Paper Discussion: Online Control for Self-Management in Computing Systems

1
Paper DiscussionOnline Control for
Self-Management in Computing Systems

• Authors Sherif Abdelwahed, Nagarajan Kandasamy,
  and Sandeep Neema

2
Key Points

• Use of control-theory for the design of
  self-managing computer systems
• Focuses on resource management problems, usually
  given as multi-variable optimization problems
  under resource constraints
• Differs from classical feedback control in its
  use of predictive models

3
Problem domain

• Deals with Switching Hybrid Systems
• Not necessarily linear model
• Exhibit both discrete-event and time-based
  dynamics
• Finite set of possible control inputs

4
Motivation

• Recent successful application of
  control-theoretic methods for resource management
  problems (task scheduling, bandwidth allocation
  and QoS adaptation, load balancing, flow control,
  and power management)
• Existence of well-established techniques to
  analyze controller stability and convergence

5
Overview
Predictive filter
Forecast values
Future inputs
Predicted states
Environment parameters
State
Physical system
Control inputs
6
Optimization
7
Power Management Case Study

• State (x) Processor frequency (f ), length of
  the queue (q)
• Environment parameters (w) Task arrival rate (l)
• Control inputs (u) Change of frequency (fi ?
  fj)
• Optimization Minimize queue utilization and
  energy consumption

8
Distributed Control

• Extend concept to systems with several
  distributed components with global goal, but
  local restrictions
• Proposes hierarchical structure, with local
  controllers whose environment parameters are
  managed by a global controller(s)
• Global controllers observe aggregate behavior and
  use a (simplified) system model

9
Main contributions

• Predictive instead of reactive feedback model
• Extension to control of distributed systems
  (global management of local parameters should
  lead to better accuracy of local predictions)

10
Conditions / limitations

• Requires complete and accurate model of the
  system.
• Requires accurate predictive filter Prediction
  model for environment (input) parameters.
• Exhaustive evaluation of available control
  inputs! Suitable for systems having a small
  number of control inputs.

11
Discussion

• When possible control inputs grow, it is
  necessary to apply other optimization techniques
  or heuristics (e.g. local search), but this can
  become domain-dependent.
• Technique doesnt seem amenable to
  generalization.
• Online control seems to be suitable for
  lower-level inner-loop part of an autonomic
  system.