Real-Time Sensor Networks with Applications in Cyber-Physical Systems

Real-Time Sensor Networks with Applications in Cyber-Physical Systems Manimaran Govindarasu Dept. of Electrical and Computer Engineering Iowa State University Ames, IA 50011, USA gmani@iastate. edu http: //ecpe. ece. iastate. edu/gmani

TALK OUTLINE n System-level Energy Management n End-to-End Energy Management n Cyber-Physical System applications n Conclusions 2

Embedded System Battery Embedded Device • Processor - computation • Network Interface -communication • Others: • Memory • I/O Energy is the most important resource It needs to be managed efficiently 3

Sensor Net Applications Wireless Industrial Networks Border Security Traffic monitoring system • Sense • Encrypt • Decrypt • Aggregate • Communicate 4

Wireless Sensor Network (WSN) – Data Aggregation Tree Model i (compute(Ti), Communicate(Mi)) En d- to - en d d ea dli ne Root/sink Leaf Nodes (sense, compute, Communicate) 5

WSN – Mesh network model Energy Management at the Computing Subsystem (considering all the tasks) C D Energy Management at the Communication Subsystem considering all the messages) A Computation Wireless Network + Communication B Energy Management at the E System-level (both messages & tasks) F G 6

WSN Challenges Limited Processing power and memory Low energy low data rate Wireless Interference Constraints Real-time Unreliable Measurements Requirements

State-of-the-Art in Energy Management Schemes Duty Cycling Sleep-Wakeup MAC with duty cycling Joint Scheduling Tasks & Msgs Data Driven Data Reduction Energy efficient Data Acquisition Online Adaptation

System-level Energy Management in Networked Real-Time Embedded Systems Computing subsystem Cross-Layer Energy-aware Task Scheduling (DVS, DPM) Communication subsystem Energy-aware Message Scheduling (DMS; Power Adaptation) System-level (Comp. + Comm. ) Energy-aware System-level Scheduling (DVS + DMS) • Single-hop • Multi-hop 9

Embedded Device Energy Model CPU Energy consumption Vdd – supply voltage f – CPU frequency CC – CPU cycles T – execution time Network interface Transmission energy consumption b – modulation level d – source-destination distance L – message length W – channel b/w in Hertz T – transmission time

Energy-Time Tradeoff (CPU) Time Energy E Vdd Time T Energy E/4 V ’dd = ½ Vdd Time 2 T 11

DMS (radio): Energy depends … Depends on distance, transmission time. Linearly dependent on transmission time Signal Transmission Energy Radio Energy Consumption Circuit Consumption Energy 12

Computation energy management Dynamic Voltage Scaling (DVS) Communication energy management Dynamic Modulation Scaling (DMS) Varying processor voltage (v) & frequency (f) (V 1, F 1) t 2 t Lower (V 2, F 2); processor slows down Computation time of a task with CC CPU cycles Varying message modulation (b) Comm. energy CPU energy Energy Model: energy vs. delay tradeoff (b 1) t’ 2 t’ lower (b 2) low trans. rate Transmission Delay of message with length L 13

Energy-aware combined scheduling of tasks & messages – the problem Wireless Network Complex periodic tasks T 1 m 2 T 3 T 2 m 4 ( single hop ) m 3 T 4 Deadline = period = D Problem Statement Given: ‘n` such complex Periodic tasks Goal: (1) Schedule Tasks & Messages (2) Assign task frequencies & msg mod. levels Objective: Minimize total system energy consumption. Constraints: Meet all the deadline, precedence & ready time constraints. 14

Energy-aware combined scheduling of tasks & messages – the solution 1. Task Mapping 2. Schedule local tasks on the nodes 3. Schedule msgs on the network Feasible schedule (Energy unaware) 4. Assign modulation levels to messages & frequency levels to tasks. Final Energy Aware Schedule Feasible Schedule T 7 P 1 T 4 M 1 M 2 Ch P 1 T 2 T 5 P 2 T 3 T 6 Use the slack to assign modulation levels to messages & frequency levels to tasks While guaranteeing: (1) Deadline (2) Precedence (3) Ready-time constraints This is an NP-Hard Problem 15

Scheduling – Static & Dynamic Offline Phase Task and message parameters Online Phase P 0 Offline energy-aware Shared wireless Static network Scheduling algorithm System-level energy-time tradeoff Analysis Statically created schedule Energy-Aware Static Sched. Algo Energy-Aware P 3 Dynamic Sched. Algo Other scheduling P 4 problems P 1

System-level Energy vs. Delay Tradeoffs Communicate Compute A Message should reach B before a deadline, D. B TA 0 ∆ MA t 1 t 2 D Comm. energy ? ? (e 1, t 1) t 1 (e 2, t 2) t 2 t 3 (e 3, t 3) Transmission Delay 17

System-level energy-delay tradeoffs 1. Subsequent gains decrease 2. All slack should not go to messages 18

Gain based Static Scheduling (GSS) Insert all messages and tasks into set Q exit Yes Is Q empty ? Remove ei from Q No Pick up the highest energy gain entity Reduce its performance mode by one level ? No Yes Reduce its performance mode by one level

Gain Based Algorithm: Example M 1 T 1 M 1 M 2 b = 7 Gaini, j f = 400 b = 10 f = 300 b = 7 M 2 0 400 300 Can I move to Complexity: (nt + nm)(ntkt+nmkm) the next col. ? 10 Message Movement Table Yes 9 8 7 6 5 4 10 3 8 6 4 T 1 2 Modulation Level (j) Task Movement Table M 1 400 300 M 2 T 1 200 20

Dynamic Slack Utilization – Distributed Algo Shared Wireless Medium P 1 P 2 • Goal: T 3 T 4 P 1 M 7 M 8 Channel P 2 T 1 Utilize dynamic slack performance scaling to further reduce energy consumption T 1 T 2 T 6 • Conditions: (1) Correctness – deadlines & M 9 M 10 precedence constraints T 5 (2) Overhead – no additional messaging Dynamic slack 21

Dynamic Slack Utilization Online Phase P 0 Rules Shared wireless network P 1 Rules P 3 Rules • Use dynamic slack locally. • Do not change the Finish times of any task/msg. P 4 22

Effect of Channel Bandwidth 2. At Low b/w, Comp-only consumes lesser energy 3. At high b/w, Comm-only consumes lesser energy 1. As b/w increases, All schemes consume lesser energy 4. Throughout, GSS performs better than comp-only and comm-only 23

Related Work Research Focus Basic Idea References Processor Energy Management DVS based Task Scheduling. DPM policies. [Aydin et al. , Pillai et al. ] CPU + Memory + I/O DVS based Task scheduling [Shin et al. ] Communication energy management Power Adaptation, DMS, sleep-wakeup [Schurgers et al. ] CPU + Network interface Node Level, DVS + sleep/wakeup [Bren at al. ] Computation sys. + communication sys. Network Level with (DMS + DVS) [Anil et. al. – TPDS 2008] 24

TALK OUTLINE n System-level Energy Management Problem n End-to-End Energy Management Problem [1] n Cyber-Physical System applications n Conclusions [1] G. Sudha Anil Kumar, G. Manimaran, and Z. Wang, "End-to-end energy management in networked real-time embedded systems, " IEEE Trans. on Parallel and Distributed Systems, Dec. 2008. 25

Data Aggregation Tree – End-to-end guarantees Problem • Given: • Aggregation tree • for each node (i) – Ti and Mi • Modulations: [bmin, bmax] • CPU Freq: [fmin, fmax] En d- to - en d d ea dli ne Root/sink Leaf Nodes • Objective: • Minimize total energy consumption (sense, compute, Communicate) • Constraints: • end-to-end deadline (D) • precedence constraints 26

Solution Approach f = 400 TD Obtain a feasible schedule b = 10 Assign message modulation levels and task frequencies MB 318 b = 10 MC MD b = 10 ME 370 Assign Task Frequencies and Message Modulation Levels While guaranteeing Precedence, ready time and end-to-end deadline constraints 27

Solutions space End-to-End Energy Management Problem Continuous Model (not realizable in practice) Optimal Solution Discrete Model (realized in practice) NP Hard Optimal: MILP formulation (worst-case: non-polynomial) Heuristics Scheduling Algorithms (GSA & EGSA) 28

Performance Evaluation n Algorithms/schemes compared n n n Optimal: MILP solved using ILOG CPLEX 10. 100 Proposed: Gain based Algorithm (GSA) Proposed: Extended gain based Algorithm (EGSA) Baseline: comp-only (only tasks are scaled) Baseline: comm. -only (only messages are scaled) Simulation Parameters n n n Bandwidth Radius factor (source – destination distance) Computational Load (cycles per task) 29

2. Energy consumptions increase as we increase radius Effect of Communication Radius 1. At low distance, Comp-only consumes lesser energy 2. Throughout, GSA & EGSA are close to MILP 30

Effect of Computation Load 1. At low Comp. Load, Comm-only consumes less energy 2. Energy consumptions increase as we increase comp. load 2. Throughout, GSA & EGSA are close to MILP 31

Summary of Results n Communication energy consumption is NOT always the dominant factor n Computation energy ~ communication energy consumption n n n At low message modulation levels Low bandwidth channels Short-distance communication High computation load In some cases, computation energy consumption > communication energy consumption System-level energy savings >> component-level savings n 20 -50% improvement for evaluated conditions 32

TALK OUTLINE n System-level Energy Management Problem n End-to-End Energy Management Problem n CPS Applications n Conclusions 33

Cyber Physical Systems (CPSs) Applications: § Critical infrastructure monitoring § Automated traffic control § Home Area Networks § Ubiquitous healthcare monitoring

Smart Grid: A Cyber-Physical System Source: http: //cnslab. snu. ac. kr/twiki/bin/view/Main/Research 35

Wireless Network Design and Fault Diagnosis Design a network for real-time data delivery in presence of latency and bandwidth constraints and an associated fault diagnosis

Wireless Network Design for Transmission line monitoring Given a directed graph G = (V, E) and a set of N flows, Find a feasible path for each flow such that the sum of the cost of all the paths is minimized while respecting the delay and bandwidth constraints of each flow.

Bayesian Network Graph Theory Fault Diagnosis Probability Theory Given: Evidence, E = (e 1, . . . , ek ), where ek is observed state of variable Xi Bayesian Networks Cause II Find: Probability of variable Xj being in a certain state x = P(Xj = x | E) Effect

Sample BN modeling a tower Fault Diagnosis

Conclusions n System Level Energy Management offers significant savings n n Commn (radio) energy is not always the dominant factor n n Dynamic slack generation and utilization are key to energy savings Cyber-Physical System poses constraints for network design n n Depends on: modulation level, Sender-Receiver distance, Bandwidth End-to-end Energy management while meeting deadlines n n CPU time, Communication, Memory, I/0 End-to-end Latency, Bandwidth constraints, legacy comm links Fault diagnosis distinguishes true faults from false positives 40

Future Work n Communication of Energy Management n n n System-level Energy Management n n Leveraging physical layer techniques (Dynamic Code Size Scaling) Network coding + Energy-aware scheduling Exploit sensing redundancy (temporal and spatial) more savings Holistic Scheme: CPU + Commn + Memory + I/O Distributed algorithms Embedded sensor network design and operation (CPS) n n Self-healing, Security, Fault diagnosis, Decision Algorithms Applications of wireless sensor networks are endless ! 41

Thank You ! Acknowlegements Sudha Anil Kumar Benazir Fateh 42