1. Large Scale Deeply Embedded Networks Jack Stankovic, Tarek Abdelzaher, Sang Son, Chenyang Lu Department of Computer Science University of Virginia Fall 2001

3. Embedded Systems Engine control Wristwatch Modems Mobile phone Internet appliances Process Control Air Traffic Control 60 Processors in Limo Smart Spaces Sensor/Actuator/CPU clouds with movable entities Smart dust

4. Smart Spaces Smart School Smart Classroom Smart City Smart Factory Pervasive Global Connectivity

5. Sensor/Actuator Clouds Heterogeneous Sensors/Actuators/CPUs (or Homogeneous but Powerful) Resource management, team formation, real-time, mobility, power battlefield awareness earthquake response tracking movements of animals Smart Dust – Biological metaphore smart paint MEMS in human bloodstream

6. Key Issues Enormous numbers of devices and amounts of software needed –flexible and tailorable –interaction with physical/distributed environment (of greater heterogeneity - not just cpus) Aggregation - system as a whole must meet requirements –individual entities not critical Real-Time, Power, Mobility, Wireless, Size, Cost, Security and Privacy

7. How the Problems Change Environment –connect to physical environment (large numbers, dense) –massively parallel interfaces –faulty, highly dynamic, non-deterministic –wireless Network –structure is dynamically changing –sporadic connectivity –new resources entering/leaving –large amounts of redundancy –self-configure/re-configure –individual nodes are unimportant

8. How the Problems Change OS/Middleware –manage aggregate performance control the system to achieve required emerging behavior –move nodes to area of interest (self-organizing) –fuzzy membership and team formation –manage power/mobility/real-time/security tradeoffs –geographically based (data centric)

9. Implications Fundamental Assumptions underlying distributed systems technology has changed –wired => wireless (limited range, high error rates) –unlimited power => minimize power –Non-real-time => real-time –fixed set of resources => resources being added/deleted –each node important => aggregate performance –... New solutions necessary

10. Implications What a single node knows is less important –iterative, diffusion, and masking type algorithms –neural net? –Adaptive control with compensation Resource Management –too many communication errors (feedback control) => move closer, increase power...

11. Example: Consensus Classical consensus: all correct processes agree on one value –No power constraints –No real-time constraints –Does not scale well to dense networks –Approximate agreement (some work here) - on sets of values (physical quantities) Solutions –diffusion and aggregation –Density/topological maps

12. Example continued 1000 nodes to produce signal strength above a threshold –500 enough –turn off others to save power –Don’t want to know which nodes have failed; individual nodes not important Topological model 100% membership 80% membership 30% membership

13. Aggregate Performance Specify and control emerging behavior to meet system-level requirements –Smart Spaces –Smart Clouds of sensors/actuators/cpus –Smart Dust

14. Motivation The emergence of soft real-time systems in unpredictable environments –agile manufacturing –command and control or other defense applications –web browsers (audio and video) –smart spaces –clouds of sensors and actuators –smart dust For Dynamic RTS in unpredictable environments –WCET too pessimistic, high variance in execution time, unbounded arrival rate, overload unavoidable

15. Motivation - What’s Available RT Scheduling Paradigms –Static - predictable all is known a priori (WCET, invocation times or worst case rates, resource needs, precedence, etc.) Cyclic scheduling, RM, table driven, … open loop –Dynamic - high degree of predictability all is known except invocation times/rates (use WCET) Spring algorithm - admission control and planning, or EDF open loop

16. Claims Despite the significant body of results in real- time scheduling many real world problems are NOT easily supported! We need a new real-time scheduling paradigm based on feedback !!!

17. Goals Theory and Practice of Feedback Control Real-Time Scheduling PID control (but not restricted to this) –does not require precise analytical model of the system being controlled –not ad hoc either Explicit use of deadline based metrics Versus new types of consensus based algorithms

18. Network Architectures - Classical HierarchicalNeighborhood

19. Network Architectures - Non-classical Clouds of sensors/actuators/cpus –network architecture dynamically changing (fast) –subject to high error rate –new resources entering and leaving due to mobility, faults, …. –Power/mobility/communication/computation/secu rity tradeoffs Smart Dust –very simple controllers