A Survey of Autonomic Computing: Degrees, Models, and Applications

Huebscher, McCann

autonomic computing self management fault tolerance robust redundancy

@article{huebscher:csur-2008,
  title={A Survey of Autonomic Computing: Degrees, Models, and Applications},
  author={Huebscher, M.C. and McCann, J.A.},
  journal={{ACM} Computing Surveys ({CSUR})},
  volume={40},
  number={3},
  pages={7},
  year={2008},
  publisher={{ACM}}
}

Autonomic computing is not just hype or a rebranding

• Yet many of the examples given in the brief history are a rebranding
  • DARPA SAS: Ad hoc routing...
  • NASA DS1, Pathfinder: Planning, mixed-initiative planning...
  • NASA ANTS: Swarms, multi-agent systems

Main properties: Self configuration, self optimization, self healing, self protection

• Self healing interpreted broadly, everything from bit errors on chip to erroneous entries in directory services
• Proactive and reactive

Inspired by properties identified by Wooldridge and Jennings:

• Autonomy
• Social Ability
• Reactivity
• Proactiveness

Big difference against previous work is dynamic adaptation to current environmental context

• Versus query optimizers, resource managers, or routing software
• But those are generally or frequently dynamic already???

Really needs to include more than one of the properties outlined above

Example: Changing codecs to match video/audio stream conditions

MAPE-K: Monitor, Analyze, Plan, Execute, Knowledge

• Looks very similar to Russell & Norvig's agent loop
• Really just adding more elements to what the agent is doing inside
  • Unnecessarily binding, e.g., explicit knowledge is not useful in many problems

Goals usually expressed via event-condition-action pairs

• How to do conflict resolution, e.g., multiple policies entailed?
• Who's specifying these in the first place?

Goal policies more flexible, but more resource intensive

• Utility functions quantify states, blending goals
  • Hard to define

Sometimes legacy systems can't be integrated with autonomic management, forcing them to be decoupled

Probes and gauges provide input to system

• Probes "are system-specific sensors that extract data from a managed element."
• Gauges filter, aggregate, and process that data to report to higher level components
• Probes and gauges need not be on the same machine
• Probes are specific to the managed element
• Gauges are specific to the management problem
• Reusable across systems and applications

Passive and active monitoring

• Probably only need a small subset of sensors to classify against large set of system states
• Adaptive monitoring policies, backing off in quiescent states, engaging more when situation is changing

May need model of system, e.g., architecture diagram

Latency is important, can't take too long to make a decision and execute or the situation will have changed

... policies are not just event-condition-action pairs, depending on what kind of policy you're talking about...

Techniques

• Planning section just lists some very similar archicture model and ECA based approaches
• Lists main forms of knowledge in autonomic systems as utility, reinforcement learning, Bayesian techniques

In multi-tier system, each tier typically implements full MAPE-K loop

Humans may need to intervene sometimes

"In many if not all of the systems we have examined for this paper, the logic of what to do when has been more or less set statically and only updated by the parameters being fed into the loop. A closed loop system would also evolve the logic that drives the system depending on how well the "old" logic did. This is akin to the work carried out in AI and is an area of research we can see becoming of more interest to the autonomic community."

• Umm...

Changes IBM's autonomous systems levels to:

• Support
• Core
• Autonomous
• Autonomic
• Closed Loop

Emerging application areas

• Power management
• Ubiquitous computing
  • "The advancement in electronics that enable the integration of complex components into smaller devices coupled with recent developments in wireless, mobile communications, have contributed to the emergence of a new class of wireless ad-hoc networks: Sensor Networks." ... "Currently, such systems are described at two levels of granularity, smart dust and general ubiquitous computing;"
    • Umm...

Hysteresis is a concern; needs some preference for stability

Software engineering practices may have to change, to better specify, test, and develop such dynamic systems

Several typos through, e.g., ridged instead of rigid, know instead of now, etc.