Cognitive Radio Implementation in ISM Bands with Microsoft SORA_PIMRC_11

2011 IEEE 22nd International Symposium on Personal, Indoor and Mobile Radio CommunicationsCognitive Radio Implementation in ISM Bands with Microsoft SORA∗ DepartmentJitin Bajaj∗ , Wooseong Kim∗ , Soon Y. Oh† and Mario Gerla∗of Computer Science, University of California, Los Angeles, USA Email: {jbajaj,wooseong, gerla}@ † UtopiaCompression Corp. 11150 Olympic Blvd. Suite 820 Los Angeles, CA 90064 Email: soon@Abstract—Cognitive radio network research has mostly focused on licensed spectrum sensing and empty spectrum use while protecting licensed band primary users. However, the wide deployment of WiFi access points in unlicensed bands in urban areas brings difficult challenges regarding co-existence of ad hoc mobile users with residential access points. In this paper, we propose a new mobile user protocol for dynamic channel selection via channel workload estimation. The mobile node senses channel usage in unlicensed ISM band and selects the maximum residual capacity channel for own packet transmission. The new protocol is implemented on Microsoft’s SORA software defined radio system that enables fast prototyping. We have evaluated the performance using a Campus deployed SORA testbed. Experimental results show the efficiency of our approach in avoiding external interference. Index Terms—Software defined radio, Cognitive radio, Spectrum sensingI. I NTRODUCTION As smart mobile devices have become prevalent, WiFi network usage has increased multifold. Many public places now provide free WiFi services and lots of wireless routers have been installed in offices and residencies. Thus, unlicensed bands where 802.11 a/b/g WiFi operates are always heavily loaded. This creates critical challenges for mobile users who want to set up a temporary MANET in an urban environment to support short lived applications such as video streaming among vehicles in a motorcade [1] or P2P content sharing among smart phone users. As mobile P2P applications (as opposed to traditional client server applications that access the Internet) are becoming increasingly popular, the coexistence of mobile P2P users and Infrastructure WiFi users (thereafter called “primary” users) is becoming problematic. However, in reality, wireless channel usage is concentrated in only a fraction of the channels in unlicensed ISM bands (i.e., 2.4.GHz and 5GHz) as shown in Figure 8. Some heavy traffic loaders capture those channels and many wireless APs in the same area are using the same channels. In other words, channels are poorly utilized and interference occurs even though there are unused channels. If traffic were fairly distributed over the multiple channels, channel utilization would significantly improve and the unlicensed ISM band congestion problem would be mitigated. Mobile users can exploit this inefficient distribution to their advantage, as discussed next.Conventional cognitive radios have been proposed and standardized in IEEE 802.22 WRAN [2] to utilize “empty spectrum holes” in licensed spectrum bands, namely TV bands, without interrupting licensed users or primary users [3]. However, we have proposed novel cognitive networks such as CoCast [1] and CoRoute [4] that exploit cognitive radios to find empty channels in ISM bands instead of scavenging the white spaces in such licensed bands. To achieve more even load distribution, they dynamically select the proper channel with low or no traffic among multiple channels by ISM band scanning. We have evaluated their performance gain in terms of scalability and capacity via simulation and modeling studies in [1], [4]. However, the practical issue remains that ISM band cognitive radio systems have not yet been developed and tested. Therefore, to extend the previous research, in this paper we examine the feasibility of intelligent channel sensing and selection in ISM band. Namely, we design and develop a cognitive wireless network system using WiFi radio in ISM band that allows nodes to sense the channel and to calculate channel workload (i.e., channel utilization). During fixed time intervals, a node measures signal power level for the received samples and classifies it as a busy sample if the power level exceeds a given threshold; or as an idle sample otherwise. Finally, workload is estimated by the number of busy samples over the total number of samples. We have implemented this system in a Software Defined Radio (SDR) that is fully programmable and ran multiple experiments to verify its performance. The contribution of this paper is two-fold: First, we develop intelligent channel sensing using received energy level measurements and workload calculations in an SDR system. Thus, we prove the feasibility of cognitive radio systems in ISM bands. Next, we experimentally prove that ISM band is not fully utilized and propose a solution that efficiently utilizes ISM band channels exploiting channel sensing and intelligent channel selection. The remainder of this work is organized as follows: Section II describes the SORA platform and Section III explains channel sensing and selection algorithm and the implementation details. Evaluation results are shown in section IVand Section V reviews the related works. Section VI concludes the paper.978-1-4577-1348-4/11/$26.00 ©2011 IEEE531II. SORA STRUCTURE SDR is a radio communication system in which elements are implemented by means of software instead of hardware, e.g., modulators and demodulators, amplifiers, filters, mixers, etc. It allows users and applications to access software physical or link layer, which have been primarily implemented on fixed and customized hardware circuits, and thus enables users to develop protocols easily. Such instant protocol implementation is not possible with traditional radio devices. Furthermore, SDR enables users to modify or to develop lower layers, e.g., link and physical layer protocols while it is a very complicated and difficult process to modify lower layers with traditional radio devices. Because of flexibility and interoperability, SDR starts to replace complex radio communication systems such as Join Tactical Radio Systems (JTRS) program in the US military. Most current SDR platforms are based on either programmable hardware such as Field Programmable Gate Arrays (FPGAs) or embedded Digital Signal Processors (DSPs). Such hardware platforms are able to satisfy the processing and timing requirements of modern high-speed wireless protocols, but there is limitation due to difficulty in FPGAs and DSPs programming. In contrast, SDR platforms based on generalpurpose processor (GPP) architectures, such as commodity personal computers (PCs), provide simple and familiar developing environment and easy-to-use tools to developers. However, existing GPP based SDR platforms only achieve limited performance due to PC hardware constraints (e.g., delayed wireless signal processing). Microsoft SORA is a GPP based SDR and thus it is comfortable to use while it is as powerful as FPGA and DSP based SDRs. We use the SORA SDR system to implement the ISM band cognitive radio system. Figure 1 illustrates the overall architecture of the SORA system. Software Radio Stack (that is left-hand side of the figure 1) is part of the PC and it consists of CPU and memory. The right-hand side of the figure 1 illustrates an RF front-ends. The PC and the RF front-ends are connected through a Radio Control Board (RCB) that is attached to the PC via a PCI-E interface and reads/writes digital signal samples from/to PC memory using Direct Memory Access (DMA). The RCB and the RF front-end board are connected by SORA Fast Radio Link (SoraFRL) where a protocol for the control of the RFend board is defined. Thus, any RF boards that are compatible with the RCB must implement SoraFRL [5], [6]. Figure 2 describes the user and kernel mode structure of SORA. A low-level RCB driver and a set of programming libraries provide SORA services. The RCB driver (i.e., Radio Control Board Driver in figure 2) manages RCB hardware resources in order to transmit and receive data. In addition, the driver controls the RF front-end connected to the RCB. Network applications, i.e., user-mode programs and kernel mode application drivers, access RCB and RF front-end resources via APIs provided by SORA SDK. Say, the kernel application drivers use SORA Core API (i.e., kernel mode API) and the user-mode applications use SORA User-mode Extension API.Fig. 1: SORA overall system architecture figure from [6].Fig. 2: SORA user and kernel mode architecture figure from [6]. Figure 3 depicts a typical SORA SDR application driver architecture that implements the IEEE 802.11 radio structure. The SORA application driver normally acts as a virtual Ethernet driver located under the Windows NDIS framework. Thus, upper layers, for example TCP/IP, view it as a normal Ethernet card so that no modification of existing network applications is required while using SDR network. Link, MAC, and physical layer should be implemented on the SORA application driver. The link layer converts and encapsulates frames. For instance, if a user wants to implement 802.11 on SORA, Ethernet frames for upper layers must be converted to 802.11 frames before sending and after receiving at the link layer. The MAC layer operates according to a finite state machine (FSM) to handle media access. To support users’ MAC programming, SORA provides FSM APIs. The physical layer (PHY) includes all baseband signal processing implementations that support modulation, demodulation, channel estimation, channel coding, etc. SDR_CONTEXT (lefthand side in figure 3) provides cross-layer message passing over layer boundaries. It has pointers to other required data structures and is typically used as the sole parameter for many SORA library APIs. III. SPECTRUM SENSING AND CHANNEL SELECTION Most of the researches on cognitive radios have focused on spectrum sensing, spectrum management, and interference prediction in the TV bands. To support the opportunistic type532Fig. 4: Energy detection mechanism in time domain.Fig. 5: Energy detection mechanism in frequency domain. signals and sine waves, compare to the frequency domain mechanism. In these mechanisms, the decision statistics for energy detection can be represented,N −1Fig. 3: SORA application driver figure from [6]. of channel access in licensed bands, channel sensing is a very crucial functionality. Furthermore, various licensed users have different spectrum assignments and access policies (i.e., different power level, modulation schemes, various propagation environments, etc) so that high quality spectrum sensing devices and algorithms are required. Several different spectrum sensing techniques have been proposed in literatures and have been implemented in experimental and commercial systems. Among them [7], we mention: • Energy Detection: it measures RF energy or RSSI to test occupancy of a channel. It is simple to implement and does not depend on the prior knowledge of primary signal. However, it cannot distinguish between primary signal and other secondary users without quiet period among secondary users. • Matched Filter : it compares a known signal or template with received signals to detect the present of primary users. It has low computational cost, but a prior knowledge, i.e., primary user signal information, is required. • Cyclostationary Feature Detection: it distinguishes primary user signal from noise using periodicity within the modulated signal. The detection is robust to random noise and interference, but it incurs high computation cost. • Wavelet Detection: it models the entire wideband as a train of consecutive frequency sub-band and then detects the primary signal. This scheme also requires high computation and high sampling rates. Among these techniques, we employ energy detection for SORA implementation since it is simple and does not require prior knowledge of primary signal. The implemenation is based on Neyman-Pearson approach [8]. The detector compares the calculated energy of the received signal and threshold value and decides whether external traffic exists [9]. The energy detection mechanism is represented by both time (i.e., figure 4) and frequency domain (i.e., figure 5). Both mechanisms employ Nyquist sampling A/D converter, squarelaw device and integrator, but the time domain mechanism requires a pre-filter matched to the bandwidth of the signal. Theoretically, results from both schemes should be the same, but in the time domain mechanism, the pre-filter implementation makes the method inflexible, particularly for narrowbandT=n=0∑ |x(n)|2(1)where n is the number of the samples, such as n = 0, 1, 2, . . . , N − 1, N . The decision of energy detection is the test of the following two hypotheses. H0 : x[n] = w[n] (2) H1 : x[n] = s[n] + w[n] (3)w[n] is assumed additive, white and Gaussian noise (AWGN) and s[n] is external signal. In here, H0 is null hypothesis that represents the received signal has only noise while H1 means that the received signal comprises of signal and noise. Say, formula (2) indicates signal absent and thus the decision value from formula (1) is smaller than the threshold value γ : otherwise, signal is present and the decision value is larger than the γ . Based on channel sensing information, channel workload is calculated. To do this, nodes periodically perform channel sampling and use the sensing results as a key input metric for channel workload calculation. If the sample satisfies formula (3), it is a busy sample; otherwise, it is idle. The channel workload is calculated as a ratio of the samples measured above threshold and total number of samples counted as follows: w(c) = Number o f busy samples Total number o f samples monitored (4),where c is channel and workload is denoted as w(c). The channel sensing is done in two different modalities. First is full scanning where nodes scan all channels and record a channel workload history, which keeps track of the amount of workload measured in each channel. It is repeated periodically over a long interval in order to reduce scanning overhead. Next one is in-band scanning where nodes scan and update workload only for the channel that nodes select for communication, which is performed more frequently. These sensing modalities are illustrated in the Figure 6. At the beginning of each sensing period, a node performs the full spectrum sensing over all n channels (i.e., n = 11 for 802.11b/g) and updates the channel workload history. Based533Fig. 6: Spectrum Sensing CyclesFig. 8: Average Workload that is averaged over 9 full spectrum scan cycles with scan time for a single channel 4 seconds. node 1 runs the channel sensing and selection algorithm while node 2 and node 3 mainly act as traffic generators to create interference. In the first experiment, we measure channel workload in a normal university laboratory environment at night with minimal external interference and without any artificial traffic generation. Channels are sensed during nine full spectrum scans; a single channel is scanned for 4 seconds. Channel workload is averaged in terms of percentage of channel busy time and it is reported in Figure 8. Channels 2 (2.417 GHz), 3 (2.422 GHz) and 4 (2.427 GHz) have over 10% channel workload and Channels 1 (2.412GHz), 5 (2.432 GHz) and 6 (2.437GHz) are fairly medium loaded while Channels 7 (2.442 GHz), 8(2.447 GHz), 9 (2.452 GHz), and 10 (2.457 GHz) have very low workload in most of the experiments. The reason of low overhead is that most desktops in the lab are using the wired network and there were only a few students in the lab at nighttime when the workload was measured. Next, to verify the workload estimation accuracy, node 2 and 3 start generating traffics, i.e., 30%, 50%, and 70% workload, using Iperf. To remove interference from outside of the experiment, we use channel 8, which has 0.028% workload during the previous experiment, and node 1’s channel is fixed to channel 8. The values are averaged over seven full spectrum scanning in which single channel is sensed for 3 seconds. Figure 9 presents averaged workload measured at node 1. It estimates 28.2%, 40.5% and 59.7% workload respectively so that 2 ∼ 10% errors occur. While traffic increases, workload calculation error becomes larger. Now interference is fixed at 50% on channel 8 and workload is measured during eight full spectrum scans with 3 second single channel sensing. Calculated workload is always less than the actual value as shown in Figure 10. Measured minimum, average, and maximum values are 36%, 40.5%, and 48%, respectively. Figure 11 reports measured workload in terms of different scan times for a single channel. 50% workload traffic is generated on channel 8 and workload is averaged over eight full spectrum scanning cycles while single channel scan times varies 2, 3, and 4 seconds. When the single channel scanFig. 7: MAC Finite State Machine on the calculated channel workload, the node selects the best channel cbest for communication, where cbest = argminc w(c). The node continuously monitors and computes the channel workload of cbest . If its workload increases and thus the minimum workload channel is changed, the node switches its own channel. This process is illustrated in Figure 6. First, the minimum workload channel is channel 10, but after time t , the best channel switches to channel 2. Now the node moves to the new channel, that is channel 2. Aforementioned sensing scheduling, full sensing with long interval and in-band sensing with short interval, can reduce overhead, but it decreases workload calculation accuracy. A node is not able to keep up with the workload changes in other channels so that it is not easy that the node finds the minimum workload channel on the fly. To mitigate this problem, the full sensing period and the threshold must be carefully adjusted. In addition, cooperative sensing, where nodes exchange sensing information with each other, is also desirable to get other spectrum sections’ workloads and to avoid channel selection conflicts. Figure 7 briefly describes channel sensing and selection process via FSM (Finite State Machine) at MAC. The system starts at the CHannel Sensing and Selection (CHSS) state where full spectrum sensing is performed and after completing the sensing, a node selects the best channel. After then, a node reaches the Carrier Sensing (CS) state and normal communication is proceeding on the selected channel. Upon reaching timeout, a node goes back to the CHSS state to repeat full channel scanning. IV. E VALUATION We have implemented the energy detection and workload calculation algorithm on a Microsoft SORA. Our implementation is added on top of SORA standard 802.11 b/g implementation so that it is compatible with normal 802.11 b/g interfaces. To evaluate our implementation, we use a small testbed consisting of three SORA nodes. During experiments,534Fig. 9: Average Measured Workload v/s Actual Workload for single channel scan time 3 seconds.tion pairs using control messages while SSCH uses channel hopping sequence to have a rendezvous channel with neighbor nodes to exchange packets in certain time slot. CoCast (Cognitive Multicast) [1] and CoRoute (Cognitive Unicast) [4] are two realizations of cognitive ad hoc networks, in which each node dynamically selects the most effective channel among multiple channels based on channel sensing. They have a common control channel as a time slot or a designated spectrum to exchange channel information among nodes. We have run a CoCast simulation experiment with 100 vehicles, 2Mbps WiFi channels. In this experiment, delivery ratio (packets received / packet sent) results report performance gain from 35% to 75% compare to conventional routing protocol, ODMRP, in a mobile environment. CoRoute also shows 30% improvement in terms of delivery ratio using a cognitive approach. VI. C ONCLUSION This project highlighted the importance of software defined radios for wireless networking research. We have used Microsoft SORA radios, arguably the state-of-the-art platforms based on GPPs. We designed and implemented Intelligent Spectrum Sensing and channel selection based on a channel workload calculation algorithm. Results were very encouraging and show great promise for ad-hoc networks. The next step will be to coordinate the best channel information between various SORA nodes. This will enable us to measure the throughput performance gains of our cognitive protocol. In the future, we also plan to test the spectrum sensing protocol for VANET environment. Finally, a more extensive evaluation may reveal additional information that can be used for further optimization of the protocol. R EFERENCES[1] W. Kim, S. Oh, and M. Gerla, “Cocast: Multicast in mobile ad hoc network using cognitive radio,” in in Proc. of IEEE MILCOM, 2009. [2] “Ieee 802.22 wg on wireless regional area networks,” in /22/, 1999. [3] J. Mitola and G. Q. Maguire, “Cognitive radio: Making software radios more personal,” IEEE Personal Communications, vol. 6, no. 1, pp. 13– 18, 1999. [4] W. Kim, S. Oh, K. Lee, and M. Gerla, “Coroute: A new cognitive anypath vehicular routing protocol,” in in Proc. of IEEE IWCMC, 2011. [5] K. Tan, H. Liu, J. Zhang, Y. Zhang, J. Fang, and G. M. Voelker, “Sora: high-performance software radio using general-purpose multi-core processors,” Commun. ACM, vol. 54, pp. 99–107, January 2011. [Online]. Available: /10.1145/1866739.1866760 [6] “Microsoft sora manual,” in /apps/pubs, 2009. [7] V. K. Bhargava and E. Hossain, Cognitive Wireless Communication Networks. Secaucus, NJ, USA: Springer-Verlag New York, Inc., 2007. [8] S. Kay, Statistical Signal Processing, Detection Theory. Prentice Hall, 1998. [9] S. Maharjan, K. Po, and J. ichi Takada, “Energy detector prototype for cognitive radio system,” IEICE Technical Report, SR2007-52, vol. 107, no. 352, pp. 43–48, 2007. [10] J. So and N. Vaidya, “Multi-channel mac for ad hoc networks: handling multi-channel hidden terminals using a single transceiver,” in In Proc. of ACM MOBIHOC, 2004. [11] P. Bahl, R. Chandra, and J. Dunagan, “Ssch:slotted seeded channel hopping for capacity improvement in ieee 802.11 ad-hoc wireless networks,” in In Proc. of ACM MOBICOM, 2004.Fig. 10: Workload of different spectrum scan cycle and average workload. Red color bars are average value of 8 full scans and green bars are value of each scan cycle.Fig. 11: Average Measured Workload v/s Single channel scan times for an actual workload of 50% time is 4 seconds, measured workload is 42% so that 8% error exists. However, when scan time reduces to 2 seconds, deviation increases to 12.6% and thus measured workload becomes 37.4%. As we expected, longer sensing time increases workload accuracy. However, still calculated workload is always less than actual value. V. R ELATED W ORK Channel diversity to increase the capacity of wireless networks with a single radio (e.g., MMAC [10] or SSCH [11]) requires channel synchronization among nodes; for example, MMAC exchanges channel information between communica-535。