Self Managing Systems: A Control Theory Foundation

1
Self Managing Systems A Control Theory Foundation

• Y. Diao, J. L. Hellerstein, Sujay Parekh, R.
  Griffith, G. Kaiser and D. Phung
• Presented
• Viraj Bhat
• virajb_at_caip.rutgers.edu

2
Background

• Performance Management Group, T.J. Watson
  Research Center
• Controls Group, University of Michigan College of
  Engineering
• Software
• Apache Web Server
• IBM Lotus Notes/Domino
• DB2
• Control System Concepts
• www.research.ibm.com/fbcs/

3
Key Concepts

• Use of classical feedback control theory to build
  self-managing systems
• Explains basics of feedback control theory for
  Digital Systems
• Proposes an architecture for Deployable Testbed
  Autonomic Computing (DTAC)

4
Architecture for Autonomic Computing with Control

• Remains same
• Control Perspective
• Target system is the resource
• Sensors measure output
• Autonomic Managers are controllers
• Objective is a part of the knowledge

5
Feedback Control Theory

• Reference I/P Desired O/P (as specified by the
  human)
• Control Error (Reference I/P Measured O/P)
• Control Input Parameters which affect behavior
  of the system
• Disturbance I/P affects Control I/P
• Controller Change Control I/P to achieve
  Reference I/P
• Measured O/P Measurable feature of the system
• Noise I/P affects Measured O/P
• Transducer Transforms measured O/P to compare
  with Reference I/P

6
Example Apache Web Server

• System cluster of 3 Apache Web Server
• Objective Utilization lt 66
• Measured Output CPU utilization
• Control Input MaxClients
• Disturbances
• Changes in Arrival Rate
• Types of Requests (Static/Dynamic pages)

7
Objectives of Controllers

• Regulatory Control
• Utilization lt 66
• Disturbance Rejection
• Backup/Virus Scan is run Max Utilization lt 66
• Optimization
• Optimize Setting of MaxClients which maximize
  response times

8
Properties of Control Systems

• SASO
• Stable
• Bounded Input produces bounded output
• Unstable systems not usable in mission critical
  work
• Accurate
• Measure Output converges to Reference (Desired)
  Input
• Short Settling Times
• Converges to the Stable Value quickly
• No Overshoot
• Achieves objectives in a steady manner

9
Stable and Unstable Systems
10
Control Analysis and Design
Actual RIS
11
Poles of a Transfer Function

• Poles values of z which make denominator of
  transfer function zero
• Absolute value of pole greater than one System
  is unstable
• Negative/Imaginary Oscillations

Oscillations (higher-freq)
Im(s)
Faster Decay
Faster Blowup
Re(s)
(e-at)
(eat)
12
Effect of Discrete Poles
Im(s)
Higher-frequency response
Longer settling time
Re(s)
Stable
z1
Unstable
13
Modeling Approaches and Challenges

• Modeling Resource Dynamics
• Static System Models
• Do not address dynamics of the systems
• Single Input-Output relationships
• Effective in IBM mainframe systems
• Black Box Method
• Suitable for Multiple Inputs and Multiple Outputs
• Apache HTTP server
• Two control Inputs MaxClients, KeepAlive
• Two measured Outputs CPU, MEM
• Special Purpose Models
• Adaptive Queue Management in Network Routers
• Online Control??

14
Sensors Approaches and Challenges

• Multiple Measurement Sources
• Interval
• Event Data
• Metric Desirable to regulate is not available
• Example Response Tim
• Surrogate Metrics used
• Substantial Overheads in Metrics Collection
• Resource Consumption Measurement consume
  resources
• Measurement Systems have built in Delays
• Measurement tasks not scheduled

15
Effectors Challenges

• Effectors/Actuators complex relationship with
  Measured Output
• IBM Lotus Domino Server (Mail Server)
• Control Active RIS (RPCs in Server) by
  Adjusting MaxUsers
• Connected Users not same as Active users
• CPU priority using nice
• nice affects priorities of a task in Non-linear
  relationship
• Start Fair Queuing (SFQ)
• Loads are heavy dead time is substantial
• Compensation

16
Deployable Testbed for Autonomic Computing (DTAC)

• Test Harness creates requests that is sent to
  HTTP server
• HTTP server process requests
• Application Server forward requests to Data base
  Server
• Sensors and Effectors constitute Managebility
  Middleware
• Control Layer is responsible for making decisions

17
Comments

• Paper gives a overview of Feedback Control
  Systems for Autonomic Computing
• DTAC architecture is too general for any
  computing system for Software Engineering point
  of view
• Linear Model of Resource Dynamics Considered
• System consists of both discrete and event based
  parameters.
• Could have been modeled as a hybrid system

18
Discussions

• Differences between Online Control and Feedback
  Control (Tabular Form)
• Pros and Cons of each approach
• How do you apply this for System of Systems
• Distill out control aspects
• Can Control Theory only suffice for Self-Managing
  Systems
• Hybrid Approach Rule and Control Based approach
  valid
• Reinforced Learning and AI techniques ??
• Can this technique be applied Workflows,
  Coordination and Composition??