Frames/sec; X-axis: Simulation time in hours) shows the impact on primary , FTP and HTTP applications respectively when model based approach is used in Scn- 3 with different values of COD parameters, whereas 8 shows the average medium access delay experienced by the WLAN Server in Scn-3 with different values of COD. Figure 9 shows the secondary network throughput obtained for these COD values in Scn-3. The exact values of COD for the observed data in our simulations are indicated within parenthesis in the above-mentioned figures. We note from these figures that for COD values and value-1, the achieved secondary network throughput is relatively high (Figure 9) but their impact on primary applications performance is extremely adverse (Figure 5, Figure 6, Figure 7 and Figure 8). With COD parameter set to value-2 and value-3, the impact on primary applications Fig. 8: WLAN Server Average Media Access Delay (on Y axis) in Scn-3 for Different COD values (Secondary Node Poisson Traffic Rate 5 Frames/sec; X-axis: Simulation time in hours) is minimal, but the achieved secondary network throughput is less than that of value and value-1 case. However, the decrease in secondary throughput (with respect to value or value-1) in case of value-2 is much smaller compared to that of value-3. Hence COD value-2 is a favorable COD value in terms of striking a balance between impact on primary network and drop in throughput of secondary network. In rest of our simulations for Scn-3, we use COD value-2. Next, we investigate the impact of secondary node s opportunistic transmissions on the performance of WLAN applications, and compare it with the base case in which no

9 Fig. 9: Secondary Network Throughput (Frames/Sec) in Scn- 3 for Different COD values (Secondary Node Poisson Traffic Rate 5 Frames/sec). For interpretation of value-1, value- 2 and value-2, see Subsection VI-A. Fig. 11: FTP Download Response Time (on Y axis) for Different Secondary Node Poisson Traffic Rates in Modelbased Approach (Scn-3); X-axis: Simulation time in hours Fig. 1: Download Response Time (on Y axis) for Different Secondary Node Poisson Traffic Rates in Modelbased Approach (Scn-3); X-axis: Simulation time in hours secondary node is operational (Scn-1). In Figure 1, Figure 11, Figure 12 and Figure 13, we observe that for low traffic rates, the HED model-based secondary nodes (Scn-3) exhibit less impact on the WLAN applications and MAC performance. The impact on primary performance starts increasing when the secondary network traffic increases. Therefore, it is of more interest to study the performance of the proposed opportunistic access scheme when secondary traffic rate is high. We consider the secondary sender node s Poisson traffic rate of 5 frames/sec. Figure 14, Figure 15 and Figure 16 show the impact on primary applications performance in scenarios Scn-1, Scn-2 and Scn-3 (with COD value-2), when the secondary Poisson traffic rate (in Scn-2 and Scn-3) is 5 Fig. 12: HTTP Page Response Time (on Y axis) for Different Secondary Node Poisson Traffic Rates in Model-based Approach (Scn-3); X-axis: Simulation time in hours frames/sec. Figure 17 compares the impact of WLAN-based and model-based secondary node transmission on the average media access delay of WLAN server. Figure 18 shows the secondary network throughput obtained in Scn-2 and Scn-3 (for secondary Poisson traffic rate of 5 frames/sec). We note from these figures that response time of applications as well as average media access delay (shown for primary WLAN server) increase marginally when secondary nodes use modelbased channel access (Scn-3) during last 3 hours of simulation

10 Fig. 13: WLAN Server Average Media Access Delay (on Y axis) for Different Secondary Node Poisson Traffic Rates in Model-based Approach (Scn-3); X-axis: Simulation time in hours Fig. 15: FTP Download Response Time (on Y axis) in all Scenarios (Secondary Node Poisson Traffic Rate 5 Frames/sec; X-axis: Simulation time in hours) Fig. 14: Download Response Time (on Y axis) in all Scenarios (Secondary Node Poisson Traffic Rate 5 Frames/sec; X-axis: Simulation time in hours) time, as compared to the case when secondary nodes use WLAN protocol (Scn-2). On the other hand, the secondary network throughput obtained in Scn-3 is significantly larger (approximately 5 times) than secondary network throughput obtained in Scn-2 (Figure 18). The marginal increase in impact on primary applications performance in Scn-3 is too small to have noticeable effect on the application performance as perceived by the primary end user. This marginal increase is not very unexpected as the model-based scheme in Scn- 3 uses stochastic channel model and is likely to make some Fig. 16: HTTP Page Response Time (on Y axis) in all Scenarios (Secondary Node Poisson Traffic Rate 5 Frames/sec; X-axis: Simulation time in hours) impact on primary network performance because of its modelbased probabilistic estimations. For primary networks, some amount of impact on primary performance can be tolerated, especially when significant secondary network throughput (and consequently, higher channel utilization) can be achieved using model-based scheme. VII. RELATED WORK Model-based approach to spectrum utilization in cognitive radio networks have recently gained a lot of attention in research community. The authors in [6] have proposed a proactive spectrum access approach where secondary nodes take input from spectrum sensing modules, and build predictive

11 Fig. 17: WLAN Server Average Media Access Delay (on Y axis) in all Scenarios (Secondary Node Poisson Traffic Rate 5 Frames/sec; X-axis: Simulation time in hours) Fig. 18: Secondary Network Throughput (in Frames/Sec) for Scn-2 and Scn-3 Secondary Node Poisson Traffic Rate 5 Frames/sec) statistical models of spectrum availability on each channel. Though the goals of this work are similar to ours, the modeling approach is different. The authors have proposed a three-tier model build using the availability statistics from the three different types of observation windows. The overall usability of the channel is computed as the wighted sum of the availability values from all the three observation windows. The authors have also used the concept of usability filter to eliminate unreliable channels with heavy and frequent appearance of primary users. The work on proactive spectrum access reported in [6] has been extended in a work in progress paper [7]. In this paper, the authors have proposed: (i) proactive channel availability prediction by secondary users using the past channel observations, and (ii) intelligent channel switching using the channel availability prediction results. The authors have used renewal theory on past channel observations to estimate the probability that a channel will be idle in the next time slot. This prediction is used to switch channel and avoid collision with any primary user transmissions. The authors have considered both exponential distribution and fixed periodic values for idle and busy time. Our work differs from this work in the sense that we use the alternating renewal process based channel occupancy model to estimate the residual idle time on the channel and transmit back-to-back as many frames as possible in the estimated residual idle time. Alternating renewal theory is also used to analyse how often to sense the availability of licensed channel and in which order to sense those channel [8]. Unlike our work, where the mean residual idle time duration is used to transmit multiple frames on the channel, the authors in [8] have used the concepts of residual busy and idle time durations to devise optimal channel sensing periods. In [9], authors address the issue of opportunistic spectrum access when multiple heterogeneous primary users are active simultaneously. Authors in [1] study the impact of various design options, such as sensing, packet length distribution and back-off time, on opportunistic spectrum access in cognitive radio networks. In [11], authors design strategies that decide, based only on knowledge of the channel bandwidths and data rates, which channels to probe in a multichannel wireless network for opportunistic transmission. The authors do not assume that the user has a priori knowledge regarding the statistics of channel states. Coexistence of cognitive radio based devices and WLAN nodes has recently been studied in [12]. In this work, the WLAN behavior is predicted based on a continuous-time Markov chain model. The cognitive medium access is derived from the CTMC model by casting the channel access problem as a constrained Markov decision process. In a related work reported in [1], the authors construct WLAN behavior model by fitting various distributions to the busy and idle periods recorded by a spectrum analyzer. The authors conclude that Phase type distributions such as Hyper-Erlang distribution show much better fit than widely used exponential distribution. Opportunistic transmission (by slotted secondary network) in unslotted primary networks has also been recently studied in [13] and [14]. There have been several other proposals on MAC protocols, sensing strategies, and channel selection for cognitive radio enabled dynamic spectrum access networks. In [15], authors formulate the problem of optimal sensing decision for a single secondary transmission pair as an optimal stopping problem. Some of the other cognitive MAC protocols reported in the literature are [16], [17], [18], [19], [2], [21] and [22]. VIII. CONCLUSION AND FUTURE WORK The results obtained in this paper indicate that the modelbased approach to medium access has potential to deliver good secondary network throughput without significantly affecting the performance of primary WLAN network. Moreover, the

12 idle channel observation duration value has considerable impact on the performance of model-based secondary network. High secondary throughput can be achieved without affecting the primary network significantly by choosing appropriate value of channel sensing duration. The secondary network performance is better if the model is reasonably accurate. In this paper, we have described the core idea of modelbased channel access mechanism for a single WLAN channel. In our future work, we plan to (i) investigate mechanism to guarantee upper bound on the impact on primary s performance, (ii) make the model online and adaptive based on changing channel occupancy statistics, (iii) extend and check the proposed algorithms for multiple pairs of secondary nodes, and (iii) develop a multi-channel opportunistic MAC protocol based on the channel access mechanism proposed in this paper. Additionally, we plan to study the impact of data outliers on model accuracy. [17] L. Ma, X. Han, and C.-C. Shen, Dynamic Open Spectrum Sharing MAC Protocol for Wireless Ad Hoc Networks, in Proceedings of 1st IEEE Symposium on New Frontiers in Dynamic Spectrum Access Networks (IEEE DySPAN 25), 25. [18] S. Sankaranarayanan, P. Papadimitratos, A. Mishra, and S. Hershey, A Bandwidth Sharing Approach to Improve Licensed Spectrum Utilization, in Proceedings of 1st IEEE Symposium on New Frontiers in Dynamic Spectrum Access Networks (IEEE DySPAN 25), 25. [19] S. Huang, X. Liu, and Z. Ding, Optimal Sensing-Transmission Structure for Dynamic Spectrum Access, in Proceedings of 3rd IEEE Symposium on New Frontiers in Dynamic Spectrum Access Networks (IEEE DySPAN 28), Chicago, USA, October 28. [2] A. W. Min and K. G. Shin, Eexploiting Multi-Channel Diversity in Spectrum-Agile Networks, in Proceedings of IEEE INFOCOM 28), Brazil, 28. [21] C. Cordeiro and K. Challapali, C-MAC: A Cognitive Mac Protocol for Multi-Channel Wireless Networks, in Proceedings of 2nd IEEE Symposium on New Frontiers in Dynamic Spectrum Access Networks (IEEE DySPAN 27), Ireland, 27. [22] H. Nan, T. ln Hyon, and S.-J. Yoo, Distributed Coordinated Spectrum Sharing MAC Protocol for Cognitive Radio, in Proceedings of 2nd IEEE Symposium on New Frontiers in Dynamic Spectrum Access Networks (IEEE DySPAN 27), Ireland, 27. REFERENCES [1] S. Geirhofer, L. Tong, and B. M. Sadler, Dynamic Spectrum Access in the Time Domain: Modeling and Exploiting White Space, IEEE Communications Magazine, May 27. [2] K. S. Trivedi, Probability and Statistics with Reliability, Queuing and Computer Science Applications, 2nd ed. John Wiley and Sons, Inc., 22. [3] A. Feldmann and W. Whitt, Fitting mixtures of exponentials to long tail distributions to analyze network performance models, Performance Evaluation (Elsevier), vol. 31, [4] R. E. A. Khayari, R. Sadre, and B. R. Haverkort, Fitting world-wide web request traces wth the EM-algorithm, Performance Evaluation (Elsevier), vol. 52, 23. [5] Opnet network simulator, [6] P. A. K. Acharya, S. Singh, and H. Zheng, Reliable Open Spectrum Communications Through Proactive Spectrum Access, in IEEE TAPAS, Boston MA, August 26. [7] L. Yang, L. Cao, and H. Zheng, Proactive channel access in dynamic spectrum networks, in Proceedings of CROWNCOM, 27. [8] H. Kim and K. G. Shin, Efficient Discovery of Spectrum Opportunities with MAC-Layer Sensing in Cognitive Radio Networks, IEEE Transactions on Mobile Computing, vol. 7, no. 5, May 28. [9] E. Jung and X. Liu, Opportunistic Spectrum Access in Heterogeneous User Environments, in Proceedings of 3rd IEEE Symposium on New Frontiers in Dynamic Spectrum Access Networks (IEEE DySPAN 28), 28. [1] S. Huang, X. Liu, and Z. Ding, Opportunistic Spectrum Access in Cognitive Radio Networks, in Proceedings of IEEE INFOCOM 28, 28. [11] N. B. Chang and M. Liu, Competitive Analysis of Opportunistic Spectrum Access Strategies, in Proceedings of IEEE INFOCOM 28, 28. [12] S. Geirhofer, L. Tong, and B. M. Sadler, Cognitive Medium Access: Constraining Interference Based on Experimental Models, IEEE Journal on Selected Areas in Communications, vol. 26, no. 1, January 28. [13] S. Shetty, M. Song, C. Xin, and E. K. Park, A Learning-based Multiuser Opportunistic Spectrum Access Approach in Unslotted Primary Networks, in Proceedings of IEEE INFOCOM, 29. [14] Q. Zhao and K. Liu, Detecting, Tracking, and Expoiting Spectrum Opportunities in Unslotted Primary Systems, in Proceedings of IEEE Radio and Wireless Symposium (RWS), 28. [15] J. Jia, Q. Zhang, and X. S. Shen, HC-MAC: A Hardware-Constrained Cognitive MAC for Efficient Spectrum Management, IEEE Journal on Selected Areas in Communications, vol. 26, no. 1, January 28. [16] Q. Zhao, L. Tong, A. Swami, and Y. Chen, Decentralized cognitive MAC for opportunistic spectrum access in ad hoc networks: A POMDP framework, IEEE Jornal on Selected Areas in Communications: Special Issue on Adaptive, Spectrum Agile and Cognitive Wireless Networks, vol. 25, no. 3, April 27.