Experimental Network Management

1
Experimental Network Management

• Jennifer Rexford
• Princeton University
• http//www.cs.princeton.edu/jrex

2
Elephant in the Room

• Networks are hard to manage
• More than half the financial cost
• And, around half of the outages
• But, no where near half of the research
• Why?
• Hairy, and preoccupied with todays protocols
• Evaluating and deploying solutions is hard

3
Clean-Slate Network Management

• A contradiction in terms?
• No, design protocols and mechanisms with
  management in mind from the beginning
• That is, design the knobs and dials
• Optimization tuning the network operation
• Select knobs that make optimization tractable
• and have small gap from optimal solutions
• Tomography inferring network properties
• Select dials for simple inference of properties
• balancing trade-off of accuracy and overhead

4
Evaluating and Deploying New Ideas

• Theoretical work
• Computational complexity vs. optimality gap
• Inference accuracy vs. measurement overhead
• Experimental facilities
• Emulab controlled experiments
• VINI mixture of control and realism
• Deployment paths
• Using legacy protocols to talk to the routers
• PC routers running open-source software
• Campus networks eat your own cooking
• End host network troubleshooting tools