Cognitive Wireless Networking Kang G. Shin Real-Time Computing Laboratory EECS Department The University of Michigan Ann Arbor, MI 48109-2121 http://www.eecs.umich.edu/~kgshin Today’s Wireless Networking Exponential growth of wireless access demands ▪ Multimedia & other QoS applications ▪ Diverse network uses – commercial, public, military Wireless medium “Paradigm shift” in network design ▪ Static, environment/app-agnostic dynamic and adaptive 2 Cognition: key to future networking What is cognition? ▪ Awareness of surrounding environment and apps, which are often subject to: ▪ Random noise, fading, heterogeneous signal attenuation ▪ Diverse app types and criticalities Why cognition? ▪ Spectrum is a limited resource ▪ Traditional network designs are not efficient ▪ New research directions, e.g., Dynamic Spectrum Access ▪ DSA requires cognition ▪ One-fits-all doesn’t apply 3 Elements of Cognition Spectrum Sensing ▪ Monitors signal activities ▪ Detects signals ▪ Energy or feature detection Environmental/App Learning ▪ Learns network dynamics and app requirements ▪ Channel quality and usage patterns (e.g., ON/OFF, SNR) ▪ Apps needs (e.g., delay, bw, jitter) System/App Adaptation ▪ Adapts system/app configurations/parameters ▪ Adapts sensing period/time/frequency, stopping rule, etc. 4 What to Expect from Cognition? Technically, ▪ Efficient spectrum utilization ▪ Smarter spatial reuse ▪ Coexistence of heterogeneous networks Economically, ▪ Extra benefit to legacy users ▪ Opportunistic spectrum auction/leasing ▪ Cheaper service to CR users ▪ CR Hotspots – cheaper Internet access CR HotSpot 5 Software-Defined Radios (SDRs) Key to cognition! ▪ Reconfigurable in real time (e.g., USRP, SORA, WARP) Today’s SDR Devices USRP1 USRP2 SORA Throughput 400 bytes/s 16 kbps 15 Mbps Bandwidth 2Mhz 768 kHz Standard 802.11g PHY CSMA Pulse-modulated CDMA Interconnect USB OFDM Gigabit Ethernet PCI ▪ Different PHY layers cannot account for the throughput differences ▪ Slow USB interface results in significant lag between carrier sense and transmission ▪ PHY and MAC layers need to tolerate processing delays 6 Cognition-based Network Design Cognition Engine Integration Architecture of Cognition Elements with Legacy Systems Cognition Engine Cognition Engine Better QoS Support Enhanced Utilization Environment/App Monitoring Environment/App Learning System/App Adaptation Optimal Decision Includes key elements in achieving awareness Enables unified cognition for wireless networks 8 Environment/App Monitoring Signal detection ▪ PHY-layer monitoring of signal activities Adaptive selection of method for signal detection ▪ Energy detection – more sensitive to SNRwall ▪ Feature detection – usually longer sensing-time Monitoring of application QoS needs ▪ Applications can provide QoS hints, e.g., bandwidth, e2e delay, jitter 9 Environment/App Learning Spectrum-usage pattern inference ▪ Infer ON/OFF channel-usage patterns ▪ Methods: ML, Bayesian, and entropy-based estimation Signal profiling ▪ Based on received signal strength (RSS) Application QoS estimation & prediction ▪ Applications may have stringent & diverse QoS needs ▪ History-based estimation/prediction using explicit hints and network-state awareness 10 System/App Adaptation Spectrum-sensing scheduling ▪ Policy-aware: ▪ Meet FCC’s requirement on sensing for primary user protection ▪ Bandwidth-aware: ▪ Maximal or fast discovery of idle channels Spectrum-aware user admission/eviction control ▪ Commercial CR Access Points ▪ Multiple user classes (with different spectrum demands) ▪ Time-varying spectrum resources (ON OFF) ▪ Optimal user admission/eviction control ▪ To maximize profits 11 System Adaptation, cont’d Application-aware DSA optimization ▪ DSA parameters (e.g., sensing time & interval) are adaptively updated based on applications’ QoS demand Collision-aware transmission scheduling ▪ Collision resolution, instead of collision avoidance DSA transmission scheduling ▪ Goal: Achieve good PU-safety vs. SU-efficiency tradeoff ▪ Dual (safe vs. aggressive) mode transmission scheduling based on PU channel-usage pattern estimation 12 Optimal Decision Existence of opportunities ▪ H0: no primaries exist there are opportunities ▪ H1: primaries exist no opportunity Reliable distributed sensing ▪ Attack-tolerant cooperative sensing ▪ Protect sensor networks from (e.g., spoofing) attacks ▪ Detection/filtering of abnormal sensing reports ▪ Mal-functioning or compromised sensors 13 System Integration Architecture Implementation & deployment of cognition ▪ Needs a well-defined integration architecture ▪ Different from traditional (full layer-based) design 14 System Integration Arch, cont’d Integration architecture consists of: Cognition Interface (CI) ▪ Provides interface API to each cognition mechanism ▪ Seamlessly integrates with OS protocol stack, applications, and other cognition mechanisms Cross-layer Interaction Framework (CLIF) ▪ Provides “awareness” management in system/network ▪ Consists of Repository, Parameter Mapper, and Trigger Manager 15 Cognition Interface Defines communication mechanisms between cognition engine and existing network stack API functions provided for ▪ Export/import & management of awareness parameters ▪ Registering trigger events 16 Cross-Layer Interactions Provides abstraction for cognition protocol implementation & deployment Consists of: ▪ Repository - stores awareness parameters ▪ Trigger Manager - registers predicates of parameters, ▪ and generates notification events Parameter Mapper - manages routines that define relationship between awareness parameters 17 Cognitive Networking Research in RTCL at Michigan CR Components & Architecture Maximal Opportunity Discovery via Periodic Sensing Fast Opportunity Discovery via Periodic Sensing Incumbent Protection via In-band Sensing Optimization Framework for Cooperative Sensing Attack-Tolerant Distributed Sensing in CRNs Spectrum-Aware User Control Collision-Aware Transmission Scheduling Context-Aware Spectrum Agility (CASA) Spectrum-Conscious WiFi (SpeCWiFi) System Integration Architecture (SIA) CNR Group @RTCL • Current Members PhD students: Eugene Chai, Hyoil Kim, Ashwini Kumar, Alex Min, Michael Zhang, Xinyu Zhang Post docs: Jaehuk Choi • Recent Alums PhD graduates: Chun-Ting Chou Post docs: Young-June Choi, Bechir Hamdaoui CR Components & Architecture Main Components RME: Resource Management Entity MME: Measurement Management Entity GCE: Group Coordination Entity PEE: Policy Enforcement Entity Resource Management Entity (RME) Maintains Spectral Opportunity Map (SOM) Status of each channel SOM is updated by scanning (MME) and exchanging SOMs (b/w RMEs) 20 GCE PEE RME MLME (MAC) MME PLME (PHY) CR Components & Architecture, cont’d Group Coordination Entity (GCE) Synchronize channel vacation Exchange spectrum-usage information Described by three states SCAN: scan a channel (MME) LISTEN: check returning incumbent (MME) VACATE: vacate channel (GCE) VACANCY VACATE GCE 3 states SCAN 21 LISTEN Maximal discovery via periodic sensing Sensing-time TIi Sensing-period TPi Ch 1: Ch 2: Ch 3: sensing: logical ch: 1 Periodic sensing 3 1,3 1 1,3 1 Discovered opportunities 1 2 2 2 Disrupted reuse 1,2 2 3 3 time Find optimal Tpi ’s – Tradeoff b/w discovery & disruption: ▪ Frequent sensing (1) more idle channels discovered, but (2) more disruption in utilizing opportunities 22 Performance Evaluation Discovered ≥98% of the analytical maximum (AORmax) ≤22% more opportunities than non-optimal schemes 23 Fast discovery via reactive sensing Reactive sensing – discover opportunities at channel vacation reactive sensing ON reused channel channel vacation opportunity found Ch 1 OFF Ch 2 Ch 3 Opportunity discovery latency seamless service provisioning Find: optimal sensing sequence for minimal latency 24 Optimal sensing sequence At channel vacation: ▪ N out-of-band channels ▪ Capacity Ci ▪ Pidlei : channel availability (probability of idleness) ▪ B : amount of bandwidths to discover at channel vacation N! possible sequences (NP-hard) Homogeneous case (Ci=C) optimal sequence Sorting channels in ascending order of TIi / Pidlei Heterogeneous case suboptimal sequence Satisfying necessary condition for optimality 25 Backup channel management Goal: manage a list of backup channels ▪ A subset of out-of-band channels channel export Backup Channel List (BCL) Candidate Channel List (CCL) channel swap out-of-band channel channel import Q1: How to form BCL Initially? Q2: How/When to update BCL? 26 Performance Evaluation (1) Optimal Sensing Sequence (2) BCL Update 76% 47% (enhanced) 91% 40% Delay Type-I: opportunities discovered at first round search Delay Type-II: opportunities discovered at successive retries 27 Incumbent Protection via In-band sensing GOALS Broadband wireless access in rural area 1) Protect incumbents (DTV, uPhone) Detectability requirements: IDT, CDT, PMD/PFA 2) Promote QoS (for CPEs) Minimal sensing overhead TV transmitter WE PROPOSED (MobiCom’08) CPEs 1) 2-tiered clustered sensor networks To support collaborative sensing Maximal cluster size (radius) Maximal sensor density 2) In-band sensing scheduling algorithm Optimal sensing-time Optimal sensing-period Better detection method: (energy vs. feature) BS 33(typical) -100km 28 Performance Evaluation minimal sensing overhead Results 29 Energy detection vs. Feature detection, applying optimal sensing time/period aRSSthreshold: avg. RSS, above which energy detection is better aRSSenergymin: avg. RSS, above which energy detection is feasible, to overcome SNRwall Optimization Framework for Cooperative Sensing GOAL To detect the existence of a primary signal as fast as possible with high detection accuracy with minimal sensing overhead KEY IDEA Exploit spatio-temporal variations in received primary signal strengths (RSSs) among sensors HOW? Optimal sensor selection Use sensors with high performance RSS-profile-based detection rule Base station manages spatial RSS profile of sensors Measured RSSs are compared to the profile Optimal stopping time for sensing Sequential analysis based on measured RSSs 30 Spatio-Temporal Diversity in RSSs OBSERVATIONS Location-dependent sensor heterogeneity Temporal variations due to measurement error How to select sensors and schedule sensing? 31 Optimal Sensing Framework SEQUENTIAL HYPOTHESIS TESTING PROBLEM Find an optimal set of cooperating sensors At each sensing period n, update decision statistic Λn, compare it with predefined thresholds Stop scheduling sensing when Λn reaches the thresholds Minimize sensing overhead while guaranteeing the detection requirements 32 Performance Evaluation SENSING SCHEDULING SENSOR SELECTION Reduce sensing while meeting detectability requirement Sensor selection further reduces the sensing overhead 33 Attack-Tolerant Distributed Sensing in CRNs THREAT Malicious/malfunctioning sensors can manipulate sensing results, thus obscurinhg the existence of a primary signal Waste of spectrum opportunities (Type-1 Attack) or excessive interference to primaries (Type-2 Attack) CHALLENGE Openness of PHY/MAC layer in SDR devices No cooperation between primary and secondary networks OBJECTIVE To withstand falsified sensing reports from malicious or faulty sensors KEY IDEA Leverage spatial RSS correlation due to shadow fading to filter abnormal sensing reports 34 Spatially-correlated shadow fading REMARKS RSSs are spatially-correlated under shadow fading Large deviations can be easily detected Form sensor clusters among sensors in proximity and cross-check validity of the reports 35 Attack-Tolerant Sensing FRAMEWORK MAIN COMPENENTS Sensing manager: manages sensor cluster and schedule sensing periods Attack detector: detects and discards abnormal sensing reports Data-fusion center: decides on existence of a primary signal 36 Anomaly Detection CORRELATION-BASED FILTER Derive conditional pdf of neighbors’ sensing results Cross-checks the abnormality of neighboring sensors’ reports If sensor i’s reports is flagged by more than x % of its neighbors, regard it as abnormal and discard/penalize it in the final decision 37 Performance Evaluation TYPE-1 ATTACK TYPE-2 ATTACK Successfully tolerates both type-1 and type-2 attacks 38 Spectrum-Aware User Control CR HotSpots – commercial CR APs ▪ Provide wireless access (e.g., Internet) ▪ Lease channels from PUs (for opportunistic reuse) ▪ Time-varying channel availability (ON or OFF) Goal: profit maximization ▪ Optimal admission and eviction control of CR end-users ▪ Eviction (at OFFON): which user to evict from the service? ▪ Approach: Semi-Markov Decision Process (SMDP) Departure from service CR HotSpot 39 ON OFF ON OFF Leased Channels User arrivals Performance Evaluation Test Conditions ▪ Channel capacity = 5 ▪ 2 channels ▪ 2 user classes (1) nk: # of class-k users in service (2) Spectrum demand = k Observation ▪ No threshold behavior (unlike in time-invariant resources) ▪ Intentional blocking of arrivals (unlike in complete-sharing) 40 Collision-aware Transmission Scheduling Iterative collision resolution (PHY layer) Cognitive sensing and scheduling (MAC layer) ▪ Sense the identity of the packet in the air (PA) ▪ Transmit if PA has the same identity (seq and session id) as the packet to be sent 41 Performance Evaluation In comparison with DCB, a CSMA/CA based broadcast protocol ▪ PDR and delay in lossy wireless networks: 42 Performance Evaluation, cont’d ▪ PDR and delay as a function of source rate (indicating maximum supportable throughput) 43 Context-Aware Spectrum Agility (CASA) CASA is composed of: ▪ Application Monitoring element ▪ Application QoS Estimation & Prediction element ▪ Application-aware DSA optimization CASA provides history-based DSA protocol optimization of DSA protocol parameters, ▪ e.g., reduce scanning duration according to e2e delay constraint CASA improves SU QoS fulfillment by ≥35% 44 Spectrum-Conscious WiFi (SpeCWiFi) SpeCWiFi consists of: ▪ Spectrum Sensing ▪ Spectrum-usage Pattern Estimation ▪ DSA Transmission Scheduling Preliminary evaluation on a madwifi-based testbed SpeCWiFi manages to keep PU interference low (<3%), while keeping SU utilization high (>94%) on avg. 45 System Integration Architecture (SIA) SIA implemented in Linux kernel ▪ Repository and Trigger Manager implemented as ▪ loadable kernel modules Dynamic hash-tables used for data management Cognition Interface implemented as DLL For user-level applications, Application Adaptation Layer (AAL) implemented to minimize user-kernel crossings Evaluation shows overhead to be minimal (~1¹s) for networking system calls 46 Conclusion Cognition-based network design is key to the nextgeneration wireless networking ▪ Dynamic spectrum resource management ▪ Environment/app-awareness Two directions in Cognition-based design ▪ Cognition Engine – 4 elements to achieve awareness ▪ Integration Architecture – for compatibility with legacy systems Still have a long way to go… http://kabru.eecs.umich.edu/bin/view/Main/RtclPapers 47