autonomic network management, some pragmatic considerations richard mortier, msr cambridge emre...
TRANSCRIPT
Autonomic Network Management,
Some Pragmatic ConsiderationsRichard Mortier, MSR Cambridge
Emre Kıcıman, MSR Redmond
Autonomic network management
• Automatic techniques to increase reliability and performance while reducing cost
• A very wide range of techniques exists– Simplistic control loops– Explicit modelling– Gossip/swarming/biological paradigms– …
• Some work well, some still experimental
Roadmap
• A position paper– Some thoughts on successful applications
based on a few historical examples– I will outline our position, leaving detailed
examples to paper
• Disclaimer– Neither of us are autonomic gurus– Neither of us “do” control theory
Three modest proposals
• In paper, with reference to examples
1. Beware interacting control loops
2. Be explicit (in goals and assumptions)
3. Help people help the system
• Clearly many other useful guidelines exist– Question: where are they written down?
Beware interacting control loops
• I am not a control theory person, but…
• …interacting control loops are complex
• Avoid uncontrolled/accidental interactions – Designed interactions can be very welcome– c.f. Session 1, Session 2
Be explicit
• Making assumptions and goals explicit is just good system design practice– If you don’t know what it’s supposed to do, and when,
how can you know what it will do?
• Automatic determination and handling of failure– Can we minimize intervention by self-testing?
• Design for maintenance, upgrade, extension– No large system is truly static– c.f. Session 1, 2
Help people help the system
• Operator intervention is normal behavior – Blind intervention is highly disruptive– Even partially sighted intervention unhelpful
• Design for measurement– Even in a perfect system, operators will want
to know it’s perfect!– c.f. Session 3
Concluding thoughts
• We proposed some guidelines for ANM– From examining historically successful systems– Prerequisites for modelling, composability– Yes they may seem obvious, but …
• None of them address another very tricky area– “inter-foo” problems: economics, regulations, policy conflicts
• Questions– Can we build even a single self-managing computer?– Is this any simpler than a self-managing network?
• Is it a requirement for a self-managing network?
– When should we stop? • Who are we building the self-managing network for?