Autonomic Computing A Knowledge Plane for the Internet, D. Clark, J. Ramming, J. Wroclawski, SIGCOMM, August. 2003. . David Choffnes, Winter 2006
The Internet is great, but… Intelligence is only at the edges When failures occur, takes a long time to debug and fix Difficult to configure and administer New goal for the network Understand what it’s being asked to do Take care of itself Internet needs AI/CogSci Need to abstract high-level goals from low-level details Make decisions based on incomplete/imperfect information Learn from previous experience/examples CS 395/495 Autonomic Computing Systems EECS, Northwestern University
CS 395/495 Autonomic Computing Systems EECS, Northwestern University A Knowledge Plane Distributed cognitive system Global vs. regional perspective Edge involvement Composition ability Unified approach Cognitive framework Make judgments in the face of partial/conflicting information Incorporate knowledge representation, learning, reasoning CS 395/495 Autonomic Computing Systems EECS, Northwestern University
CS 395/495 Autonomic Computing Systems EECS, Northwestern University
CS 395/495 Autonomic Computing Systems EECS, Northwestern University Why? Do we need a new construct? Data plane hides information, control plane exposes everything Need middle ground to express goals at a high level and have them automatically fulfilled by tuning at the low level Unified approach Network measurement (everyone uses same info) Tracing a hurricane to the flap of a butterfly’s wings Cognitive System “close the loop” on the network as does an ordinary control system recognize-explain cycle => recognize-explain-suggest cycle => recognize-act cycle for many management tasks the KP must be able to learn and reason model behavior, dependencies, and requirements of applications CS 395/495 Autonomic Computing Systems EECS, Northwestern University
CS 395/495 Autonomic Computing Systems EECS, Northwestern University What is it good for? Fault diagnosis/mitigation WHY, FIX constructs Automatic (re)configuration Ongoing operation to meet goals KP as assistant to network admins Overlay networks KP maintains performance information Knowledge-enhanced IDS Data gathering and correlation CS 395/495 Autonomic Computing Systems EECS, Northwestern University
Knowledge Plane Architecture Distributed organization Bottom-up Constraint-driven E.g., “no multicast” May adopt behavior not specifically constrained Compositional (moves from simple to complex) Global perspective Data/knowledge integration Expect imperfect info Reason about tradeoffs CS 395/495 Autonomic Computing Systems EECS, Northwestern University
Functional/Structural Requirements Gather/Acquire/Generate observations, assertions and explanations about network conditions Cross-regional reasoning Knowledge-driven routing w/ understanding of tradeoffs Trust/Robustness Structural Sensors and actuators Don’t do: Each region reasons about only itself Maybe: Multiple regions compete to provide info about an AS CS 395/495 Autonomic Computing Systems EECS, Northwestern University
CS 395/495 Autonomic Computing Systems EECS, Northwestern University
CS 395/495 Autonomic Computing Systems EECS, Northwestern University
CS 395/495 Autonomic Computing Systems EECS, Northwestern University Creating a KP Building blocks Epidemic algs (dist), Bayesian NWs (learning), rank aggregation (trust), constraint satisfaction algs, policy-based management. Challenges Representing and utilizing knowledge Scalability Routing knowledge Economic incentives Malicious users and trust CS 395/495 Autonomic Computing Systems EECS, Northwestern University