1. Networked Embedded Control System - Integration of control and computing

2. Introduction Cyber physical system What’s the drivers?
Computation/information processing and physical processes are tightly integrated that it is not possible to identify whether behavioral attributes are the result of computations, physical laws, or both working together.
What’s the drivers?
Decreasing cost of computation: Economic motivation for adopting IT in every industry
Environmental pressure: Introduction of IT to improve energy efficiency
Limits of open-loop approach: Though engineering relies on computer based implementations, limited methods and tools are used at the risk of unsafe and unpredictable systems.

3. Environment is changing: Networked Embedded Systems
Embedded systems are becoming increasingly networked
Controller-area-networks (CAN) bus in automobiles
Services in large buildings are now run across networks
e.g. heating, lighting, security 3

4. Cyber-Physical System (CPS)
Definition: Integrations of computation and physical processes
Defining characteristics
Cyber capability in every physical component
Networked at multiple & extreme scales
Complex at multiple temporal & spatial scales
Dynamically reorganizing/reconfiguring
High degrees of automation
Unconventional computational & physical substrates
Operation must be dependable
Goals
Integrated physical and cyber design
New science for future engineered systems (10~20 year perspective)

5. A CPS Example: Electric power grid
Current
Equipment protection devices trip locally
Cascading failure
Future?
Real-time cooperative control of protection devices
Self-healing

6. What’s new? Nothing new What’s difference? Control theory Network
Embedded system
What’s difference?
Higher complexity
More S/W
Large variations in use cases
Distributed
Increased power-awareness
Open loop  closed loop
Homogeneity  Heterogeneity 6

7. Another view from Another perspective: A DDDAS Model (Dynamic, Data-Driven Application Systems)
Discover,
Ingest,
Interact
Models
Discover, Ingest, Interact
Computations
sensors & actuators
s & a
Computational
Infrastructure
(grids, perhaps?)
Cosmological:
10e-20 Hz.
Humans
3 Hz.
Subatomic:
10e+20 Hz.
S p e c t r u m of P h y s i c a l S y s t e m s

8. A DDDAS Example: Forest Fires
Atmospheric
Model
Fire Prop.
Combustion
Policy,
Planning,
Response
Fire
Fighters
Kirk Complex Fire. U.S.F.S. photo

9. Application of computing to control - Example: Control in the Tunnel Scenario
Control over sensor network
Localization and navigation of mobile robot over sensor network
Control of sensor network resources
feedback-based adjustment of radio transmit power in sensor network nodes
Self-organizing middleware
Mobile robot acting as a mobile radio gateway 9

10. Physical network reconfiguration
Partition of network due to failure of sensor nodes
Unreachable nodes 10

11. Physical network reconfiguration
Use mobile agents to restore the communication 11

12. Why “control” is important?
Influencing the behavior of dynamical systems
Lack of control can lead to overcompensation and unstable behavior
Open-loop control
No direct connection between the output of the system and the actual conditions encountered.
 The system cannot compensate for unexpected forces
Closed-loop control
A sensor monitors the output and feeds the data to a controller which adjusts the control input to maintain the desired output.
 The system dynamically compensate for disturbances to the system
Real world (Physical world) requires dynamic adaptation

13. Good Prospects for Control
Conventional static worst-case design is becoming difficult due to increased complexity
New hardware design (MPSoC, Multi-core, etc.)
Large software
Multi-tasking in embedded systems
Dynamic environment
Dynamic load changes
Needs for context-aware services
Power consumption
DVS, DPM, temperature control
Use of feedback control

14. Network Control view of NECS Plant Node Controller Node Actuating
Sensing
Node
Actuating
Sensing
Network
Controller

15. Control systems in computing system’s perspective
Due to economic considerations, in spite of the fast development of the computing hardware, embedded systems are resource constrained
Limitations on: speed, memory size, communication bandwidth, etc.
Use of additional resource (CPU, RAM) is not economically justified.
Cost favors general-purpose computing components over specially designed hardware and software.
Quality of control (QoC) not only depends on the control theoretic methods but also efficient management of resources.
The key to manage resources is “scheduling” resources.
Conventional design of control systems does not consider resource scheduling.
Resulting in “implementation-aware control systems.”

16. Computing systems in control system’s perspective
Computing systems (H/W & S/W) are inherently control systems.
Computing systems are designed with assumptions, so there are uncertainties in resource utilization that affect the performance.
Programs’ execution time
Users’ requests
Input data
Etc.
Regarding complex computing systems as controlled dynamics with defined error terms.
Use of feedback control method to computing systems can increase the flexibility.
Virtually any of computing applications can be considered.

17. Co-Design of Control and Scheduling
Given a set of processes to be controlled and a computer with limited computational resource, design a set of controllers and schedule them as real-time tasks such that the overall control performance is optimized.
K-E. Årzén & A. Cervin. Control and embedded computing: survey of research directions.
Uni-processor case.
Alternative view
Design and schedule a set of controllers such that the least expensive implementation platform can be used while still meeting the performance specifications.

18. Control of computing systems
Conventional approach
Generation of static schedule
Problem
High complexity – longer design time
Longer response time
Hard to use in general-purpose computers
Use of periodic task model
Low utilization due to polling
Complexity in programming due to resource scheduling

19. Applying control theory to scheduling
Feedback controlled scheduling system
e.g. PID control

20. Feedback Controlled EDF
Problem: Only applicable to control relative delay

21. Applying control theory to computing systems - Example
Web-based applications
Web Application Server or HTTP server provides services upon requests from network
Users expect real-time response from server

22. Control of dynamic computing system
Utilization bound for non-periodic tasks:
Implementation: Apache server on Linux (AMD-based PC), HTTP 1.1
From “Schedulability Analysis and Utilization Bounds for Highly Scalable Real-Time Services” by T.F. Abdelzaher and C. Lu,
presented at RTAS 2001

23. Implementation-aware control - effects of computing system
A set of digital control loops
Each controller is realized as a separate period task
The primary goal of co-design approaches becomes optimizing QoC (Quality of Control) under CPU resource constraints
Optimize sampling period, input-output latency subject to performance specification and schedulability.
- maximize
Quality of control
- subject to
Schedulability

24. Effects of sampling period
Sampling frequency affects the system performance
High frequency  high control performance, but high network utilization  Less number of controllers  high cost
Low frequency  low network utilization and low cost, but low control performance
The upper bound of sampling period
Sampling period guaranteeing the stability of the system
Stability constraint
The lower bound of sampling period
Scheduling period from schedulability constraint

25. Control loop timing Main parts of control loop Timing constraints
Data collection
Control algorithm computation
Output transmission
Timing constraints
Sampling period
I/O latency (control delay)
Though sampling frequency at sensor node is fixed, sampling period at controller may have jitter depending on implementations
Scheduling theory can be used to analyze the time variations and delays in control loops.

26. Example Three control tasks T1=12ms, T2=8ms, T3=5ms Control loop
t=current time
LOOP
A/D conversion
ControlAlgorithmExecution
D/A conversion
t=t+h
WaitUntil(t)
END
Priorities are given rate-monotonic.
Execution time is 2ms.

27. Implementation awareness
Preemptive scheduling T3 T2 T1
I/O latency
Sampling period
Non-preemptive scheduling
Sampling period T3 T2 T1
I/O latency
Sampling period

28. Preemptive vs. non-preemptive
Preemptive scheduling
Responsiveness
Favor high priority tasks
(generally) Higher utilization
Non-preemptive scheduling
Introduce blocking time
(generally) Lower utilization
Shorter I/O latency  Control applications’ preference
It is hard to say which one is better

29. Application of computing to control
Networked Embedded Control System (NECS)
Feedback control system wherein the control loops are closed through RTN
Aviation system, automotive system, surveillance system, etc

30. Networked Control Systems
Competing shared network
Network bandwidth = mi/hi
mi = Tc + Tca + Tsc
Hi = Transmission period of each control system
Tc : Computation time
Tca : Controller to actuator transmission time
Tsc : Sensor to controller transmission time

31. Control approach - compensation
Develop compensation method for jitter
Building up control functions for irregular sampling interval
Offline calculation + online control
w/o compensation
With compensation

32. Scheduling approach Use of relative deadline different from period
Drawbacks
Analysis is complex.
There can be utilization loss.
Jitter is reduced

33. Use of non-preemptive scheduling
Utilization can be (virtually) arbitrarily small
Example:
19991 10
10000
20000
T2  ∞

34. Thread-X package Dual priority scheme
Two fixed priorities are assigned to a task.
Priority
Threshold: run-time priority
Only tasks with priority higher than the threshold of the currently running tasks can preempt.
It can achieve higher utilization than premptive and non-preemptive scheduling
Threshold calculation requires complex calculation – done offline (design time)

35. Quantum-based fixed priority scheduling
Combination of priority-based and quantum-based scheduling
Enhances utilization
Adopts non-preemptiveness in preemptive scheduling
Can be easily implemented on general-purpose computers

36. Quantum-based scheduling vs. Thread-X
Achieves higher utilization than Thread-X
 Shorter period can be employed

37. Reducing power consumption
Energy-limited variable voltage microprocessor on which N independent control tasks run concurrently.

38. Low-power control
Find a balanced engineering solution for a control system.
Shorter sampling period  Higher control performance, Higher power consumption
Longer sampling period  Poorer control performance, Lower power consumption
Work in progress currently

39. Conclusion and future directions
Integration of control and scheduling is an emerging field of research.
Application of control theory to computing systems.
Design of implementation-aware control systems.
Future research directions
Control perspective
Event-driven control
Dynamic models of computing systems
Modeling of embedded control systems
Computing perspective
Providing temporal determinism of control tasks
Supporting tools development
Practical implementation